chat.openai.com, or on the iOS or Android app stores: A proprietary dialog agent that is amusing and sometimes helpful. The language models underlying ChatGPT are heavily entwined with multiple layers of proprietary fine-tuning, guard-rails and content filters.

• Tech details: Unknown model size and architecture, with proprietary training data and processes.
• Access: Requires creating a free OpenAI account to access the GPT-3.5 model. Paid accounts can also access the more advanced GPT-4 model.
• Data Privacy: By default, all manual user conversations are saved for use by OpenAI. This can be disabled in settings. API interactions are not saved. See this FAQ page for more information.
• Affordances: Paid accounts can enable internet access and integration with third party services via plugins.
• User controls: Can bookmark, share, and delete prior conversations. Bulk data export and deletion controls are also available in the settings.
• Computational access: Via the OpenAI Chat Completions API (requires $), (or via lots many other 3rd party interfaces).

bing.com (click on ‘Chat’ tab): A proprietary natural language chat interface for the Bing search engine. Bing chat is generally able to provide accurate citations for information it provides.

• Tech details: Based on a customised version of one of OpenAI’s GPT models
• Access: Requires using Microsoft’s Edge browser, and creating a free microsoft account. Will not work if an Edge private browsing session is used, or if Bing’s ‘Safe Search’ setting is set to ‘strict’.
• Data Privacy: Microsoft saves all conversations for later use as per the terms of use.
• Affordances: The dialog agent can browse the Internet
• User controls: Bookmarking or sharing individual conversations is not possible. Can choose from ‘creative’, ‘balanced’ or ‘precise’ conversation styles.
• Computational access: None

ai.meta.com/llama: A second generation of the Llama language model from Meta. Llama and Llama 2 are semi-open-source, but many details of the model’s construction and training are proprietary. There are both basic LLM and fine-tuned dialog agent versions of LLama and Llama 2 available. The fine-tuned versions have undergone additional training to discourage toxic outputs and limit the topics they will discuss.

• Tech details: transformer models with either 7, 13, or 70 billion parameters. Training data and processes are largely opaque, but some details are provided in the associated report.
• Access: The 13b parameter Llama 2 dialog agent can be freely accessed at the LMSYS ChatArena – chat.lmsys.org without a user account (as well as a range of other models).
• Data privacy: LMSYS ChatArena saves all conversations for research purposes.
• Affordances: Text input and output only.
• User controls: LMSYS ChatArena allows users to control generation hyper-parameters, and also lets users compare different models side-by-side.
• Computational access: The Llama and Llama 2 models can be used programmatically using the HuggingFace libraries, but this requires pre-approval from Meta. See this page for more information.

• Hugging Face has an online database of user-created ‘spaces’ where you can interact with many different models. E.g. the Llama 2 70b chat model can be found here. Many spaces offer free compute (your prompts go into a queue and you have to wait for HuggingFace’s servers to process them and generate the response), while some spaces offer faster interaction because the creator has paid to rent a GPU server from HuggingFace.
• GPT4All is a desktop app for Windows, Mac and Linux that lets users interact with a range of dialog agents on their own computer. This works even without a GPU, but can be slow. This also requires downloading the model you want to use, which can take a while.

Gender bias is a preference or prejudice toward one gender over another. This type of bias is prevalent in modern society (for instance, recall the #GamerGate harassment campaign in 2014, or the more recent Depp v. Heard trial) and has deep seated historical roots, but is particularly relevant when dealing with text-based generative language models due to the interconnectedness of gender and language (even more so in languages that are more strongly gendered than English – e.g. German).,

There has been a lot of work in evaluating, critiquing, and removing gender bias in the area of machine learning for Natural Language Processing (for example, see the 2019 review by Sun et al., or the 2020 paper by Leavy et al.) – and even meta-critiques of gender bias in AI literature. The relevance of these historical investigations to the rapidly emerging domain of generative language models and Generative AI more generally is not yet clear though. Some of the ADM+S projects in this space include the Toxicity Scalpel project – which aims to measure and mitigate sexist tendencies in language models, Lucinda Nelson’s PhD project – which is investigating ‘every day’ misogyny in social media platforms, and Dr. Ariadna Matamors-Fernández’s DECRA project, which investigates racist and sexist humour in online spaces.

Resources

• StereoSet dataset and benchmark: used to measure stereotype bias in language models across gender, race, religion, and profession. It consists of 17,000 sentences and can be used to measure model race preferences.
• The Sentence Encoder Attribution Test (SEAT) dataset and benchmark: A collection of gender (and other) stereotyped text samples for testing for gender association biases in language models.
• The Ambivalent Sexism Inventory: A 22-question survey developed by psychologists Glick and Fisk to map sexist tendencies. See also more recent media coverage of this theory and survey by PBS.
• Reports from 2015, 2017, and 2018 mapping sexism and workplace gender inclusivity, both internationally and in the Australian context.
• Guidance document from Western Sydney University for gender inclusive language in surveys and questionnaires.

Political bias is a well-observed phenomenon in offline and online social interactions. For instance, our political leaning can impact how we perceive the trustworthiness of online news and how our social media news feeds are curated. Large language models can also exhibit political bias in a range of text-generation tasks. Some approaches for testing political bias include prompting language models with existing political statements or using validated questionnaires such as the political compass test.

Some recent investigations have revealed certain political biases of popular large language models. A pre-print article (i.e., not peer-reviewed) by Hartmann et al. (2023) reports an experiment where they prompted ChatGPT with 630 political statements from two leading voting advice applications and the nation-agnostic political compass test. The results show that ChatGPT is ‘pro-environmental’ and follows what the authors describe as, ‘a left-libertarian ideology.’ Similar political biases towards progressive and libertarian views were also observed in ChatGPT by Rutinowski et al. (2023) (pre-print) when prompted with the political compass test.

Resources

• The Political Compass – A test that profiles political personalities applicable to all democracies.
• 2023 Political Quiz – A political test that includes country specific tests presented in different languages.
• The BOLD dataset and benchmark is a recently published dataset of text prompts for testing for five types of biases in language models, including political bias.
• Vote Compass – a survey updated and re-released by ABC each federal election. This questionnaire maps political bias across two primary dimensions: Social (from progressive to conservative) and Economic (from left to right leaning).

Indigenous, racial and colonial biases refer to distinct, yet connected types of prejudice which impact the way individuals are treated or perceived. Racial bias discriminates based on an individual’s race or ethnicity. Indigenous bias is specifically directed at Indigenous people and it stems from negative attitudes towards indigenous cultures, languages, and ways of life. Simultaneously, colonial bias can manifest as the overrepresentation of perspectives, values, and languages of colonial powers, often at the expense of those from colonised or formerly colonised societies, leading to further suppression of already marginalised indigenous perspectives.

Australia’s history in particular is marked by a complex interplay of indigenous cultures and the impacts of colonialism. Unfortunately, this history has also given rise to deep-seated biases and prejudices that continue to affect the country today. At present these issues are being examined in the public square once again due to the up-coming referendum on the indigenous voice to parliament. Other international issues involving these biases include machine learning annotators being more likely to incorrectly classify African American English text as hateful or toxic, as well as discriminatory facial recognition and healthcare algorithms.

The potential for LLMs to generate outputs that reflect racial, indigenous or colonial biases is a preceding concern. For instance, we know that language models are primarily trained on text from Western sources, which means they will inevitably overrepresent Western values and languages, while marginalising indigenous or non-Western ones.

Resources

• Adult Income Dataset: Also known as the “Adult” dataset and it is extracted from the 1994 Census database. It includes features like age, workclass, education, and race, and it serves the task of predicting whether the income exceeds $50K/yr given the demographics. 
• StereoSet dataset and benchmark: used to measure stereotype bias in language models across gender, race, religion, and profession. It consists of 17,000 sentences and can be used to measure model race preferences.
• UN Declaration on the Rights of Indigenous Peoples: This document might serve as a starting point to create some questions for how an LLM understands or respects the mentioned articles.
• The Racial Bias In Data Assessment Tool developed by Chapin Hall at The University of Chicago helps assess the likelihood of racial bias in a dataset.
• Harvard University has developed a range of Implicit Association Tests, including several that test for racial biases.
• A recent survey paper on bias and fairness in machine learning gives several examples of racial bias in real world scenarios, as well as efforts to map and counteract these issues

Disability discrimination refers to situations where a person is treated “less favourably, or not given the same opportunities as others in a similar situation because of their disability” (Australian Human Rights Commission 2023). Disability itself has a broad and often contested definition, encompassing varied sensory, physical, neurodiverse, cognitive and psychosocial disabilities. Discrimination against persons with disability can happen through individual prejudice, but also can be written into legislation, hiring practices, medical diagnosis and pathologisation, and normative judgments about capacity and performance. Advocacy groups such as the Disability Rights Movement have sought to combat such discrimination through many channels over the past century. 

Artificial intelligence has the potential to revolutionise accessibility for individuals with disabilities. These tools can assist in various ways, such as providing real-time captioning for those with hearing impairments or offering voice recognition software for individuals with mobility limitations. However, they can also perpetuate biases and discrimination if not properly regulated. It is essential for policymakers, organisations, and developers to proactively address this issue by implementing robust regulations and guidelines standards such as the Web Web Content Accessibility Guidelines (WCAG) improving access to information and decision-making. As much as automation can lead to improvements in vision, hearing, mobility and interpretability, it can also embed discriminatory decision-making behind opaque algorithms and data sets. By ensuring that AI tools are developed with inclusivity in mind, we can mitigate the risk of perpetuating disability discrimination. Organisations should prioritise diversity and inclusion when training AI algorithms. By incorporating diverse datasets that accurately represent individuals with disabilities, we can reduce bias and create more equitable outcomes. The arrival of AI and large language models presents another frontier for the identification and mitigation of disability bias. How do LLMs complete sentences, for example, that contain markers of disability? Do they reproduce known societal biases, or has reinforcement learning taught models to be less discriminatory? Research on earlier models  

Resources

• Australian Human Rights Commission (2023). Disability discrimination.  
• Gadiraju, V., Kane, S., Dev, S., Taylor, A., Wang, D., Denton, E., & Brewer, R. (2023, June). “I wouldn’t say offensive but…”: Disability-Centered Perspectives on Large Language Models. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 205-216).
• Magee, L., Ghahremanlou, L., Soldatic, K., & Robertson, S. (2021). Intersectional bias in causal language models. arXiv preprint arXiv:2107.07691. 
• Whittaker, M., Alper, M., Bennett, C. L., Hendren, S., Kaziunas, L., Mills, M., … & West, S. M. (2019). Disability, bias, and AI. AI Now Institute, 8.
• The Ableist Bias Dataset is a recently published collection of prompt templates designed to elicit intersectional biases in generative language models, including ableist biases.
• Students with a disability by age and gender dataset from South Australian Government Data Directory is a ten-year data of schools students with a disability who are verified by a Department for Education psychologist or speech pathologist as eligible for the Department for Education Disability Support Program. 
• BITS is a recently published dataset of prompts designed to test for intersectional biases against people with disabilities in toxicity classifier models, but may also be useful for testing generative language models.