Everyone is talking about AI. Experts across governments, big tech and academia know decisions need to be made on how to move forward. But no one can agree how. Should we push for fast development to maximise on benefits? Or proceed cautiously to ensure it is safe and doesn’t affect people and the planet badly?
We launched findings, recommendations and reflections from our Responsible Ai UK-funded Public Voices in AI project in parliament Wednesday 26 March.
Public Voices in AI comes from the starting point that listening to people from all walks of life will improve AI. It is not a distraction, but imperative to its successful development. Asking people about their hopes and fears about AI is also the right thing to do, if it is going to change our lives as much as we are told it will. Through listening to people’s thoughts, feelings and experiences, we are more likely to make decisions that ensure AI works for all, rather than only those most likely to profit from it.
Evidence tells us what we know about something & influences decisions and discussions, including those about AI, especially in places like parliament. Policy-makers and politicians will tell you that they make evidence-based decisions. This means that who makes evidence and who is included in evidence is important. We know that research exists on how people feel about and experience AI, but not on who is doing this research about whom.
At our event in Parliament, I presented findings and recommendations of an innovative evidence review, summarising a presentation and call to action I gave around the Paris AI Action Summit in February. Sara Cannizaro and I collated and interrogated 320 pieces of evidence, that we categorised as follows:
WHO (researchers) say WHAT, in WHAT WAY, HOW, WHY and WHERE (their motivations, money, research design, data collection, findings, conclusions and recommendations)
WHO (people/publics) say WHAT about WHAT, in WHAT WAY, HOW, WHY and WHERE (how people feel about which AI, how they were asked, where they are from, who they are (demographics), why they were asked).
We found that people who are more positive about AI will find that people are also more positive about AI, whether that is AI in general, or specific AI technologies, such as facial recognition, or specific use cases, such as health. On the other side of the fence, concerned critics of AI are more likely to find that people are concerned about AI. This shows that the motivations and methods of researchers will affect what evidence they find about public opinions. Gina Neff, Professor of Responsible AI at Queen Mary University of London; Deputy CEO, RAi UK, called these sides of the argument the builders and the blockers, and suggested that we need to find a productive way to move beyond these differences. I argued that this productive space comes from people being able to state that they are not represented in the evidence, or are misrepresented.
We found little evidence of ‘evidence’ that properly included diverse populations. There was a lot of ‘inclusion washing’ - the targeting of specific groups that would enable an impression of acceptance. One example that sticks out for me is a finding that black women support AI in breast cancer diagnosis, when only black women were targeted to speak to, and asked questions that were worded in a way that enabled the researchers to make a claim like this easily. Other research on black women’s experience of healthcare suggests they suffer most from misdiagnosis, and are often not taken seriously. If this evidence is not rigorous, this finding will first, misrepresent opinions of an already misrepresented group, and second, mislead decision-makers to invest in AI that is supported by a group that in fact may not really support it, and they will do so believing they are improving systems for the public good.
Sometimes, research on which evidence is based includes diverse, under-represented groups in the sample, but doesn’t analyse what the data tells us about these different groups. Demographic data may be collected, but not used to understand differences in opinions. There may also be little consideration of how to phrase questions that can be clearly understood by people of different languages, ages or cultures. Are your questions available to those who are visually impaired, and/or deaf? What if they don’t have the internet or a mobile? Are you explaining or asking about AI in a way that makes sense to people?
How findings are presented is also important. Here is an example:
“Demographic traits explained the most variance in comfort with AI revealing that men and those with higher perceived technology competence were more comfortable with AI.”
This is not an incorrect statement, and it also incorporates people’s demographic data. But how would a policy-maker or AI developer interpret this as a call to action: what could they do? They might think let’s increase technology competence - this will increase AI comfort. Yet what is important here is understanding why some groups feel uncomfortable. More complex issues of social inequality are at play here. How we phrase findings and recommendations should also represent what people actually said about their lives.
I finished my presentation in Parliament by arguing that we need good evidence to make good AI. This means it must be of good quality in terms of social science research methods rigour, and assurance of public voice. Good evidence is good at:
- Including people who are often overlooked
- Being sensitive to needs of diverse groups
- Incorporating differences in analysis
- Reporting on inequalities in findings
- Communicating diverse perspectives clearly in recommendations
- Reflecting on how it serves the communities spoken to
This is what our project has tried to do!
The evidence review benefited enormously from input from the People’s Panel in AI, coordinated by Connected by Data. Thanks to their advice, findings will be shared through an innovative comic format called a webtoon, by a young working class illustrator / comic artist from Plymouth called Kitty McKewan.
For more information on the numerous findings and outputs, see the Public Voices in AI website, here.
The research reported here was undertaken as part of Public Voices in AI, a satellite project funded by Responsible AI UK and EPSRC (Grant number: EP/Y009800/1).
Public Voices in AI was a collaboration between: the ESRC Digital Good Network at the University of Sheffield (Grant number: ES/X502352/1), Elgon Social Research Limited, Ada Lovelace Institute, The Alan Turing Institute, and University College London.
Public Voices in AI was a year-long (2024-25) research project that aimed to ensure that public voices are attended to in artificial intelligence (AI) research, development, deployment and policy (‘AI RDD&P’). It synthesised, reviewed, built and shared knowledge about public views on AI and engaging diverse publics in AI RDD&P, with and in consultation with target beneficiaries working in (responsible) AI and members of the public, especially from groups most negatively affected by and underrepresented in AI.
Public Voices in AI team:
Helen Kennedy Professor of Digital Society, University of Sheffield. Director of the Digital Good Network
Ros Williams, Senior Lecturer in Digital Media and Society, University of Sheffield. Associate Director of the Digital Good Network.
Susan Oman, Senior Lecturer in Data, AI and Society, AI and In/equalities Lead, Centre for Machine Intelligence & Senior Researcher at the Digital Good Network, University of Sheffield
Helen Margetts, Professor of Society and the Internet, University of Oxford, Director, Public Policy Programme, Alan Turing Institute for Data Science and AI.
Octavia Field Reid, Associate Director (Public participation & research practice), Ada Lovelace Institute.
Jack Stilgoe, Professor of Science and Technology Policy, Dept of Science & Technology Studies, University College London.
Eleanor O’Keeffe, Public Participation & Research Practice Lead, Ada Lovelace Institute.
Roshni Modhvadia, Researcher, Ada Lovelace Institute.
Mhairi Aitken, Senior Ethics Fellow, Alan Turing Institute.
Cian O’Donovan, Senior Research Fellow, Department of Science and Technology Studies, University College London.
Tvesha Sippy, Researcher, Alan Turing Institute.
Sara Cannizzaro, Postdoctoral Researcher, Public Voices in AI project.
Ruth Lauener, Manager, Digital Good Network, University of Sheffield.
Sarah Givans, Research Support Project Administrator, Digital Good Network, University of Sheffield.