USA
Why voters should care about what ChatGPT tells them about history and politicsPeter Finn
Lauren C. Bell
Amy Tatum
Caroline Leicht
December 2nd, 2024
LSE
50 States or Bust! project
With the 50 States or Bust! project, we set out to explore exactly these types of questions in a systematic rather than an anecdotal fashion. Launched in Spring 2023, we began by developing a standard list of prompts for one of the most popular generative AI tools, ChatGPT, (essentially requests for ChatGPT to carry out a task) that could be tweaked for all US states and territories. These prompts were designed to provide insight into how ChatGPT would respond to inquiries about the history and politics of US states and territories, and which sources of information would be provided when requested by the prompts. Between 22 June and 27 June 2023, these prompts were entered into ChatGPT4 for all states and territories, which generated 56 profiles which varied in length from 123 to 705 words.
Following the creation of our profiles we began speaking to academic experts on state* level US politics about these profiles. We gave the experts a standard list of questions that asked them to judge the profiles that had been generated by ChatGPT in relation to the portrayal of history and politics and how well they drew from academic materials. We asked experts for qualitative responses, as well as asking them to rate the profiles on history, politics, and use of academic materials from 1-10 (with 10 being the highest). These interviews, as well as a standardised discussion of the politics of these states, can be found on our project podcast. As of writing, we have carried out 19 interviews with experts.
When it comes to information about the states, ChatGPT can’t do it all
In new research, we drew on 17 interviews with experts from a diverse range of states including Virginia, California, Utah, and Florida. Based on these interviews, we discovered the following:ChatGPT can’t do it all: Tools such as ChatGPT are clearly useful for some tasks. However, experts identified factual inaccuracies in over 40 percent of the profiles generated.
Difficult to verify: Even though our prompts asked for sources in all instances, the information provided was often so vague as to be of little use.
Hallucinated information: Our interviews demonstrated that ChatGPT tends to hallucinate false information, such as stating that Ohio has a Congressional delegation of 16 rather than 15, and incorrectly maintaining that Iowa’s current boundaries were set in 1851, when the state entered the union in 1846. Not only is it hard to judge the validity of some of the information or factual assertions ChatGPT provides, in some cases the information is simply factually wrong.
Citations that do not exist: As well as hallucinations of factual information, ChatGPT also hallucinated citations, meaning that it said it has drawn from sources that do not actually exist.
To be fair to OpenAI, the company (part owned by Microsoft) that operates ChatGPT, it is open about potential flaws with ChatGPT, noting, for instance, that ‘[s]ometimes, ChatGPT sounds convincing, but it might give you incorrect or misleading information (often called a “hallucination” in the literature)’ highlighting that it ‘can even make up things like quotes or citations’ and advising users not to rely on it as the ‘only source for research’. Yet, with an increasing amount of content generated by ChatGPT, and similar tools such as Gemini from Google, the cumulative issues illustrated by our research raise important questions about the role of such tools in society, especially as there are very few restrictions on how material generated by tools such as ChatGPT and Gemini can be used. Below we chart three of these that relate to democracy and elections.
Why voters should careFacts matter: As in academia, facts, and the evidence on which they are based, are integral to democracy. As generative artificial intelligence tools become embedded more into our work and private lives, there is a danger that voters come to see the output they generate as factually accurate. Yet, our research demonstrates that this is often not the case.
Sources, and the existence of sources, matter: As important as facts and evidence themselves is where they are drawn from. Facts and evidence drawn from questionable sources is a problem as old as democracy itself. However, generative artificial intelligence tools add to this problem not just by engaging with questionable sources, but by completely inventing sources.
Undermining the idea of truth: Differing interpretations of facts and evidence are key to democracy. Indeed, without such differing interpretations, societies would not develop and evolve. However, for such different interpretations to occur, there must be some agreement on what counts as evidence, and how one arrives at a particular fact. The potential for the almost unlimited generation of large amounts of content that is untethered from evidence has the potential to undermine faith in evidence or the idea of truthful narratives.
Given their increasing pervasiveness in our lives, it is important that voters are aware of the pitfalls, as well as the potential benefits, of tools such as ChatGPT. One avenue we are currently exploring is whether the material upon which ChapGPT is trained causes a bias in favour of nationalised topics in US politics: even when requests are given for information on state politics. Further forward, there is scope for testing our method on newer versions of ChatGPT (though these test profiles for California and Utah with ChatGPT4o appear to show many of the same issues as our initial dataset) or other tools such as Gemini. Moreover, it would be interesting to know what would occur if similar prompts were developed and tested for other countries such as Canada and the UK.* At present the authors have not managed to secure any interviews with territory level experts but would welcome such engagement. Please do get in touch if you are an expert on US territory level politics and are interested in being interviewed
Photo by Solen Feyissa on Unsplash
The past two years have seen artificial intelligence (AI) tools such as ChatGPT become widely used by those looking for information – including information to help them decide how to vote. But how accurate are AI tools’ responses when asked about political information? In new research, Peter Finn, Lauren C. Bell, Amy Tatum, and Caroline V. Leicht prompted ChatGPT to create short political and historical profiles of US states, and then asked human experts to rate them. They found that ChatGPT really can’t do it all: experts identified factual inaccuracies in over 40 percent of the profiles, and often provided vague and even false information and citations.
Key to most understandings of democracy is that voters (and citizens in general) should have sufficient evidence about the policies and actions of those who govern them to make an informed decision about their performance and who to vote for in future elections. Recent decades have seen many of the transparency mechanisms which can support this informed decision-making moving into the digital realm with examples, such as the National Security Archive based at The George Washington University, showing the value of focused oversight available to both academics and voters in general.
Increasingly, however, generative artificial intelligence (AI) tools are being embedded in our lives as the technology infrastructure, such as the initial online offerings of traditional media sources and social media sites, that dominated the first two decades of the twenty first century is superseded. What then, are the possibilities offered by such tools in terms of academic research and knowledge, and why does it matter to the average voter?
Key to most understandings of democracy is that voters (and citizens in general) should have sufficient evidence about the policies and actions of those who govern them to make an informed decision about their performance and who to vote for in future elections. Recent decades have seen many of the transparency mechanisms which can support this informed decision-making moving into the digital realm with examples, such as the National Security Archive based at The George Washington University, showing the value of focused oversight available to both academics and voters in general.
Increasingly, however, generative artificial intelligence (AI) tools are being embedded in our lives as the technology infrastructure, such as the initial online offerings of traditional media sources and social media sites, that dominated the first two decades of the twenty first century is superseded. What then, are the possibilities offered by such tools in terms of academic research and knowledge, and why does it matter to the average voter?
50 States or Bust! project
With the 50 States or Bust! project, we set out to explore exactly these types of questions in a systematic rather than an anecdotal fashion. Launched in Spring 2023, we began by developing a standard list of prompts for one of the most popular generative AI tools, ChatGPT, (essentially requests for ChatGPT to carry out a task) that could be tweaked for all US states and territories. These prompts were designed to provide insight into how ChatGPT would respond to inquiries about the history and politics of US states and territories, and which sources of information would be provided when requested by the prompts. Between 22 June and 27 June 2023, these prompts were entered into ChatGPT4 for all states and territories, which generated 56 profiles which varied in length from 123 to 705 words.
Following the creation of our profiles we began speaking to academic experts on state* level US politics about these profiles. We gave the experts a standard list of questions that asked them to judge the profiles that had been generated by ChatGPT in relation to the portrayal of history and politics and how well they drew from academic materials. We asked experts for qualitative responses, as well as asking them to rate the profiles on history, politics, and use of academic materials from 1-10 (with 10 being the highest). These interviews, as well as a standardised discussion of the politics of these states, can be found on our project podcast. As of writing, we have carried out 19 interviews with experts.
When it comes to information about the states, ChatGPT can’t do it all
In new research, we drew on 17 interviews with experts from a diverse range of states including Virginia, California, Utah, and Florida. Based on these interviews, we discovered the following:ChatGPT can’t do it all: Tools such as ChatGPT are clearly useful for some tasks. However, experts identified factual inaccuracies in over 40 percent of the profiles generated.
Difficult to verify: Even though our prompts asked for sources in all instances, the information provided was often so vague as to be of little use.
Hallucinated information: Our interviews demonstrated that ChatGPT tends to hallucinate false information, such as stating that Ohio has a Congressional delegation of 16 rather than 15, and incorrectly maintaining that Iowa’s current boundaries were set in 1851, when the state entered the union in 1846. Not only is it hard to judge the validity of some of the information or factual assertions ChatGPT provides, in some cases the information is simply factually wrong.
Citations that do not exist: As well as hallucinations of factual information, ChatGPT also hallucinated citations, meaning that it said it has drawn from sources that do not actually exist.
To be fair to OpenAI, the company (part owned by Microsoft) that operates ChatGPT, it is open about potential flaws with ChatGPT, noting, for instance, that ‘[s]ometimes, ChatGPT sounds convincing, but it might give you incorrect or misleading information (often called a “hallucination” in the literature)’ highlighting that it ‘can even make up things like quotes or citations’ and advising users not to rely on it as the ‘only source for research’. Yet, with an increasing amount of content generated by ChatGPT, and similar tools such as Gemini from Google, the cumulative issues illustrated by our research raise important questions about the role of such tools in society, especially as there are very few restrictions on how material generated by tools such as ChatGPT and Gemini can be used. Below we chart three of these that relate to democracy and elections.
Why voters should careFacts matter: As in academia, facts, and the evidence on which they are based, are integral to democracy. As generative artificial intelligence tools become embedded more into our work and private lives, there is a danger that voters come to see the output they generate as factually accurate. Yet, our research demonstrates that this is often not the case.
Sources, and the existence of sources, matter: As important as facts and evidence themselves is where they are drawn from. Facts and evidence drawn from questionable sources is a problem as old as democracy itself. However, generative artificial intelligence tools add to this problem not just by engaging with questionable sources, but by completely inventing sources.
Undermining the idea of truth: Differing interpretations of facts and evidence are key to democracy. Indeed, without such differing interpretations, societies would not develop and evolve. However, for such different interpretations to occur, there must be some agreement on what counts as evidence, and how one arrives at a particular fact. The potential for the almost unlimited generation of large amounts of content that is untethered from evidence has the potential to undermine faith in evidence or the idea of truthful narratives.
Given their increasing pervasiveness in our lives, it is important that voters are aware of the pitfalls, as well as the potential benefits, of tools such as ChatGPT. One avenue we are currently exploring is whether the material upon which ChapGPT is trained causes a bias in favour of nationalised topics in US politics: even when requests are given for information on state politics. Further forward, there is scope for testing our method on newer versions of ChatGPT (though these test profiles for California and Utah with ChatGPT4o appear to show many of the same issues as our initial dataset) or other tools such as Gemini. Moreover, it would be interesting to know what would occur if similar prompts were developed and tested for other countries such as Canada and the UK.* At present the authors have not managed to secure any interviews with territory level experts but would welcome such engagement. Please do get in touch if you are an expert on US territory level politics and are interested in being interviewed
.
This article is based on the paper, ‘Assessing ChatGPT as a Tool for Research on US State and Territory Politics’ in Political Studies Review.
This article is based on the paper, ‘Assessing ChatGPT as a Tool for Research on US State and Territory Politics’ in Political Studies Review.
Photo by Solen Feyissa on Unsplash
Note: This article gives the views of the author, and not the position of USAPP – American Politics and Policy, nor the London School of Economics.
Shortened URL for this post: https://wp.me/p3I2YF-ezi
About the author
Peter Finn
Dr Peter Finn is a multi-award-winning Senior Lecturer in Politics at Kingston University. His research is focused on conceptualising the ways that the US and the UK attempt to embed impunity for violations of international law into their national security operations. He is also interested in US politics more generally, with a particular focus on presidential power and elections. He has, among other places, been featured in The Guardian, The Conversation, Open Democracy and Critical Military Studies.
Lauren C. Bell
Lauren C. Bell is the James L. Miller Professor of Political Science and Associate Provost and Dean of Academic Affairs at Randolph-Macon College in Ashland, Virginia. She is the author of Filibustering in the US Senate (Cambria Press: 2011) and Warring Factions: Interest Groups, Money, and the New Politics of Senate Confirmation (The Ohio State University Press: 2002), and co-author of Slingshot: The Defeat of Eric Cantor (CQ Press 2015). She served as a United States Supreme Court Fellow during 2006-07, and was an American Political Science Association Congressional Fellow on the US Senate Committee on the Judiciary from November 1997 until August 1998.
Amy Tatum
Amy Tatum is a Lecturer in Communication and Media at Bournemouth University. She specialises in US politics and the role of women in politics.
Caroline Leicht
Dr Caroline Leicht is a Tutor in Media, Culture and Society at the University of Glasgow and a Visiting Fellow in the Department of Politics and International Relations at the University of Southampton. Her research focuses on gender, media, and political communication, particularly on social media. She previously covered the 2020 US Presidential Election for German public broadcaster ZDF.
Shortened URL for this post: https://wp.me/p3I2YF-ezi
About the author
Peter Finn
Dr Peter Finn is a multi-award-winning Senior Lecturer in Politics at Kingston University. His research is focused on conceptualising the ways that the US and the UK attempt to embed impunity for violations of international law into their national security operations. He is also interested in US politics more generally, with a particular focus on presidential power and elections. He has, among other places, been featured in The Guardian, The Conversation, Open Democracy and Critical Military Studies.
Lauren C. Bell
Lauren C. Bell is the James L. Miller Professor of Political Science and Associate Provost and Dean of Academic Affairs at Randolph-Macon College in Ashland, Virginia. She is the author of Filibustering in the US Senate (Cambria Press: 2011) and Warring Factions: Interest Groups, Money, and the New Politics of Senate Confirmation (The Ohio State University Press: 2002), and co-author of Slingshot: The Defeat of Eric Cantor (CQ Press 2015). She served as a United States Supreme Court Fellow during 2006-07, and was an American Political Science Association Congressional Fellow on the US Senate Committee on the Judiciary from November 1997 until August 1998.
Amy Tatum
Amy Tatum is a Lecturer in Communication and Media at Bournemouth University. She specialises in US politics and the role of women in politics.
Caroline Leicht
Dr Caroline Leicht is a Tutor in Media, Culture and Society at the University of Glasgow and a Visiting Fellow in the Department of Politics and International Relations at the University of Southampton. Her research focuses on gender, media, and political communication, particularly on social media. She previously covered the 2020 US Presidential Election for German public broadcaster ZDF.
No comments:
Post a Comment