ChatGPT summary: “The article explores the rising use of AI in everyday life and governance, highlighting concerns about trust, especially in the U.S., where skepticism toward government AI use is high. Despite AI’s rapid advancements and increasing government investment, public trust in AI for tasks like social media monitoring and welfare decisions remains low. Studies suggest that keeping human oversight in AI processes can help build trust. In the Global South, AI adoption is growing, but biased training data limits its effectiveness. MIT GOV/LAB is researching how AI can be used in governance to foster equality and trust in these regions.”
Google AI (Gemini) search summaries are now at the top of the page from Google searches, and AI tools help us write emails, text messages, and code. It’s commonplace to ask in meetings, “well, what does ChatGPT say about that? or “have you run the question by Gemini?” However, policies on acceptable use of these tools have been slow to materialize.
At MIT GOV/LAB, we are thinking about what advances in AI mean for democracy, governance, and trust between citizens and government. We’ve engaged in these conversations for several years now, including an MIT policy paper on generative AI for pro-democracy platforms and research by practitioner-in-residence Luke Jordan on AI/ML for development and governance. As part of MIT GOV/LAB’s focus on technology and governance, we’re now exploring implications for AI on governance in the Global South. While MIT GOV/LAB’s focus on AI and governance may seem like an obvious extension of our work, the following summary articulates our trajectory and current thinking.
AI is performing as well as people on some cognitive tasks.
The graph below from the 2024 Stanford Artificial Intelligence Index shows how dramatically AI has improved relative to human capacity. In the last few years, we see stunning improvement in AI capability, and the data show AI performance to be close to or better than that of humans on a range of cognitive tasks. We are now in uncharted waters.
Stanford Artificial Intelligence Index, 2024.
Governments are clearly seeing the potential of AI with increasing spending, year after year.
The graph below shows increasing spending by the U.S. government, with estimated investment in AI more than doubling between 2018 and 2023 to 3.3 billion USD (Stanford Artificial Intelligence Index, 2024).
In 2020, Engstrom et al. (2020) surveyed federal agencies and counted 157 applications of AI use. Three years later, the Federal government’s most recent list of AI use cases (Sept. 2023) reports over 700 applications in various stages of development or deployment.
Source: Govini (2023), Chart: 2024 AI Index report.
How do people feel about government use of AI?
At the same time, Americans are highly skeptical. The Schwartz Reisman Institute for Technology & Society (2024) conducted an in-depth survey on attitudes towards AI in 21 countries. Respondents were asked how much they support the use of AI to perform various government functions — from monitoring social media for public safety to determining eligibility and amounts for welfare and social security entitlements or determining eligibility for visas and immigration. The US showed the lowest support for these uses relative to nearly every other country in every use case. The graph below shows how many respondents agreed with a sample of government AI use cases in the US and countries from the Global South.
Sample of data from Schwartz Reisman Institute for Technology & Society (2024) visualized by MIT GOV/LAB.
If the public opinion data isn’t convincing enough, experimental data corroborates these findings. In one study by Kreps & Jakesch (2023), respondents were told whether AI was used or not in email replies from legislators to constituents, some of which were written either entirely by or with the assistance of large language models (LLMs). Though respondents tended to trust the legislator the most when responses were in fact written entirely or partially by LLMs, telling respondents that the information was written by AI caused trust to go down in every case. Other studies find similar trust penalties for AI use in tax fraud identification programs (Ingrams et al., 2021) or reviews of social welfare applications (Gaozhao et al., 2023).
Kreps & Jakesch (2023).
So, why are US citizens so distrustful of AI? It’s not the case that Americans simply have the lowest trust in government (2024 Edelman Trust Barometer), nor do Americans report the highest levels of direct experience with AI applications (e.g., ChatGPT; Schwartz Reisman Institute for Technology & Society). It could have something to do with negative hype in the press; it could be that the US is at the leading edge of implementation; or it could be something else entirely (cue the Hollywood killer robot movies). Right now, we don’t have a good answer.
What could increase people’s trust in AI? Some experimental evidence.
This raises the question of what aspects of government AI applications implemented in good faith do earn trust. Explainability (i.e., knowing how something works) is often stated as a core determinant of AI trustworthiness (Lipton, 2018), and it is tempting to assume that providing information about how AI applications work would increase trust in government AI use (de Fine Licht & de Fine Licht, 2020). Surprisingly, the evidence is mixed. In one study, telling respondents that government AI applications had been built to ensure explainability and positive social outcomes had no effect on trust in those applications (Kleizen et al., 2023). Another study found reduced trust in a non-governmental AI application with explainability (Leichtmann et al., 2023), while still others found no measurable effect of explainability on trust (Duarte et al., 2022) or task performance (Alufaisan et al., 2021).
On the other hand, there is indirect support that a different factor increases trust in government AI use: retaining a “human in the loop” (i.e., there is a real human exercising oversight in the decision-making process). Haesevoets et al. (2024) conducted several experiments examining policy decisions made by humans consulting AI (see the chart below). Using multiple methods, they found that a 75%-human-made / 25%-AI-made ratio engendered the greatest citizen acceptance, even relative to making decisions with no AI component whatsoever. Similarly, retaining a human in the loop increased trust in health insurance program recommendations (Aoki, 2021), preference for criminal recidivism prediction programs (Kennedy et al., 2022), and perceived legitimacy of decisions for a variety of government programs (Waldman & Martin, 2022). These experiments demonstrate that human involvement in AI is critical for building trust, but it remains unclear exactly how this will work in practice.
Haesevoets et al. (2024).
AI use by governments in the Global South? A growing gap.
In addition to the trust issue, there is also a significant bias in how AI tools are being built — the data used to train AI models is often in English and from U.S. or other WEIRD sources (western, educated, industrialized, rich and democratic; Henrich et al., 2010). Furthermore, the research on AI, like the papers included in this piece, is primarily coming out of U.S. and European institutions and journals. This is a problem because it means that it is difficult to use AI in government contexts like supporting decision-making and determining eligibility for social welfare programs if there are no data contextualized to the local geographies.
At the same time, it’s clear that governments in the Global South are starting to use AI (see reports from OECD on Latin America and the Caribbean and Brookings and GSMA on AI in Africa). But the rate at which governments are beginning to use AI may be outpacing the process of diversifying AI models using training data from the Global South.
At MIT GOV/LAB, we are interested in understanding how AI is being used in the Global South around key governance issues and by government and civil society actors. Importantly, we aim to understand how AI use can affect trust between citizens and government. As public sector AI use becomes more common around the world, we aim to produce research and cases to demonstrate ways of using AI that encourage more equality and build trust.
Photo by Google DeepMind on Unsplash. An artist’s illustration of artificial intelligence (AI). This image explores how multimodal models understand a users input and generate an output.