Authors Aaron J. Snoswell, Kevin Witzenberger, Rayane El Masr
Date 15 April 2025

Earlier this year, scientists discovered a peculiar term appearing in published papers: “vegetative electron microscopy”.

This phrase, which sounds technical but is actually nonsense, has become a “digital fossil” – an error preserved and reinforced in artificial intelligence (AI) systems that is nearly impossible to remove from our knowledge repositories.

Like biological fossils trapped in rock, these digital artefacts may become permanent fixtures in our information ecosystem.

The case of vegetative electron microscopy offers a troubling glimpse into how AI systems can perpetuate and amplify errors throughout our collective knowledge.

A bad scan and an error in translation

Vegetative electron microscopy appears to have originated through a remarkable coincidence of unrelated errors.

First, two papers from the 1950s, published in the journal Bacteriological Reviews, were scanned and digitised.

However, the digitising process erroneously combined “vegetative” from one column of text with “electron” from another. As a result, the phantom term was created.

Decades later, “vegetative electron microscopy” turned up in some Iranian scientific papers. In 2017 and 2019, two papers used the term in English captions and abstracts.

This appears to be due to a translation error. In Farsi, the words for “vegetative” and “scanning” differ by only a single dot.

An error on the rise

The upshot? As of today, “vegetative electron microscopy” appears in 22 papers, according to Google Scholar. One was the subject of a contested retraction from a Springer Nature journal, and Elsevier issued a correction for another.

The term also appears in news articles discussing subsequent integrity investigations.

Vegetative electron microscopy began to appear more frequently in the 2020s. To find out why, we had to peer inside modern AI models – and do some archaeological digging through the vast layers of data they were trained on.

Empirical evidence of AI contamination

The large language models behind modern AI chatbots such as ChatGPT are “trained” on huge amounts of text to predict the likely next word in a sequence. The exact contents of a model’s training data are often a closely guarded secret.

To test whether a model “knew” about vegetative electron microscopy, we input snippets of the original papers to find out if the model would complete them with the nonsense term or more sensible alternatives.

The results were revealing. OpenAI’s GPT-3 consistently completed phrases with “vegetative electron microscopy”. Earlier models such as GPT-2 and BERT did not. This pattern helped us isolate when and where the contamination occurred.

We also found the error persists in later models including GPT-4o and Anthropic’s Claude 3.5. This suggests the nonsense term may now be permanently embedded in AI knowledge bases.

Author ADM+S Centre
Date 17 April 2025

Ever wondered why certain videos show up in your TikTok Feed? Does TikTok know exactly what you like, or does it nudge you to like things? Whether you’re scrolling, liking, or making content, the TikTok algorithm is learning from you.

Using a new tool created by researchers from the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) at QUT and The University of Sydney, TikTok users can now see themselves through the eyes of the algorithm. 

Launched ahead of the upcoming federal election, the For You Research Project explores how TikTok’s powerful recommendation algorithm is influencing culture, creativity, and public discourse in Australia.

“We’re using a new and exciting approach which is based on citizen science,” says lead researcher Patrik Wikström.

“TikTok is a highly personalised platform,” says Professor Wikström, and it’s “shaping what Australians see online”.

Participants are invited to join the project to learn what is shaping their TikTok Feed and get a new perspective on the role of TikTok in their lives.

Author ADM+S Centre
Date 12 April 2025

On 9 April, Associate Prof Julia Stoyanovich, Director of the Center for Responsible AI at NYU Tandon School of Engineering and Partner Investigator at the ARC Centre of Excellence for Automated Decision-Making and Society, testified at the Research & Technology Subcommittee Hearing – DeepSeek. A Deep Dive.

Her testimony focused on the national security and competitive advantage implications of DeepSeek for the US.

“It was an honor and a privilege to testify at the U.S. House of Representatives today, at a Research & Technology Subcommittee Hearing of the Committee on Science, Space, and Technology,” said Prof Stoyanovich.

In her remarks, Professor Stoyanovich offered three key recommendations with regards to the technology implications of DeepSeek:

Recommendation 1: Foster an Open Research Environment
To close the strategic gap, the federal government must support an open, ambitious
research ecosystem. This includes robust funding for fundamental AI science, public datasets, model development, and compute access. The National AI Research Resource (NAIRR) is essential here—providing academic institutions, startups, and public agencies with tools to compete globally. Federal support for the National Science Foundation and other agencies is vital to sustaining open research and building a skilled AI workforce.

Recommendation 2: Incentivise Transparency Across the AI Lifecycle
Transparency drives progress, safety, and accountability. The government should require public disclosure of model architecture, training regimes, and evaluation protocols in federally funded AI work—and incentivize similar practices in commercial models. Public benchmarks, shared leaderboards, and reproducibility audits can raise the floor for all developers.

Author T.J. Thomson
Date 11 April 2025

You may have seen ads by Apple promoting its new Clean Up feature that can be used to remove elements in a photo. When one of these ads caught my eye this weekend, I was intrigued and updated my software to try it out.

The feature has been available in Australia since December for Apple customers with certain hardware and software capabilities. It’s also available for customers in New Zealand, Canada, Ireland, South Africa, the United Kingdom and the United States.

The tool uses generative artificial intelligence (AI) to analyse the scene and suggest elements that might be distracting. You can see those highlighted in the screenshot below.

You can then tap the suggested element to remove it or circle elements to delete them. The device then uses generative AI to try to create a logical replacement based on the surrounding area.

Easier ways to deceive

Smartphone photo editing apps have been around for more than a decade, but now, you don’t need to download, pay for, or learn to use a new third-party app. If you have an eligible device, you can use these features directly in your smartphone’s default photo app.

Apple’s Clean Up joins a number of similar tools already offered by various tech companies. Those with Android phones might have used Google’s Magic Editor. This lets users move, resize, recolour or delete objects using AI. Users with select Samsung devices can use their built-in photo gallery app to remove elements in photos.

There have always been ways – analogue and, more recently, digital – to deceive. But integrating them into existing software in a free, easy-to-use way makes those possibilities so much easier.

Using AI to edit photos or create new images entirely raises pressing questions around the trustworthiness of photographs and videos. We rely on the vision these devices produce in everything from police body and traffic cams to insurance claims and verifying the safe delivery of parcels.

If advances in tech are eroding our trust in pictures and even video, we have to rethink what it means to trust our eyes.

How can these tools be used?

The idea of removing distracting or unwanted elements can be attractive. If you’ve ever been to a crowded tourist hotspot, removing some of the other tourists so you can focus more on the environment might be appealing (check out the slider below for an example).

But beyond removing distractions, how else can these tools be used?

Some people use them to remove watermarks. Watermarks are typically added by photographers or companies trying to protect their work from unauthorised use. Removing these makes the unauthorised use less obvious but not less legal.

Others use them to alter evidence. For example, a seller might edit a photo of a damaged good to allege it was in good condition before shipping.

As image editing and generating tools become more widespread and easier to use, the list of uses balloons proportionately. And some of these uses can be unsavoury.

AI generators can now make realistic-looking receipts, for example. People could then try to submit these to their employer to get reimbursed for expenses not actually incurred.

Author ADM+S Centre
Date 10 April 2025

Users of Amazon’s Alexa-enabled Echo devices in Australia and around the world may have noticed something different — or perhaps they haven’t. That’s part of the concern, according to experts.

In a recent interview with ABC Radio National’s Life Matters, Prof Daniel Angus, Chief Investigator at the ARC Centre of Excellence for Automated Decision-Making and Society, explained that Amazon has made a significant and controversial change: all audio captured by its Echo smart speakers is now automatically sent to the cloud by default.

Users can opt out, but doing so limits the device’s ability to personalise responses and learn user preferences. The move has sparked a new wave of concern over consumer privacy, AI hype, and the growing power of Big Tech.

“This move is diabolical,” Prof Angus told Life Matters. “It breaks that fundamental trust.”

A Symptom of the AI Hype Cycle

Prof Angus argues that Amazon’s decision is not just a privacy issue, but part of a broader, more concerning trend.

“We’re in a hype cycle around AI,” he said. “Companies need us to believe in the idea of AI to maintain growth. This is not just about functionality — it’s about market dominance and feeding the myth of inevitable AI revolution.”

At the core of this trend is data. More data means better AI models, and smart speaker interactions — even something as simple as setting a timer — are a rich source of training material.

He pointed to Amazon’s market saturation and reliance on growth-at-all-costs as motivations for expanding data collection practices without clear consumer benefit.

Author ADM+S Centre
Date 4 April 2025

A new autistic-led project, Autistic Supports for Comfort, Care, and Connection, reveals the everyday and creative ways autistic adults use objects, services, and creatures to support their wellbeing.

Conducted by Dr Megan Rose, research fellow, and Prof Deborah Lupton, from the ARC Centre of Excellence for Automated Decision-Making and Society at UNSW, the study interviewed 12 autistic Australians about the non-human supports they rely on for entertainment, social connection, special interests, burnout recovery, sensory challenges, and overall wellbeing.

Participants also imagined their ideal new support system tailored to their needs.

To visually represent these experiences, autistic graphic illustrator Sarah Firth was commissioned to create unique ‘portraits’ of each participant. Using anonymised interview transcripts, Sarah crafted illustrations that depict the challenges, coping strategies, and special interests of each individual—without ever seeing or meeting them.

Author T.J. Thomson
Date 2 April 2025

You wouldn’t usually associate Pikachu with protest.

But a figure dressed as the iconic yellow Pokémon joined a protest last week in Turkey to demonstrate against the country’s authoritarian leader.

And then a virtual doppelgänger made the rounds on social media, raising doubt in people’s minds about whether what they were seeing was true. (Just to be clear, the image in the post shown below is very much fake.)

This is the latest in a spate of incidents involving AI-generated (or AI-edited) images that can be made easily and cheaply and that are often posted during breaking news events.

Doctored, decontextualised or synthetic media can cause confusion, sow doubt, and contribute to political polarisation. The people who make or share these media often benefit financially or politically from spreading false or misleading claims.

How would you go at telling fact from fiction in these cases? Have a go with this quiz and learn more about some of AI’s (potential) giveaways and how to stay safer online.

Authors T.J. Thomson, Michelle Riedlinger, Phoebe Matich, Ryan J. Thomas
Date 2 April 2025

Generative artificial intelligence (AI) has taken off at lightning speed in the past couple of years, creating disruption in many industries. Newsrooms are no exception.

A new report published today finds that news audiences and journalists alike are concerned about how news organisations are – and could be – using generative AI such as chatbots, image, audio and video generators, and similar tools.

The report draws on three years of interviews and focus group research into generative AI and journalism in Australia and six other countries (United States, United Kingdom, Norway, Switzerland, Germany and France).

Only 25% of our news audience participants were confident they had encountered generative AI in journalism. About 50% were unsure or suspected they had.

This suggests a potential lack of transparency from news organisations when they use generative AI. It could also reflect a lack of trust between news outlets and audiences.

Who or what makes your news – and how – matters for a host of reasons.

Some outlets tend to use more or fewer sources, for example. Or use certain kinds of sources – such as politicians or experts – more than others.

Some outlets under-represent or misrepresent parts of the community. This is sometimes because the news outlet’s staff themselves aren’t representative of their audience.

Carelessly using AI to produce or edit journalism can reproduce some of these inequalities.

Our report identifies dozens of ways journalists and news organisations can use generative AI. It also summarises how comfortable news audiences are with each.

The news audiences we spoke to overall felt most comfortable with journalists using AI for behind-the-scenes tasks rather than for editing and creating. These include using AI to transcribe an interview or to provide ideas on how to cover a topic.

But comfort is highly dependent on context. Audiences were quite comfortable with some editing and creating tasks when the perceived risks were lower.

The problem – and opportunity

Generative AI can be used in just about every part of journalism.

For example, a photographer could cover an event. Then, a generative AI tool could select what it “thinks” are the best images, edit the images to optimise them, and add keywords to each.

These might seem like relatively harmless applications. But what if the AI identifies something or someone incorrectly, and these keywords lead to mis-identifications in the photo captions? What if the criteria humans think make “good” images are different to what a computer might think? These criteria may also change over time or in different contexts.

Even something as simple as lightening or darkening an image can cause a furore when politics are involved.

AI can also make things up completely. Images can appear photorealistic but show things that never happened. Videos can be entirely generated with AI, or edited with AI to change their context.

Generative AI is also frequently used for writing headlines or summarising articles. These sound like helpful applications for time-poor individuals, but some news outlets are using AI to rip off others’ content.

AI-generated news alerts have also gotten the facts wrong. As an example, Apple recently suspended its automatically generated news notification feature. It did this after the feature falsely claimed US murder suspect Luigi Mangione had killed himself, with the source attributed as the BBC.

What do people think about journalists using AI?

Our research found news audiences seem to be more comfortable with journalists using AI for certain tasks when they themselves have used it for similar purposes.

For example, the people interviewed were largely comfortable with journalists using AI to blur parts of an image. Our participants said they used similar tools on video conferencing apps or when using the “portrait” mode on smartphones.

Likewise, when you insert an image into popular word processing or presentation software, it might automatically create a written description of the image for people with vision impairments. Those who’d previously encountered such AI descriptions of images felt more comfortable with journalists using AI to add keywords to media.

AuthorADM+S Centre
Date 31 March 2025

Are you a renter in Australia with experience using digital rental platforms? A new research project is looking for participants to share their experiences with ‘RentTech’ and its impact on housing justice.

PhD researcher Samantha Floreani from the ARC Centre of Excellence for Automated Decision-Making and Society at Monash University is conducting a study on the growing influence of digital technologies in the residential real estate sector.

These technologies—sometimes referred to as ‘RentTech’—include online rental application platforms (such as 2Apply, Sorted, Ignite, and Snug), property management apps (Kolmeo, Cubbi, ConsoleTenant), and rent payment platforms (Rental Rewards, Ailo, SimpleRent), among others. The research aims to explore how these technologies affect renters’ experiences and housing justice in Australia.

Samantha says, “Against the backdrop of the ongoing housing crisis, renters are increasingly interacting with digital technologies at every stage of their housing experience.

“These tools come with promises of increased convenience, efficiency, and profit for real estate agents and landlords—but what do they mean for renters? Through this study, I aim to find out.”

Authors Simon Coghlan, Damiano Spina, Falk Scholer and Hui Chia
Date 26 March 2025

Every day, users ask search engines millions of questions. The information we receive can shape our opinions and behaviour.

We are often not aware of their influence, but internet search tools sort and rank web content when responding to our queries. This can certainly help us learn more things. But search tools can also return low-quality information and even misinformation.

Recently, large language models (LLMs) have entered the search scene. While LLMs are not search engines, commercial web search engines have started to include LLM-based artificial intelligence (AI) features into their products. Microsoft’s Copilot and Google’s Overviews are examples of this trend.

AI-enhanced search is marketed as convenient. But, together with other changes in the nature of search over the last decades, it raises the question: what is a good search engine?

Our new paper, published in AI and Ethics, explores this. To make the possibilities clearer, we imagine four search tool models: Customer Servant, Librarian, Journalist and Teacher. These models reflect design elements in search tools and are loosely based on matching human roles.

The four models of search tools

Customer Servant

Workers in customer service give people the things they request. If someone asks for a “burger and fries”, they don’t query whether the request is good for the person, or whether they might really be after something else.

The search model we call Customer Servant is somewhat like the first computer-aided information retrieval systems introduced in the 1950s. These returned sets of unranked documents matching a Boolean query – using simple logical rules to define relationships between keywords (e.g. “cats NOT dogs”).

Librarian

As the name suggests, this model somewhat resembles human librarians. Librarian also provides content that people request, but it doesn’t always take queries at face value.

Instead, it aims for “relevance” by inferring user intentions from contextual information such as location, time or the history of user interactions. Classic web search engines of the late 1990s and early 2000s that rank results and provide a list of resources – think early Google – sit in this category.

Journalist

Journalists go beyond librarians. While often responding to what people want to know, journalists carefully curate that information, at times weeding out falsehoods and canvassing various public viewpoints.

Journalists aim to make people better informed. The Journalist search model does something similar. It may customise the presentation of results by providing additional information, or by diversifying search results to give a more balanced list of viewpoints or perspectives.

Authors Ned Watt and Michelle Riedlinger
Date 25 March 2025

The loss of Australia’s go-to political fact-checker and the rise of AI tools has created a crisis for political accountability just as the nation’s voters prepare to go to the polls.

Professional fact-checkers have never been under more pressure and social media users face a complex and fast-evolving misinformation landscape.

It’s crucial voters understand the situation in the lead-up to the vote.

This federal election will be the first without ABC RMIT Fact Check, which completed its first fact check during the Rudd-Abbott election of 2013.

It will also be Australia’s first federal election since the release of ChatGPT and other generative AI tools that have heralded a new normal of AI-generated political advertising and propaganda.

The risk to accountability is a win for vested interests in Australia’s political and media systems. It means there is even more potential for those vested interests to manipulate information for their own benefit rather than the public good.

Australia needs political parties to commit to ethical use of AI in their campaigning and bipartisan support for improved human-AI detection tools created by fact-checkers and for fact-checkers and journalists, to improve media information integrity systems.

What happened to political fact-checking

Independent fact-checking has faced a public legitimacy crisis in the past few years, mirroring similar crises of trust in news.

The crisis is driven in part by politicians’ denigrations of online investigative research activities, which is related to distrust of the fact-checking movement by far-right politicians and their allies around the world.

In Australia, the end of fact-checking arrangements between the ABC and RMIT University happened in 2024 amid a media furore in the lead-up to the Voice to Parliament referendum. Conservative media depicted RMIT Fact Lab, another entity under RMIT’s professional fact-checking wing, as grossly biased.

Claims of fact-checkers’ political bias hinge on observations that right-leaning voices tend to share news content that diverges from established consensus more often, resulting in a relatively high proportion of their claims being fact-checked.

The suspension of RMIT Fact Lab’s membership of Meta’s third-party fact-checking program cast a long shadow over the credibility of fact-checking, reflecting similar questions to those recently posed in the United States about the role of truth in politics.

Australia still has two locally-owned fact-checking units – ABC’s in-house fact-checker ABC News Verify and Australian Associated Press (AAP) fact-checking service – as well as AFP Australia, the local division of Agence France-Presse’s (AFP) fact-checking operation.

Australian fact-checkers have been part of a push for political accountability and depolarisation, responding to the concerns of Australians about the interplay between private interests in politics and media organisations and the public interest, including the roles of big tech in moderating information online.

There have been calls for greater accountability and transparency in news reporting, but fact-checkers worldwide have experienced setbacks.

At the start of the year, Meta announced that it was ending its third-party fact-checking program in the United States and making changes to its content-moderation policies. The changes would amplify political content and allow content targeting vulnerable minorities that it previously considered contentious and divisive.

Author ADM+S Centre
Date 24 March 2025

As Australia approaches its federal election, concerns are mounting over the spread of misinformation and disinformation on Chinese social media platform RedNote or known as the little red book for Mandarin-speakers. RedNote is a platform increasingly used by Australian politicians to connect with Chinese Australians.

In addition to informational and educational content, deepfake videos, political-or-commerical-driven misleading content, and shadow banning are emerging as key issues in the digital landscape, raising alarm over the integrity of online political discourse.

In a recent investigation, the ABC uncovered a deepfake video featuring a manipulated clip of Opposition Leader Peter Dutton speaking Mandarin on RedNote.

The video uses legitimate footage from an interview where Dutton discusses the Indigenous flag, but AI has altered it to make it appear as though he is speaking Mandarin. In the video, Dutton appears to suggest that Indigenous flags should not be displayed at press conferences, a claim that is misleading and taken out of context.

ARC Centre of Excellence for Automated Decision-Making and Society researchers Dr Fan Yang, from Melbourne University and Dr Robbie Fordyce from Monash University discussed the issue in an interview on ABC’s World Today.

Dr Fan Yang studies Australian political information on Chinese-language social media services and warns that such deepfake videos are not isolated incidents. She notes that other misleading content has spread on the platform, such as videos implying the Albanese government is arresting temporary migrants and commercial-driven threatening messages about Australia’s new policies on immigration and housing.

In these cases, the videos are often taken out of context, with captions misrepresenting the events.

Further complicating matters, Dr Yang points out the lack of official voices, such as those from the Australian Electoral Commission (AEC) on RedNote to prebunk and debunk false or misleading information; and that the narrow scope of what public agencies classify as “misinformation” and “disinformation” limits their capacity for effective intervention. She highlights troubling instances of shadow banning, where Australian politicians’ accounts and content are hidden from Chinese Australian users.

“If you search for the name of a politician, you wouldn’t even be able to find their account,” Dr Yang explains.

“This raises concerns that Chinese Australians are being exposed to an increasingly one-sided view of political events.”

Following the publication of an ABC investigative report, on 24 March, ADM+S affiliated PhD researcher Dan Dai identified platform intervention of the latest content in relation to the hashtags #澳大利亚大选 or #澳洲大选（meaning Australian election) on RedNote. No recent content appears in search results for the term.

Author Kathy Nickels
Date 17 March 2025

The Office of the National Broadcasting and Telecommunications Commission of Thailand (Office of the NBTC), and RMIT University, Australia, have formalised a new partnership with the signing of a Memorandum of Understanding (MOU) on 17 February 2025.

This landmark agreement aims to foster international collaboration in academic and research endeavors, contributing to the shared goals of addressing global challenges related to digital access and policy-oriented research.

Office of the NBTC, a leading independent state body that regulates broadcasting, television, radiocommunications, and telecommunications across Thailand and the 10 other countries in the Association of Southeast Asian Nations (ASEAN), will collaborate with RMIT.

The partnership will also involve researchers from the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) to advance research and development in areas critical to shaping future policies and regulatory decisions.

Author ADM+S Centre
Date 14 March 2025

A recent study has uncovered significant concerns surrounding the use of Large Language Models (LLMs) to assess the relevance of information, particularly in passage labelling tasks.

This research investigates how LLMs label passages of text as “relevant” or “non-relevant,” raising new questions about the accuracy and reliability of these models in real-world applications, especially when they are used to train ranking systems or replace humans for relevance assessment.

The study, which received the “Best Paper Honorable Mention” at the SIGIR-AP Conference on Information Retrieval in Tokyo in December 2024, compares the relevance labels produced by various open-source and proprietary LLMs with human judgments.

It finds that, while some LLMs agree with human assessors at similar levels of human-to-human agreement as measured in past research, they are more likely to label passages as relevant. This suggests that while LLMs’ “non-relevant” labels are generally reliable, their “relevant” labels may not be as dependable.

Marwah Alaofi, a PhD student at the ARC Centre of Excellence for Automated Decision-Making and Society, supervised by Prof Mark Sanderson, Prof Falk Scholer, and Paul Thomas, conducted the study as part of her research into measuring the reliability of LLMs for creating relevance labels.

Authors Katharina Esau, Axel Bruns and Tariq Choucair
Date 13 March 2025

Australian voters are being targeted by divisive ‘them vs us’ strategies that overshadow policy debate. Here are the signs and ways to move past the soundbites.

Politicians and media organisations are setting the stage for an Australian election where division is a deliberate strategy to mobilise supporters, discredit opponents and split undecided voters.

Polarisation is already shaping the national conversation and it’s a tactic born out of much more than just differing views.

Democratic debate thrives on differing opinions but excessive polarisation pushes discussion away from constructive engagement and into entrenched conflict that has negative consequences for democracy and broader society.

Voters need to know how to spot the signs of those ‘conflict strategies’ and to question them; to look past the soundbites for information they can trust that doesn’t break every debate down to ‘us vs them’.

Negative campaigning — attacking instead of selling your own policies — has been a feature of democratic elections for centuries. Such tactics are designed to stir emotional reactions about an opponent.

It has become a standard election strategy.

Australians might particularly associate it with former Liberal Prime Minister Tony Abbott. Abbott was known for his ruthless negativity as opposition leader from 2009 and his attack-ad-driven campaign in 2013.

How polarisation turns destructive
Now tactics have shifted to a form of strategic polarisation designed to do much more than merely discredit opponents — now the aim is to stoke all-encompassing divisions across society.

These tactics were seen in spectacular fashion in the 2016 US general election.

They were mirrored globally, with examples including Jair Bolsonaro in Brazil, Rodrigo Duterte in the Philippines and the 2017 presidential election in France.

Politicians framed their opponents as an existential threat, encouraging voters not just to support their own side but to despise their opponents.

In Australia, US influence looms large and Trump’s return to the presidency has emboldened politicians here to double down on similar strategies.

In a healthy democracy, competing parties debate ideas, disagree strongly and propose diverging solutions.

However, when polarisation becomes destructive it has potentially severe consequences for democracy and societal cohesion.

There are five key symptoms of destructive polarisation, all seen in recent Australian and global political contests.

When dialogue becomes impossible
A key symptom is that communication between opposing sides ceases to function. Rather than engaging in constructive debate, political actors, media producers and the public either avoid meaningful interaction or reduce their exchanges to misrepresenting, insulting and attacking each other.

This can be seen when party leaders trade insults and shout slogans during campaign debates, rather than debating policy, and in quips or ‘fails’ later pushed on social media.

Australia’s winner-takes-all political system — where coalitions between parties are rare — further exacerbates this. For political leaders, the ability to find compromise and build consensus is seen as a weakness rather than a strength.

When facts don’t matter
Political actors and supporters might also dismiss information outright, based on the source rather than the content.

This might target think tanks or media outlets that might be seen as aligned with one side of politics. Even independent institutions or entire professions such as researchers, public servants or journalists might be dismissed as inherently biased.

Social media users then employ the same strategy, rejecting information based on its source rather than engaging with the information.

When policy becomes a slogan war
Destructive polarisation thrives on reducing nuanced debates to misleading black-and-white choices.

Translating a complex problem or policy into simple terms is one thing, but something else is at play here.

Instead of explaining the issue and their proposed policies, candidates often oversimplify by attacking their political opponents or sometimes by blaming minorities. The message becomes: “If you support us, the problem will be solved. If you support them, we are all doomed.”

The goal of this kind of messaging is to reduce a complex issue to partisan blame or the scapegoating of entire social groups — such as migrants — while ignoring the factors contributing to a policy problem.

Authors T.J Thomson and Aimee Hourigan
Date 13 March 2025

Australians are swamped by misinformation every day but they’re smart enough to know they need help to better navigate an untrustworthy online world.

False online claims about business and the economy top the list of misinformation concerns for Australians and research indicates they are screaming out for help on how to deal with it.

In some ways, it’s not surprising misinformation on the economy rates so highly during a cost-of-living crisis and with a federal election looming — finance-related scams are also a concern — but they’re just a few areas highlighted as Australians drown in a sea of questionable claims every day.

Online misinformation and disinformation have been labelled bigger short-term global threats than climate change or war, so improving media literacy is a critical step in fighting it.

That need is stark considering researchers — whose report, Online misinformation in Australia, was published late last year — found more than half of dodgy information is reported as coming from news sources, whether that be traditional or alternative forms of media.
Australians encounter hundreds of claims each day through channels that might include listening to a podcast, scrolling social media or reading the news and when surfing the Internet to shop, learn or seek entertainment.

The challenge is assessing how many of these claims are true and how confident Australians can be in their abilities to separate fact from fiction.

Half of us encounter misinformation weekly
The researchers found more than half of Australians encounter misinformation in a typical week and 97 percent of Australians have poor or limited ability to verify claims they encounter online.

The research has shone a light on the sources of everyday misinformation, the topics covered and how and where those claims are communicated. It also offers suggestions on how to respond.

Research participants were asked to document online news and information they saw each day for a week and rate its trustworthiness.

More than 20 percent of the 1,600 examples provided were perceived by the participants to have false or misleading claims.

Those misleading claims weren’t limited to the usual suspects such as health or political information, but ranged across other topics that included celebrity news, entertainment and sports.

False or misleading claims about business and economics were the most prevalent. A cost-of-living crisis and heightened focus on money can attract both those who don’t have it as well as those who want to exploit the vulnerable for financial, political or other gain.

It’s only logical then that scams feature high on the list of misinformation threats worrying Australians.

Author Natalie Campbell
Date 7 March 2025

The Joint Committee of Public Accounts and Audit has published its report on the Inquiry into the Use and Governance of AI by Public Sector Entities, citing the ADM+S submission throughout.

Responding to the steep increase in AI adoption by public sector entities that was found during the 2022-23 Commonwealth Financial Statements, the Committee established a specific Inquiry into the Use and Governance of AI by Commonwealth Entities in September 2024.

Chair of the Committee, Hon Linda Burney MP explained, “The issue that was fundamental to this inquiry was whether the existing governance and oversight of this technology matches its rapid and continuing advancement.

“Policy frameworks must be equipped to adequately assess the great promise that AI brings but also understand the inherent and significant risks that accompany its use.”

In February 2025 the Committee released a report titled ‘Proceed with Caution’, which provides four key recommendations.

1. The Australian Public Service Commission to introduce questions on the use and understanding of artificial intelligence and other emerging technologies into its annual APS Employee Census.
2. The Australian Government convenes a whole of Government working group within 12 months of this report to develop key frameworks for managing sovereign risks, and biases that result from the adoption of these technologies can be effectively mitigated.
3. The Australian Government establishes a statutory Joint Committee on Artificial Intelligence and Emerging Technologies to provide effective and continuous Parliamentary oversight of the adoption of these systems across the Australian government and more widely.
4. Any guidance issued by the Digital Transformation Agency, or any other Australian Government agency, should clearly define all AI systems and applications.

In addition to addressing the Inquiry’s terms of reference, the ADM+S submission led by Prof Kimberlee Weatherall included three other areas of research that raise important considerations around the use of AI in the public sector;Disability and accessibility, Environmental impact, and Trauma-informed approaches.

The 24 October submission reads, “The public sector should, in its use of AI, demonstrate the positive impacts that technology can have in achieving important public goals, such as promoting access, inclusion, and better public services.”

Author ADM+S Centre
Date 5 March 2025

Author Sam Whiting
Date 5 March 2025

Author Kevin Witzenberger and Michael Richardson
Date 3 March 2025

After a year of shoehorning generative AI into its flagship products, Microsoft is trying to recoup the costs by raising prices, putting ads in products, and cancelling data centre leases. Google is making similar moves, adding unavoidable AI features to its Workspace service while increasing prices.

Is the tide finally turning on investments into generative AI? The situation is not quite so simple. Tech companies are fully committed to the new technology – but are struggling to find ways to make people pay for it.

Shifting costs

Last week, Microsoft unceremoniously pulled back on some planned data centre leases. The move came after the company increased subscription prices for its flagship 365 software by up to 45%, and quietly released an ad-supported versionof some products.

The tech giant’s CEO, Satya Nadella, also recently suggested AI has so far not produced much value.

Microsoft’s actions may seem odd in the current wave of AI hype, coming amid splashy announcements such as OpenAI’s US$500 billion Stargate data centre project.

But if we look closely, nothing in Microsoft’s decisions indicates a retreat from AI itself. Rather, we are seeing a change in strategy to make AI profitable by shifting the cost in non-obvious ways onto consumers.

The cost of generative AI

Generative AI is expensive. OpenAI, the market leader with a claimed 400 million active monthly users, is burning money.

Last year, OpenAI brought in US$3.7 billion in revenue – but spent almost US$9 billion, for a net loss of around US$5 billion.

Microsoft is OpenAI’s biggest investor and currently provides the company with cloud computing services, so OpenAI’s spending also costs Microsoft.

What makes generative AI so expensive? Human labour aside, two costs are associated with AI models: training (building the model) and inference (using the model).

While training is an (often large) up-front expense, the costs of inference grow with the user base. And the bigger the model, the more it costs to run.

Smaller, cheaper alternatives

A single query on OpenAI’s most advanced models can cost up to US$1,000 in compute power alone. In January, OpenAI CEO Sam Altman said even the company’s US$200 per month subscription is not profitable. This signals the company is not only losing money through use of its free models, but through its subscription models as well.

Both training and inference typically take place in data centres. Costs are high because the chips needed to run them are expensive, but so too are electricity, cooling, and the depreciation of hardware.

To date, much AI progress has been achieved by using more of everything. OpenAI describes its latest upgrade as a “giant, expensive model”. However, there are now plenty of signs this scale-at-all-costs approach might not even be necessary.

Chinese company DeepSeek made waves earlier this year when it revealed it had built models comparable to OpenAI’s flagship products for a tiny fraction of the training cost. Likewise, researchers from Seattle’s Allen Institute for AI (Ai2) and Stanford University claim to have trained a model for as little as US$50.

In short, AI systems developed and delivered by tech giants might not be profitable. The costs of building and running data centres are a big reason why.

Author Natalie Campbell
Date 3 March 2025

ADM+S Affiliate Dr Christopher O’Neill, who recently completed a Research Fellowship with Prof Mark Andrejevic at the Monash University node of ADM+S has been awarded a 2025-2026 Fulbright Scholarship at the University of Southern California.

Commemorating the achievement at Parliament House in Canberra on 27 February, Dr O’Neill was presented his Fulbright Scholarship from the Governor General, Sam Mostyn.

Dr O’Neill will spend four months working with Assoc Prof Mike Ananny, Associate Professor of Communication and Journalism at the USC Annenberg School of Journalism, studying automation, work and error.

The Fulbright Program is the largest educational scholarship of its kind, created by US Senator J. William Fulbright and the US Government in 1946, and is the flagship foreign exchange scholarship program of the United States.

Successful Fulbright recipients are interviewed and selected by panels of experts from academia, government, professional organisations and the U.S. Embassy in a competitive process which assesses academic and professional merit, a strong program proposal with defined potential outcomes, and ambassadorial skills.

Dr O’Neill is currently a Research Fellow at the Alfred Deakin Institute, where his work draws upon science and technology studies and critical media theory to study the place of automation in contemporary biopower.

Prior to his role at Deakin, he spent three years as a Postdoctoral Research Fellow working with ADM+S Chief Investigator Prof Mark Andrejevic at the Monash University node of ADM+S, where among other projects, he developed a critical analysis of the role of the human in automated work and surveillance systems.

Author Natalie Campbell
Date 3 March 2025

The 2025 ADM+S Summer School, hosted by the University of Melbourne Law School, brought together over 120 students, researchers and mentors for a curated program spanning research methodologies, ethics advice, writing and publishing, and more.

Bringing together higher degree research students (HDRs) and early career researchers (ECRs) from all nine ADM+S nodes, the annual Summer School provides a perfect opportunity for community members to ask questions, share concerns, learn from one another, and get the most out of their research journey in the ADM+S community.

ADM+S Manager of Research Training and Development and member of the Summer School working group Sally Storey, said “This event would not be possible without the incredible generosity of our Centre’s research community.”

“I want to say a huge thank you to all our presenters and mentors for sharing your knowledge and expertise with our attendees, and the time leading up to the Summer School preparing presentations, materials, wrangling, scheduling… the effort is outstanding!”

The program encourages PhD students to engage with topics across disciplines, learn about different research methods, and create connections with peers and mentors from across the national ADM+S network – an invaluable experience for all early career researchers.

PhD Student Tace McNamara from Monash University explained, “I’m looking at AI and its capacity to understand art and music as an audience.

“It’s been really interesting talking to people from other disciplines because I think what I’m doing is inherently interdisciplinary, so hearing about law, media, culture, that’s something I don’t do on a daily basis in my lab, and it’s been really valuable.”

Sessions ranged from ‘Ethical uses of GenAI in research’, to ‘Unpacking ideas animating technology governance’, ‘Interviewing with digital trace data’, ‘How to study socio-technical networks’, ‘Harnessing technology for remote research, and more.

Author Natalie Campbell
Date 28 February 2025

The ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) is thrilled to congratulate members of the Professional Staff team, who have been recognised for their service excellence in the annual RMIT Research Service Awards.

The awards ceremony was held on 21 February 2025 at The Capitol Theatre in Melbourne, dedicated to celebrating the achievements of the RMIT research community and research support staff.

The Research Service Awards invited peers to nominate those in their community who demonstrate tremendous effort in supporting and delivery successful research outcomes.

ADM+S Chief Operating Officer Nicholas Walsh was awarded the Service Excellence award, which honours an individual who demonstrate excellence in research and innovation support.

Announcing the award, Jane Holt, Executive Director of RMIT’s Research Strategy & Services, highlighted Nick’s pioneering role as the first COO of RMIT’s inaugural centre of excellence, and commended his coordination and delivery of a remarkable ARC mid-term report working with ADM+S’ global partners.

“His dedication to meeting the centre’s needs and resolving challenges within RMIT’s enterprise systems demonstrates his outstanding service and leadership,” she said.

Author Natalie Campbell
Date 27 February 2025

On 25 February 2025, the University of Sydney unveiled its new Centre for AI, Trust and Governance (CAITG), appointing ADM+S Chief Investigator Prof Kimberlee Weatherall as co-director alongside Prof Terry Flew.

As co-director, Prof Weatherall will lead groundbreaking research to ensure AI is transparent, fair, and accountable, championing the critical role of law and policy in shaping ethical AI.

“Universities have a critical role to play in ensuring that AI develops for the benefit of everyone, all the way across society,” says Prof Weatherall.

 “I’m proud to be co-directing CAITG that can bring together the University of Sydney’s outstanding researchers and students, from different research disciplines, to understand how the technology is developing, its impacts in the world and how to shape it for the better.”

CAITG’s research agenda is focused on AI’s relationship to digital creative industries platforms and information, law and policy, education and equity, organisations and work, and civic technology and participation.

Some of the themes being investigated include:

• how to restore trust in social institutions, and whether AI presents new threats to trust and social cohesion
• how laws and regulations need to change in order to ensure that AI systems serve the public interest
• how the community can be better involved in decisions about the uses of AI in secondary and tertiary education
• foreign actors using AI to undermine democracy in Australia and in the Asia-Pacific region.

Author Daniel Angus and Mark Andrejevic
Date 25 February 2025

The rise of ‘dark advertising’ — personalised advertisements increasingly powered by artificial intelligence that evade public scrutiny — means Australians face a murky information landscape going into the federal election.

It’s already happening and, combined with Australia’s failure to enact truth-in-advertising legislation and big tech’s backtracking on fact-checking, means voters are left vulnerable to ad-powered misinformation campaigns. And that’s not good for democracy.

Tackling misinformation requires legislative action, international collaboration and continued pressure on platforms to open their systems to scrutiny.

The failures of US tech platforms during their own elections should serve as a clear warning to Australia that industry self-regulation is not an option.

Political advertising plays a pivotal role in shaping elections, even while it is shrouded in opacity and increasing misinformation.

In the lead-up to the 2025 federal election, a significant volume of deceptive advertising and digital content has already surfaced. That’s not surprising, given the Australian Electoral Commission (AEC) limits its oversight to the official campaign period, meaning false claims can proliferate freely before the official campaign.

At the heart of this challenge lies the evolution of digital political advertising.

What is ‘dark advertising’?
Modern campaigns rely heavily on social media platforms, leveraging associative ad models that tap into beliefs or interests to deliver digital advertising. Unlike traditional media, where ads are visible and subject to better regulatory and market scrutiny, digital ads are often fleeting and hidden from public view.

Recent AI developments make it easier and cheaper to create false and misleading political ads in large volumes with multiple variations increasingly difficult to detect.

This ‘dark advertising’ creates information asymmetries, in this case one where groups have access to information and can control and shape how it’s delivered. That leaves voters exposed to tailored messages that may distort reality.

Targeted messaging makes it possible to selectively provide voters with very different views of the same candidate. In the recent US presidential election, a political action committee linked to X owner Elon Musk targeted Arab-American voters with the message that Kamala Harris was a diehard Israel ally, while simultaneously messaging Jewish voters that she was an avid supporter of Palestine.

Ad targeting online also lets political advertisers single out groups more likely to be influenced by selective, misleading or false information. Conservative lobby group Advance Australia’s recent campaign basically followed this playbook, disseminating outdated news articles on Facebook, a tactic known as malinformation, where factual information is deliberately spread misleadingly to harm individuals or groups.

The vulnerabilities
The Albanese government recently withdrew a proposed truth-in-political-advertising bill, leaving voters vulnerable to misleading content that undermines democratic integrity.

The bill was never introduced to parliament and its future remains uncertain.

The transparency tools provided by Meta, which covers Facebook and Instagram, and Google parent company Alphabet — which include ad libraries and “Why Am I Seeing This Ad?” explanations — also fall woefully short of enabling meaningful oversight.

These tools reveal little about the algorithms that determine ad delivery or the audiences being targeted. They do include some demographic breakdowns, but say little about the combination of ads an individual user might have seen and in what context.

Recent findings from the US highlight the vulnerabilities of political advertising in the digital age. An investigation by ProPublica and the Tow Center for Digital Journalism revealed that deceptive political ads thrived on platforms like Facebook and Instagram in the lead-up to the 2024 US elections.

Ads frequently employed AI-generated content, including fabricated audio of political figures, to mislead users and harvest personal information. One ad account network has run about 100,000 misleading ads, significantly exploiting Meta’s advertising systems.

Author ADM+S Centre
Date 19 February 2025

A new industry report has found audiences and journalists are growing increasingly concerned by generative artificial intelligence (AI) in journalism.

Summarising three years of research, the Generative AI & Journalism report was launched at the ARC Centre of Excellence for Automated Decision-Making and Society this week.

Report lead author, Dr T.J. Thomson from RMIT University in Melbourne, Australia, said the potential of AI-generated or edited content to mislead or deceive was of most concern.

“The concern of AI being used to spread misleading or deceptive content topped the list of challenges for both journalists and news audiences,” he said.

“We found journalists are poorly equipped to identify AI-generated or edited content, leaving them open to unknowingly propelling this content to their audiences.”

This is partly because few newsrooms have systematic processes in place for vetting user-generated or community contributed visual material.

Most journalists interviewed were not aware of the extent to which AI is increasingly and often invisibly being integrated into both cameras and image or video editing and processing software.

“AI is sometimes being used without the journalists or news outlet even knowing,” Thompson said.

While only one quarter of news audiences surveyed thought they had encountered generative AI in journalism, about half were unsure or suspected they had.

“This points to a potential lack of transparency from news organisations when they use generative AI or to a lack of trust between news outlets and audiences,” Thomson said.

News audiences were found to be more comfortable with journalists using AI when they themselves have used it for similar purposes, such as to blur parts of an image.

“The people we interviewed mentioned how they used similar tools when on video conferencing apps or when using the portrait mode on smartphones,” Thomson said.

“We also found this with journalists using AI to add keywords to media since audiences had themselves experienced AI describing images in word processing software.”

Thomson said news audiences and journalists alike were overall concerned about how news organisations are – and could be – using generative AI.

“Most of our participants were comfortable with turning to AI to create icons for an infographic but quite uncomfortable with the idea of an AI avatar presenting the news, for example,” he said.

Author Ash Watson
Date 19 February 2025

When we’re uncomfortable we say the “vibe is off”. When we’re having a good time we’re “vibing”. To assess the mood we do a “vibe check”. And when the atmosphere in the room changes we call it a “vibe shift”.

In a broad sense, a “vibe” is something akin to a mood, atmosphere or energy.

But this is an imperfect definition. Often, we’ll use this term to describe something we feel powerfully, but find hard to articulate.

As journalist and cultural critic Kyle Chayka described in 2021, a vibe is “a placeholder for an unplaceable feeling or impression, an atmosphere that you couldn’t or didn’t want to put into words”.

Being able to understand the subtleties of social interactions – that is, to “feel the vibes” – is extremely valuable, not just for our social interactions, but also for researchers who study people.

What’s behind the rise of vibes? And how can sociologists like myself unpack “vibe culture” to make sense of the world?

A history of vibes

The nuance and complexity of vibes makes them an interesting cultural trend. Vibes can be very specific, but can also totally resist specificity.

Australians (and fans of Australiana) will remember the iconic line from the beloved 1997 film The Castle: “It’s just the vibe of the thing… I rest my case.”

While it may seem like a recent cultural development, vibe isn’t the first example of cryptic language being used to express an ambiguous thing or situation. There are similar concepts with long histories, such as “quintessence” in Ancient Greek philosophy and “auras” in mysticism.

More recently, vibes rose in popularity through music including 1960s rock, epitomised by the Beach Boys (“pickin’ up good vibrations”) and Black American rap vernacular from the 1990s, such as in the song Vibes and Stuff by A Tribe Called Quest (“we got, we got, we got the vibes”).

While we don’t know when the term was first used as it is today, it seems to have taken hold in the 1970s.

I trawled the online archive of The New Yorker and found an early mention of vibes in a 1971 report about communes in New York City.

One interviewee spoke about the “vibration of togetherness” that drew them to the commune. Ending the day on the subway, the author Hendrik Hertzberg (now a senior editor at the magazine) “just sat there and soaked up the good vibes”.

New uses and meanings have emerged in the years since.

Author ADM+S Centre
Date 19 February 2025

Generative artificial intelligence (AI) has taken off at lightning speed in the past couple of years, creating disruption in many industries. Newsrooms are no exception.

A new report published today finds that news audiences and journalists alike are concerned about how news organisations are – and could be – using generative AI such as chatbots, image, audio and video generators, and similar tools.

The report draws on three years of interviews and focus group research into generative AI and journalism in Australia and six other countries (United States, United Kingdom, Norway, Switzerland, Germany and France).

Only 25% of our news audience participants were confident they had encountered generative AI in journalism. About 50% were unsure or suspected they had.

This suggests a potential lack of transparency from news organisations when they use generative AI. It could also reflect a lack of trust between news outlets and audiences.

Who or what makes your news – and how – matters for a host of reasons.

Some outlets tend to use more or fewer sources, for example. Or use certain kinds of sources – such as politicians or experts – more than others.

Some outlets under-represent or misrepresent parts of the community. This is sometimes because the news outlet’s staff themselves aren’t representative of their audience.

Carelessly using AI to produce or edit journalism can reproduce some of these inequalities.

Our report identifies dozens of ways journalists and news organisations can use generative AI. It also summarises how comfortable news audiences are with each.

The news audiences we spoke to overall felt most comfortable with journalists using AI for behind-the-scenes tasks rather than for editing and creating. These include using AI to transcribe an interview or to provide ideas on how to cover a topic.

But comfort is highly dependent on context. Audiences were quite comfortable with some editing and creating tasks when the perceived risks were lower.

The problem – and opportunity

Generative AI can be used in just about every part of journalism.

For example, a photographer could cover an event. Then, a generative AI tool could select what it “thinks” are the best images, edit the images to optimise them, and add keywords to each.

These might seem like relatively harmless applications. But what if the AI identifies something or someone incorrectly, and these keywords lead to mis-identifications in the photo captions? What if the criteria humans think make “good” images are different to what a computer might think? These criteria may also change over time or in different contexts.

Even something as simple as lightening or darkening an image can cause a furore when politics are involved.

AI can also make things up completely. Images can appear photorealistic but show things that never happened. Videos can be entirely generated with AI, or edited with AI to change their context.

Generative AI is also frequently used for writing headlines or summarising articles. These sound like helpful applications for time-poor individuals, but some news outlets are using AI to rip off others’ content.

AI-generated news alerts have also gotten the facts wrong. As an example, Apple recently suspended its automatically generated news notification feature. It did this after the feature falsely claimed US murder suspect Luigi Mangione had killed himself, with the source attributed as the BBC.

What do people think about journalists using AI?

Our research found news audiences seem to be more comfortable with journalists using AI for certain tasks when they themselves have used it for similar purposes.

For example, the people interviewed were largely comfortable with journalists using AI to blur parts of an image. Our participants said they used similar tools on video conferencing apps or when using the “portrait” mode on smartphones.

Author Natalie Campbell
Date 17 February 2025

The House of Representatives Standing Committee on Employment, Education and Training has published The Future of Work: Inquiry into the Digital Transformation of Workplaces, following their Inquiry into the Digital Transformation of Workplaces, citing contributions from ADM+S Affiliate Emmanuelle Walkowiak’s 19 June submission.

The inquiry found that Imminent support is required for employers, workers, students, and regulators, and that Australia needs to increase investment in research and development to ensure the safe, responsible and effective use of ADM and AI in the workplace.

The report explains, “digital transformation has exposed significant risks, including gaps in Australia’s regulatory frameworks and workplace protections. This is especially the case with data and privacy.”

A Vice-Chancellor Senior Research Fellow in Economics at RMIT’s Blockchain Innovation Hub, Dr Walkowiak’s research primarily focuses on technology driven inclusion at work, and the changing nature of work in a digital economy.

Her submission to the Inquiry outlined evidence-based recommendations on harnessing AI for productivity, skill development, and job creation in Australia while addressing risks like impacts on hiring, job design, and work quality. It explored AI’s effect on labour rights, fairness, and dignity at work, as well as its influence on small businesses and vulnerable groups, including neurodiverse workers.

Dr Walkowiak’s submission is cited in the report’s discussion of Regulating Technology: Public views (p.19), Opportunities in productivity and efficacy (p.28), and Data and Privacy: Disclosure and breach of privacy (p.49).

Author Suvradip Maitra, Lyndal Sleep, Paul Henman, Suzana Fay
Date 10 February 2025

Late last year, ChatGPT was used by a Victorian child protection worker to draft documents. In a glaring error, ChatGPT referred to a “doll” used for sexual purposes as an “age-appropriate toy”. Following this, the Victorian information commissioner banned the use of generative artificial intelligence (AI) in child protection.

Unfortunately, many harmful AI systems will not garner such public visibility. It’s crucial that people who use social services – such as employment, homelessness or domestic violence services – are aware they may be subject to AI. Additionally, service providers should be well informed about how to use AI safely.

Fortunately, emerging regulations and tools, such as our trauma-informed AI toolkit, can help to reduce AI harm.

How do social services use AI?

AI has captured global attention with promises of better service delivery. In a strained social services sector, AI promises to reduce backlogs, lower administrative burdens and allocate resources more effectively while enhancing services. It’s no surprise a range of social service providers are using AI in various ways.

Chatbots simulate human conversation with the use of voice, text or images. These programs are increasingly used for a range of tasks. For instance, they can provide mental health support or offer employment advice. They can also speed up data processing or help quickly create reports.

However, chatbots can easily produce harmful or inaccurate responses. For instance, the United States National Eating Disorders Association deployed the chatbot Tessa to support clients experiencing eating disorders. But it was quickly pulled offline when advocates flagged Tessa was providing harmful weight loss advice.

Recommender systems use AI to make personalised suggestions or options. These could include targeting job or rental ads, or educational material based on data available to service providers.

But recommender systems can be discriminatory, such as when LinkedIn showed more job ads to men than women. They can also reinforce existing anxieties. For instance, pregnant women have been recommended alarming pregnancy videos on social media.

Recognition systems classify data such as images or text to compare one dataset to another. These systems can complete many tasks, such as face matching to verify identity or transcribing voice to text.

Such systems can raise surveillance, privacy, inaccuracy and discriminationconcerns. A homeless shelter in Canada stopped using facial recognition cameras because they risked privacy breaches – it’s difficult to obtain informed consent from mentally unwell or intoxicated people using the shelter.

Risk-assessment systems use AI to predict the likelihood of a specific outcome occurring. Many systems have been used to calculate the risk of child abuse, long-term unemployment, or tax and welfare fraud.

Often data used in these systems can recreate societal inequalities, causing harm to already-marginalised peoples. In one such case, a tool in the US used for identifying risk of child mistreatment unfairly targeted poor, black and biracial families and families with disabilities.

A Dutch risk assessment tool seeking to identify childcare benefits fraud was shut down for being racist, while an AI system in France faces similar accusations.

Author ADM+S Centre
Date 10 February 2025

In a new submission to the Attorney-General Department’s Automated Decision-Making (ADM) Reform consultation, experts from the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) urge the government to adopt a more comprehensive regulatory framework for ADM in the public sector. 

The submission argues that the current focus of the consultation paper on legislation and regulation overlooks essential aspects like enforcement and accountability, which are critical to ensuring responsible use of technology in government decision-making.

The response highlights that the existing approach is too narrow, focusing primarily on AI-based systems, while neglecting broader systemic issues. The authors contend that the government should lead by example, setting a standard for safe and accountable technology use that applies to all technical systems, not just AI.

Among the key recommendations outlined in the submission are calls for stronger enforcement mechanisms, including active monitoring and independent oversight. The experts also emphasize the need for transparency in the acquisition of ADM systems, urging the government to adopt robust measures to prevent misuse and ensure accountability across all public sector applications of automated decision-making.

Author Fan Yang and Sukhmani Khorana
Date 6 February 2025

Migrants in Australia often encounter disinformation targeting their communities. However, disinformation circulated in non-English languages and within private chat groups often falls beyond the reach of Australian public agencies, national media and platform algorithms.

This regulatory gap means migrant communities are disproportionately targetedduring crises, elections and referendums when misinformation and disinformation are amplified.

With a federal election just around the corner, we wanted to understand how migrants come across disinformation, how they respond to it, and importantly, what can be done to help.

Our research

Our research finds political disinformation circulates both online and in person among friends and family.

Between 2023 and 2024, we carried out a survey with 192 respondents. We then conducted seven focus groups with 14 participants who identify as having Chinese or South Asian cultural heritage.

We wanted to understand their experiences of political engagement and media consumption in Australia.

An important challenge faced by research participants is online disinformation. This issue was already long-standing and inadequately addressed by Australian public agencies and technology companies, even before Meta ended its fact-checking program.

Lack of diversity in news

Our study finds participants read news and information from a diverse array of traditional and digital media services with heightened sense of caution.

They encounter disinformation in two ways.

The first is information misrepresenting their identity, culture, and countries of origin, particularly found in English-language Australian national media.

The second is targeted disinformation distributed across non-English social media services, including in private social media channels.

From zero (no trust) to five (most trusted), we asked our survey participants to rank their trust towards Australian national media sources. This included the ABC, SBS, The Age, Sydney Morning Herald, 9 News and the 7 Network.

Participants reported a medium level of trust (three).

Our focus groups explained the mistrust participants have towards both traditional and social media news sources. Their thoughts echoed other research with migrants. For instance, a second-generation South Asian migrant said:

it feels like a lot of marketing with traditional media […] they use marketing language to persuade people in a certain way.

Several participants of Chinese and South Asian cultural backgrounds reported that Australian national media misrepresent their culture and identity due to a lack of genuine diversity within news organisations. One said:

the moment you’re a person of colour, everyone thinks that you’re Chinese. And we do get painted with the same paintbrush. It is very frustrating […]

Another added:

Sri Lanka usually gets in the media for cricket mainly, travel and tourism. So apart from that, there’s not a lot of deep insight.

For migrants, the lack of genuine engagement with their communities and countries of origin distorts public understanding, reducing migrants to a one-dimensional, often stereotypical, portrayal. This oversimplification undermines migrants’ trust in Australian national media.

Author Haiqing Yu
Date 5 February 2025

DeepSeek has emerged on the front line of debates determining the future of AI, but its arrival poses questions over who decides what ‘intelligence’ we need.

Chinese company DeepSeek stands at the crossroads of two major battles shaping artificial intelligence development: whether source code should be freely available and whether development should happen in free or controlled-information environments.

That also highlights the DeepSeek paradox. It champions open-source AI — where the source code of the underlying model is available for others to use or modify — while operating in China, one of the world’s most-controlled data environments.

That means DeepSeek prompts obvious questions about who decides what kind of ‘intelligence’ we need. Such questions are obviously front of mind for some governments, with several already placing restrictions on the use of DeepSeek.

DeepSeek, a Chinese startup, unveiled its AI chatbot late last month. It seemed to equal the performance of US models at a fraction of the cost and the news triggered a massive sell-off of tech company shares on the US sharemarket.

It also sparked concerns about data security and censorship. In Australia, DeepSeek has been banned from all federal government devices, the NSW government has reportedly banned it from its devices and systems and other state governments are considering their options. The Australian ban followed similar action by Taiwan, Italy and some US government agencies.

The Australian government says the bans are not related to DeepSeek’s country of origin, but the issues being raised now are similar to those discussed when Chinese-based social media app TikTok was banned on Australian government devices two years ago.

Yet aside from those concerns and DeepSeek’s role in reshaping the power dynamics in the US-China AI rivalry, it also gives hope to less well-resourced countries to develop their own large language models using DeepSeek’s model as a starting point.

For those seeking Chinese-related pop culture references, DeepSeek is a Monkey King moment in the global AI landscape.
Monkey King, or Wukong in Chinese, was a character featured in the 16th century novel Journey to the West.

The story was popularised in the 1980s television series Monkey and later iterations. In these stories, Wukong was the unpredictable force challenging established power, wreaking havoc in the Heavenly Palace and embodying both defiance and restraint.

That’s a pretty apt description for where DeepSeek stands in the AI world in 2025.

A new benchmark
As the author of a recent Forbes piece rightly points out, the real story about DeepSeek is not about geopolitics but “about the growing power of open-source AI and how it’s upending the traditional dominance of closed-source models”.

Author Kolawole Samuel Adebayo says it’s a line of thought that Meta chief AI scientist Yann LeCun also shares.

The AI industry has long been divided between closed-source titans like OpenAI, Google, Amazon, Microsoft and Baidu and the open-source movement, which includes Meta, Stability AI, Mosaic ML as well as universities and research institutes.

DeepSeek’s adoption of open-source methodologies — building on Meta’s open-source Llama models and the PyTorch ecosystem — places it firmly in the open-source camp.

While closed-source large language models prioritise controlled innovation, open-source large language models are built on the principles of collaborative innovation, sharing and transparency.

DeepSeek’s innovative methods challenge the notion that AI development is backed by vast proprietary datasets and computational power, measured by the number and capacity of chips.

It also demonstrates a point made by the Australian Institute for Machine Learning’s Deval Shah three months before DeepSeek made global headlines: “The future of LLM [large language model] scaling may lie not just in larger models or more training data, but in more sophisticated approaches to training and inference.”

The DeepSeek case illustrates that algorithmic ingenuity can compensate for hardware and computing limitations, which is significant in the context of US export controls on high-end AI chips to China. That’s a crucial lesson for any nation or company restricted by computational bottlenecks.

It suggests that an alternative path exists — one where innovation is driven by smarter algorithms rather than sheer hardware dominance.

Author Data and Society 
Date 30 January 2025

Data & Society (D&S) today announced the launch of its Climate, Technology, and Justice program. Climate change is perhaps the most urgent social issue of our time and is only accelerating in importance. Already disproportionately impacting communities in the majority world, energy-intensive technologies like artificial intelligence only worsen the problem. Data & Society has spent a decade building an empirical research base on data-driven technologies, and fostering a network that is influencing how these technologies are studied and governed. The organization is uniquely well-positioned to examine the social and environmental repercussions of the expanded global infrastructures and labor practices needed to sustain the growth of digital technologies, from AI and blockchain to streaming and data storage.

The new program will be led by Tamara Kneese, who joined D&S in 2023 as senior researcher and project director of the Algorithmic Impact Methods Lab (AIMLab), and whose experience in human-centered technology and climate activism in the tech industry make her an ideal leader for this work. Joining her are two affiliates: Zane Griffin Talley Cooper, who studies data, resource extraction, and the Arctic; and Xiaowei R. Wang, whose body of multidisciplinary work, over the past 15 years, sits at the intersection of tech, digital media, art, and environmental justice.

Succeeding Kneese as AIMLab project director is D&S Senior Researcher Meg Young, whose leadership of the Lab’s participatory efforts and impact engagement since its 2023 launch have been key to its early successes. A champion for participatory methods in the AI impact space and for making technology more accountable to the public, Young’s work with communities across the country has positioned AIMLab for the future.

Before joining D&S, Kneese was lead researcher at Green Software Foundation (GSF), where she was part of the policy working group and the author of GSF’s first State of Green Software Report, which provided insight into the people and planet impacts of AI. Earlier, she was director of developer engagement on the green software team at Intel and assistant professor of media studies and director of gender and sexualities studies at the University of San Francisco. She and Young recently co-authored ”Carbon Emissions in the Tailpipe of Generative AI” in the Harvard Data Science Review, offering an overview of the current state of measuring, regulating, and mitigating AI’s environmental impacts and underscoring that the real existential threat posed by AI is its impact on climate.

“While this program will first tackle the environmental impacts of AI, we have expansive visions of how D&S’s considerable skillset can help us understand the complex relationships between climate, the environment, climate change, technology, and justice — areas like e-waste and tech reuse, algorithmic disaster prediction, and low-carbon tech adoption, centering the experiences and voices of the communities most affected,” said Alice E. Marwick, Data & Society’s director of research. “I am thrilled about the new body of scholarship that we will develop under Tamara’s leadership.”

Author ADM+S Centre
Date 30 January 2025

A newly released evaluation report highlights the success of the Victorian Women’s Trust’s Rural Women Online program in addressing digital exclusion among women in regional Victoria.

Involving hands-on digital skills workshops on a range of topics, a help desk for one-on-one support, stands from local services providers, and keynotes from leading thinkers and writers on digital inclusion in Australia, Rural Women Online was delivered in August and September 2024 in Greater Shepparton and North East Victoria following extensive community consultation. Hundreds of women from across regional Victoria participated in the program, gaining new skills, confidence, and forging new social connections and opportunities.

Key outcomes include

1. Boosting confidence: The program saw significant increases in participants’ confidence with digital technologies, with 43% reporting they felt more capable using digital tools and navigating online platforms. Workshops focused on practical skills like managing passwords, identifying scams, and safely using online services, helping participants overcome fears and avoid common pitfalls.
2. Tailored support: Over half of participants sought personalised assistance from local mentors at the program’s help desks. Mentors were not necessarily ‘tech experts’ but were relatable and were happy to learn alongside participants, setting up a space for mutual empowerment for a range of often highly personal tasks.
3. Strengths-based learning: By focusing on participants’ existing capabilities and reframing digital challenges as opportunities for empowerment, the program created a supportive environment. This approach empowered women to see themselves as capable digital users, shifting the narrative from vulnerability to resilience.
4. Social connection: The program fostered a sense of community among participants, enabling them to share experiences and build networks for ongoing support. Informal workshops and “chat corners” encouraged open dialogue and connection, reducing feelings of isolation and promoting collaborative learning.

Rural women are among the most digitally excluded groups in Victoria, facing barriers like limited access to technology, low digital confidence, and a lack of locally relevant resources.

Author ADM+S Centre
Date 28 January 2025

On 20 January 2025 U.S. President Donald Trump revoked a 2023 executive order aimed at regulating artificial intelligence (AI), prioritising innovation over regulation of the rapidly advancing technology.

The executive order previously signed by former President Joe Biden, required AI companies to submit safety testing data to federal authorities, to establish safety standards around AI development.

ADM+S Affiliate and Director of the Centre for AI and Digital Ethics at the University of Melbourne Prof Jeannie Paterson joined ABC radio last week to discuss some of the implications of this decision.

“It’s definitely a statement that the guardrails have come off for the development on AI,” she explains.

“The Executive order said that anybody who was releasing AI to be used with government had to put in place safeguards to prevent bias, to protect privacy, to reduce error, and to keep it cyber secure.

Those requirements aren’t there anymore so it’s hard to say what sort of safety measures will be reduced.”

Author Mark Andrejevic
Date 20 January 2025

In the wake of Donald Trump’s election victory, Meta chief executive Mark Zuckerberg fired the fact-checking team for his company’s social media platforms. At the same time, he reversed Facebook’s turn away from political content.

The decision is widely viewed as placating an incoming president with a known penchant for mangling the truth.

Meta will replace its fact-checkers with the “community notes” model used by X, the platform owned by avid Trump supporter Elon Musk. This model relies on users to add corrections to false or misleading posts.

Musk has described this model as “citizen journalism, where you hear from the people. It’s by the people, for the people.”

For such an approach to work, both citizen journalists and their readers need to value good-faith deliberation, accuracy and accountability. But our new research shows social media users may not be the best crowd to source in this regard.

Our research

Working with Essential Media, our team wanted to know what social media users think of common civic values.

After reviewing existing research on social cohesion and political polarisation and conducting ten focus groups, we compiled a civic values scale. It aims to measure levels of trust in media institutions and the government, as well as people’s openness to considering perspectives that challenge their own.

We then conducted a large-scale survey of 2,046 Australians. We asked people how strongly they believed in a common public interest. We also asked about how important they thought it was for Australians to inform themselves about political issues and for schools to teach civics.

Importantly, we asked them where they got their news: social media, commercial television, commercial radio, newspapers or non-commercial media.

What did we find?

We found people who rely on social media for news score significantly lower on a civic values scale than those who rely on newspapers and non-commercial broadcasters such as the ABC.

By contrast, people who rely on non-commercial radio scored highest on the civic values scale. They scored 11% higher than those who rely mainly on social media and 12% higher than those who rely on commercial television.

The lowest score was for people who rely primarily on commercial radio.

Author Natalie Campbell
Date 17 January 2025

Congratulations to ADM+S Partner Investigator Assoc Prof Julia Stoyanovich from New York University, a 2025 recipient of a Presidential Early Career Award for Scientists and Engineers in the United States.

Assoc Prof Stoyanovich is amongst nearly 400 scientists and engineers who received the award on 15 January, the highest honour bestowed by the U.S. government on outstanding scientists and engineers early in their careers.

“I am immensely grateful to my mentors and long-term collaborators, students, and postdocs for making this possible.

And I am thrilled to be able to call New York University my home, where all doors are open and the sky is the limit,” said Assoc Prof Stoyanovich.

Announced via the White House website, the media release reads, “Established by President Clinton in 1996, PECASE recognizes scientists and engineers who show exceptional potential for leadership early in their research careers.

Authors  Ned Watt, Michelle Riedlinger and Silvia Montaña-Niño
Date 8 January 2025

Meta has announced it will abandon its fact-checking program, starting in the United States. It was aimed at preventing the spread of online lies among more than 3 billion people who use Meta’s social media platforms, including Facebook, Instagram and Threads.

In a video, the company’s chief, Mark Zuckerberg, said fact checking had led to “too much censorship”.

He added it was time for Meta “to get back to our roots around free expression”, especially following the recent presidential election in the US. Zuckerberg characterised it as a “cultural tipping point, towards once again prioritising speech”.

Instead of relying on professional fact checkers to moderate content, the tech giant will now adopt a “community notes” model, similar to the one used by X.

This model relies on other social media users to add context or caveats to a post. It is currently under investigation by the European Union for its effectiveness.

This dramatic shift by Meta does not bode well for the fight against the spread of misinformation and disinformation online.

Independent assessment

Meta launched its independent, third-party, fact-checking program in 2016.

It did so during a period of heightened concern about information integrity coinciding with the election of Donald Trump as US president and furore about the role of social media platforms in spreading misinformation and disinformation.

As part of the program, Meta funded fact-checking partners – such as Reuters Fact Check, Australian Associated Press, Agence France-Presse and PolitiFact – to independently assess the validity of problematic content posted on its platforms.

Warning labels were then attached to any content deemed to be inaccurate or misleading. This helped users to be better informed about the content they were seeing online.

A backbone to global efforts to fight misinformation

Zuckerberg claimed Meta’s fact-checking program did not successfully address misinformation on the company’s platforms, stifled free speech and lead to widespread censorship.

But the head of the International Fact-Checking Network, Angie Drobnic Holan, disputes this. In a statement reacting to Meta’s decision, she said:

Fact-checking journalism has never censored or removed posts; it’s added information and context to controversial claims, and it’s debunked hoax content and conspiracy theories. The fact-checkers used by Meta follow a Code of Principles requiring nonpartisanship and transparency.

A large body of evidence supports Holan’s position.

In 2023 in Australia alone, Meta displayed warnings on over 9.2 million distinct pieces of content on Facebook (posts, images and videos), and over 510,000 posts on Instagram, including reshares. These warnings were based on articles written by Meta’s third-party, fact-checking partners.

Numerous studies have demonstrated that these kinds of warnings effectively slow the spread of misinformation.