Conversational AI increases political knowledge as effectively as self-directed internet search

Abstract

Conversational AI systems are increasingly being used in place of traditional search engines to help users complete information-seeking tasks. This has raised concerns in the political domain, where biased or hallucinated outputs could misinform voters or distort public opinion. However, in spite of these concerns, the extent to which conversational AI is used for political information-seeking, as well as the potential impact of this use on users’ political knowledge, remains uncertain. Here, we address these questions: First, in a representative national survey of the UK public (n = 2,499), we find that in the week before the 2024 election, as many as 32% of chatbot users—and 13% of eligible UK voters—have used conversational AI to seek political information relevant to their electoral choice. Second, in a series of randomized controlled trials (n = 2,858 total), we find that across issues, models, and prompting strategies, task-directed conversations with AI to research-specific political topics increase political knowledge (increase belief in true information and decrease belief in misinformation) to the same extent as self-directed Google search. Taken together, our results suggest that people in the United Kingdom are increasingly turning to conversational AI for information about politics. These findings substantially extend prior work by demonstrating that conversational AI's effects on political knowledge generalize across multiple topics, political perspectives, and model families, suggesting that the shift toward AI-assisted political information-seeking may not lead to increased public belief in political misinformation.

Significance statement.

As conversational AI systems rapidly replace traditional search engines, concerns have emerged about their potential to spread political misinformation and bias electoral outcomes. This study provides a broad epistemic health assessment in the context of conversational AI's actual use and impact on political knowledge. Using a representative UK survey during the 2024 election, we reveal widespread adoption: 13% of eligible voters used AI chatbots for election-related information-seeking. Through randomized controlled trials, we demonstrate that conversational AI strengthens belief in true information while reducing belief in misinformation as effectively as traditional Google search, with effects generalizing across topics and models. While both AI- and search-guided participants toward greater agreement with progressive-leaning factual statements, this shift was comparable across the ideological spectrum. These findings suggest the transition to AI-assisted political information-seeking may not exacerbate misinformation problems, informing ongoing debates about AI governance and democratic integrity.

Introduction

Conversational AI systems (or chatbots) such as ChatGPT, Claude, and Gemini are now used regularly by hundreds of millions of people across the world. Analysis of consumer usage reveals that among the most common use cases is the seeking of information for private and professional purposes, including both general knowledge and practical advice (1). As public usage of chatbots continues to increase, a particular concern is that large language models (LLMs) will produce unreliable or biased answers when queried about current affairs or political issues. This could be damaging to democracy, which relies on the electorate having access to reliable information, especially before elections (2).

The concerns about AI's potential impact on political knowledge are situated within a longstanding debate in political science about whether factual information actually matters for democratic decision-making. Some scholars have argued that voters can effectively compensate for being relatively poorly informed by relying on heuristics, party affiliation, and opinion leaders (3, 4), suggesting that the quality of political information sources may be less consequential than commonly assumed. However, this optimistic view depends on the assumption that voters “know” they are uninformed and respond accordingly. Kuklinski et al. (5) have shown that misinformed citizens, those who hold confident but incorrect beliefs, represent a qualitatively different challenge: they make systematically different policy judgments than both informed and uninformed citizens, and their false confidence makes them more resistant to correction of these beliefs. Indeed, attempts to correct political misinformation can fail or even backfire among ideologically committed individuals (6), though more recent work suggests that such backlash effects may be less prevalent than initially thought (7). The broader information environment compounds these challenges: the rapid acceleration of content production and consumption has intensified competition for limited collective attention (8), while exposure to untrustworthy sources, though perhaps less prevalent than commonly feared (9), remains a persistent concern, particularly as social media has become a major outlet for political information, with false news stories reaching wide audiences across the ideological spectrum (10). Moreover, exposure to opposing viewpoints does not straightforwardly reduce polarization but can rather reinforce preexisting attitudes (11). Against this backdrop, the emergence of conversational AI as a new source of political information raises the question of whether these systems will exacerbate or ameliorate existing challenges to an informed electorate.

Two main uncertainties make the extent of potential risks unclear. First, it is unclear whether people trust AI enough to use it for political information-seeking in high-stakes political moments, such as in the lead up to national elections. Among the general public, trust in AI is low: surveys consistently report that a majority of people do not trust AI to be used as tools to provide medical care, legal advice, or to assist human journalists (12–14). One recent survey showed that levels of trust in AI were comparable to those for politicians, advertising executives, and social media influencers who are themselves the least trusted of all professions (15). Given this, it is an open question whether people will turn to conversational AI for political guidance during elections, even when such tools are available.

Second, it is unclear whether conversational AI systems present functional substitutes for traditional search engines on common political issues. On the one hand, it is widely acknowledged that AI models have a problem with factuality (AI developers have devoted considerable energy to studying what they call “hallucinations” in AI models (16–19)), and there has been concern that models may be politically biased, especially in a progressive or libertarian direction (20–22). On the other hand, significant progress has been made in tackling these risks: developers have implemented techniques such as retrieval-augmented generation (23), knowledge graphs (24), fine-tuning on factuality rankings (25), and semantic entropy methods for detecting confabulations (26). As a result, it remains unclear if reliability issues would substantively impact users’ political knowledge in a real-world information-seeking setting.

A growing literature has begun to characterize how conversational AI differs from traditional internet search as an information source, with implications that cut both ways. On the demand side, evidence suggests that users already blend AI assistants into their information-seeking rather than substituting them fully, often acting on AI-provided information even when they doubt it and intending to verify it later (27). On the supply side, experimental work shows that LLM-based search can speed up decisions at accuracy comparable to conventional search but can also induce overreliance when the model is wrong, an effect attenuated when supporting evidence or model confidence is made visible (28). Relatedly, identical content is judged more credible when delivered in a conversational format than as static, search-style text, with users correspondingly less likely to catch inaccuracies (29). Together, these findings motivate a direct test of whether conversational AI serves as a knowledge-equivalent substitute for internet search on common political issues, rather than merely a faster or more persuasive but potentially less scrutinized interface.

A growing body of work has examined AI's potential to influence political attitudes, demonstrating that AI-generated messages can persuade humans on policy issues (30, 31) and that AI-generated propaganda can be as persuasive as human-written content (32). Recent work has further shown that the persuasive power of conversational AI stems primarily from posttraining and prompting techniques rather than model scale or personalization, though with a concerning trade-off: optimizing AI systems for persuasion systematically reduces the factual accuracy of their outputs (33). However, conversational AI's persuasive capabilities may also be leveraged for positive ends: Costello et al. (34) demonstrated that personalized dialogues with an LLM durably reduced conspiracy beliefs, with effects persisting for at least two months, suggesting that AI's capacity to generate tailored, evidence-based arguments can help correct misinformation rather than spread it. However, these studies examine scenarios in which AI-generated content is delivered to users, rather than the common scenario in which users actively seek out political information from AI systems. Recent work has begun to unravel the role of conversational AI in this latter use case (35), providing initial evidence that conversational AI can improve factual knowledge on specific political topics. However, this preliminary work was limited to single-issue investigations favoring liberal-aligned factual information with a narrow scope of political issues, without built-in controls for general knowledge acquisition or examination of effects across the political spectrum.

In the present investigation, we first show in a representative survey (study 1) 1 week after the UK 2024 general election that as many as 13% of eligible voters may have used conversational AI to find information relevant to their electoral choice. Second, in a randomized controlled trial (RCT; study 2), we find that participants who used a chatbot (GPT-4, Claude, or Mistral) to research factual information related to issues of concern for UK voters increased their political knowledge to the same extent as participants who researched the same issues using Google search. In a follow-up RCT (study 3), we find that even when LLMs were explicitly prompted to use sycophantic or persuasive techniques, participants’ belief formation did not differ from those interacting with standard unprompted AI models. Taken together, these results suggest that although people in the United Kingdom are increasingly turning to AI for political information, this shift may not lead to increased public belief in political misinformation.

Our work makes three key contributions: (1) Replication: We replicate recent findings that conversational AI can improve political knowledge on specific issues (35); (2) Generalization: We demonstrate that these effects generalize across multiple political topics, model families, and prompting strategies and remain consistent across politically balanced information representing both progressive and conservative viewpoints; (3) Broader epistemic health assessment: We provide the first assessment of AI's effects on broader epistemic health indicators (trust, private beliefs, and extremity) and the first nationally representative data for the United Kingdom on real-world chatbot usage for political information during a national election. Importantly, our work examines a specific but common use case: structured, task-directed political information-seeking. Real-world interactions with conversational AI are more varied: people may use chatbots to seek validation for existing views, to discuss politics in less structured ways, or may be redirected away from political engagement entirely. Our findings speak to the direct effects of using AI as a research tool for political information but should not be interpreted as capturing the full equilibrium effects of AI on political knowledge or engagement more broadly.

Results

Survey (study 1)

A representative sample of UK adults eligible to vote (n = 2,499) was surveyed in the 4 days immediately following the UK general election that took place on July 4, 2024 (see Supplementary material for details on survey demographics and survey questions). Respondents primarily relied on traditional media for political information over the preceding 4 weeks: television remained the most common source (54%), followed by social media (36%), internet websites (33%), internet search (29%), radio (28%), and newspapers (27%). By comparison, 9% used AI chatbots as a source of political information. Among chatbot users (n = 1,024), our central finding is that one-third (32%) used chatbots in the lead up to the UK 2024 election to research information relating to current affairs and political issues (Fig. 1A). This comprises the most popular use case, on par with work or educational uses (McNemar's χ²(1) = 0.1, P = 0.753).

Figure 1.

Experimental design for measuring the impact of conversational AI on political knowledge. Participants completed baseline assessments of misinformation belief (primary outcome) across four political topics (criminal justice, COVID-19, immigration, and climate change) using 7-point Likert scales. To assess generalization to broader measures of epistemic health, participants also completed assessments of trust levels, private political beliefs, and extremity indicators. Participants were randomized to using conversational AI chatbots (Claude, GPT-4, or Mistral) or Google search. During the research phase, participants investigated two randomly assigned topics (researched topics) while two others serve as within-subject controls (nonresearched topics). Following the research phase, all measures were re-administered to assess pre-post changes.

Participants who used chatbots for political information found them significantly more useful than nonuseful (89% vs. 11%, n = 430; χ²(1) = 256.3, P < 0.001) and more accurate than inaccurate (87% vs. 13%, n = 417; χ²(1) = 223.1, P < 0.001). Most respondents viewed chatbots as politically neutral rather than showing partisan bias (62% vs. 38%, n = 404; χ²(1) = 24.8, P < 0.001). Among those perceiving bias (n = 152), there was an equal split between right- and left-leaning ideologies (58% vs. 42%, χ²(1) = 3.8, P = 0.052). While respondents were evenly divided on whether chatbots influenced their perspective overall (47% vs. 53%, n = 422; χ²(1) = 1.9, P = 0.173), among those reporting a directional influence (n = 124), liberal influence significantly exceeded conservative influence (63% vs. 37%, χ²(1) = 8.3, P = 0.004). The majority of respondents felt no influence on their voting intentions (60% vs. 40%, n = 443; χ²(1) = 18.7, P < 0.001), but among those reporting an influence (n = 176), most were encouraged rather than discouraged to vote (79% vs. 21%, χ²(1) = 59.1, P < 0.001).

RCTs (studies 2 and 3)

In study 2 (RCT), we recruited an independent sample of UK residents (n = 1,147 final sample) online via Prolific.com. Participants were then linked to a custom-built online experiment app and instructed to research true or false information related to issues of concern for UK voters (each participant was pseudo-randomly assigned to researching two out of four topics: climate change, immigration, criminal justice, and COVID-19 policy; see Supplementary material for details on sourcing and political balancing of the material). The two other topics served as within-subject controls (nonresearched topics) to rule out nonspecific temporal changes to beliefs.

Participants researched the two assigned topics consecutively, either using conversational AI (GPT-4o, Claude-3.5, or Mistral) or Google search (Fig. 2). Both setups were embedded within the online experiment app—participants interacted with conversational AI through an embedded chat window, and Google search was available as an embedded browser window within the app. The experiment app also recorded the time participants spent researching each topic.

Figure 2.

Conversational AI usage patterns and influence on belief in true versus false information. A) Survey (study 1) results: Self-reported use cases for AI chatbots among UK users. B) RCT (study 2) results: Change in agreement with true vs. false information from pre- to post-researching. Left panel shows the conversational AI condition; right panel shows the Google search control condition. Solid lines indicate researched topics, and dotted lines denote nonresearched topics. Error bars represent 95% CIs. C) RCT (study 2) results, left: Bayesian GLM parameter estimates, error bars denote HPDI. Shaded area depicts an a priori defined ROPE, where effect sizes are considered to be negligible/practically 0; Right: GLM comparison using WAIC, as a measure of out-of-sample predictive accuracy of the GLMs (closer to zero is better). Full model = GLM1: GLM including parameters to quantify differences in change effects between conversational AI and Google search conditions, No ConvAI Term = GLM2: GLM not including parameters to quantify differences between different conversational AI models, Null model = GLM3: GLM not including parameters to quantify differences in change effects between conversational AI and Google search conditions or different conversational AI models.

We assessed beliefs in true and false information on a 7-point Likert scale before and after researching all topics. Since participants answered questions about all four topics both before and after the research phase, regardless of whether they researched them, we could compare belief change on topics participants actively researched against belief change on topics they were merely re-assessed on, isolating the effect of the research activity from generic effects of questionnaire repetition or participant engagement. Additionally, we measured other markers of epistemic health as secondary outcomes: trust, private political beliefs, and extremity change. After completing the study, participants were fully debriefed about the aims and hypotheses of the research.

To test our research questions, we fitted and compared three Bayesian Generalized Linear Models (GLM1–3, Eq. 1, see Methods for details). There was no model evidence of differences between conversational AI and Google search conditions.

Model fit metrics suggested no better fit of GLMs that included parameters for differences in change effects between conversational AI and Google search conditions (GLM1: Full model) in comparison to a GLM that did not include these terms (GLM3: Null model; Fig. 1C, right panel). True information received on average approximately one Likert scale point higher agreement ratings than false information (Fig. 1B, purple vs. orange lines), resulting in a nonzero difference parameter in GLM2 (β_TRUE = 0.67, 95%-Highest Posterior Density Interval [HPDI] [0.65; 0.68]; Fig. 1C, right panel). Researching (vs. not researching) political issues increased belief in true information and decreased belief in misinformation across time points (Fig. 1B, solid vs. dotted lines; (β_{POST  ×  TRUE  ×  RESEARCHED} = 0.26, 95%-HPDI [0.19; 0.32], Fig. 1C). Importantly, we found that belief change was nearly identical for participants who researched using conversational AI or Google search (Fig. 1B, left vs. right panel), reflecting in a close to zero parameter estimate (β_{POST × TRUE × RESEARCHED × CONVAI} = 0.02, 95%-HPDI [−0.09; 0.12]; Fig. 1C). This null difference between conditions was further qualified by the fact that the HPDI fully encloses a region of practical equivalence (ROPE; gray-shaded zone), an a priori specified interval of effect sizes that are negligible. This pattern also held separately for each of the different model families tested (GPT, Claude, and Mistral; Fig. 4C–E).

Figure 4.

Belief in true and false information across prompting techniques (study 3) and different conversational AI models (study 2). Change in agreement with true vs. false information from pre- to post-researching for: (A) GPT-4o prompted to be persuasive, (B) GPT-4o prompted to be sycophantic, (C) GPT-4o with standard prompting, (D) Claude, and (E) Mistral. Solid lines indicate researched topics, and dotted lines denote nonresearched topics. Error bars represent 95% CIs.

We repeated the above analyses for trust, private political beliefs, and extremity change and found highly similar results as for beliefs in true and false information (Fig. 3). Trust was measured by asking participants to state their agreement with statements on trust or distrust in politicians, experts, media, and technology (see Supplementary material for details). We found that distrust statements on average produced higher agreement than trust statements (β_DISTRUST = 0.35, 95%-HPDI [0.32; 0.39]). There was no difference in change of trust ratings from before to after researching for researched vs. not researched topics (β_{POSTxDISTRUSTxRESEARCHED} = −0.08, 95%-HPDI [−0.17; 0.01]). Importantly, we found that the patterns of trust and trust change were highly similar for participants who researched using conversational AI or Google search (β_{POSTxDISTRUSTxRESEARCHEDxCONVAI} = 0.026, 95%-HPDI [−0.09; 0.14], Fig. 3A). Refitting the best-fitting GLM excluding the items on trust in politicians (which had slightly different framing than other items) yielded qualitatively identical results. All reported nonzero effects remained nonzero and all null effects remained null. Specifically, the effect of distrust remained robust (β_DISTRUST = 0.12, 95%-HPDI [0.08; 0.15]; 0% ROPE overlap). There was still no difference in change of trust ratings from before to after researching for researched vs. not researched topics (β_{POSTxDISTRUSTxRESEARCHED} = −0.07, 95%-HPDI [−0.15; 0.01]), and crucially, the patterns of trust and trust change remained highly similar for participants who researched using conversational AI or Google search (β_{POSTxDISTRUSTxRESEARCHEDxCONVAI} = 0.02, 95%-HPDI [−0.07; 0.11]). This confirms that the inconsistent phrasing of this single item did not meaningfully affect our conclusions.

Figure 3.

Agreement with trust and distrust statements and private beliefs. Top row: A) Change in agreement with trust vs. distrust statements from pre- to post-researching. B) Change in private beliefs: agreement with progressive and conservative statements from pre- to post-researching. Error bars in top row represent 95% CIs. Bottom row: Bayesian GLM parameter estimates for (A) agreement with trust/distrust statements and (B) agreement with private belief statements. Extremity was computed based on pre- to post-researching change in the sign of the difference to the center point of the Likert scale (4), indicating a flip on more agreement with progressive to more agreement with conservative beliefs (or vice versa). Error bars denote HPDI. Shaded area depicts an a priori defined ROPE, where effect sizes are considered to be negligible/practically 0. Solid lines indicate researched topics, and dotted lines denote nonresearched topics.

Private political beliefs were assessed by asking participants to state their agreement with progressive and conservative views on the 4 topics (see Supplementary Information for details). We observed that progressive statements on average produced higher agreement ratings than conservative statements (β_PROGRESSIVE = 0.66, 95%-HPDI [0.63; 0.70]). Additionally, researching topics increased agreement with progressive views and decreased agreement with conservative views (β_{POSTxPROGRESSIVExRESEARCHED} = 0.23, 95%-HPDI [0.12; 0.34]). Importantly, we found that the effect of researching topics on private political beliefs was highly similar for participants who researched using conversational AI or Google search (β_{POSTxPROGRESSIVExRESEARCHEDxCONVAI} = 0.02, 95%-HPDI [−0.11; 0.15], Fig. 3B).

To investigate whether the observed progressive shift is an artifact of our predominantly progressive-voting sample (progressive: 50.1%; conservative: 23.7%; no clear self-placement: 26.1%), we added a binary ideological self-placement variable (progressive: Labour, Green, Liberal Democrats, SNP, Sinn Féin, Plaid Cymru; conservative: Conservative, Reform UK, Unionist parties; based on self-identification) and its interactions with all experimental predictors. Model comparison confirmed ideological self-placement as a strong predictor of beliefs overall (ΔWidely Applicable Information Criterion [WAIC] = 1,266), driven by “partisan congruence,” i.e. participants agreed more with statements aligned with their political identity (β_{PROGRESSIVExVOTE} = 2.00, 95%-HPDI [1.91; 2.08]). Critically, however, none of the other higher-order interactions involving both pre-post treatment and ideological self-placement were meaningfully different from zero, including the key β_{POSTxPROGRESSIVExRESEARCHEDxVOTE} = 0.04 (95%-HPDI [−0.17; 0.26] and β_POSTxVOTE = 0.003, 95%-HPDI [−0.08; 0.08]), confirming that the progressive shift was comparable across the ideological spectrum. In other words, ideological self-placement (unsurprisingly) strongly predicts “what” people believe on average, but not “how they change” in response to researching information. Both conservative and progressive participants experienced a similar progressive shift, suggesting a general effect rather than a sample composition artifact. Additionally, we observed that exposure to information sources was linked to reduced partisan congruence (β_{PROGRESSIVExRESEARCHEDxVOTE} = −0.30, 95%-HPDI [−0.44; −0.16]). However, we interpret this finding with caution, as the observed effect is not a temporally specific change but rather represents an average effect over time.

Extremity change was defined as sign flips in private political beliefs from before to after researching/not researching an issue—with reference to the center point of the Likert scale (4, values below this value being negative, and values above being positive). Within a binomial GLM, there was no difference in extremity change for progressive or conservative statements (β_PROGRESSIVE = 0.007, 95%-HPDI [−0.07; 0.08]). Additionally, there was no reliable evidence that researching topics changed the extremity of views differentially for progressive or conservative information (β_{PROGRESSIVExRESEARCHED} = −0.16, 95%-HPDI [−0.28; −0.03]; nonzero, but overlapping with ROPE). Again, we found that the effect of researching topics was nearly identical for participants who researched using conversational AI or Google search (β_{PROGRESSIVExRESEARCHEDxCONVAI} = 0.07, 95%-HPDI [−0.14; 0.27]; Fig. 3B, bottom).

The above results were obtained using LLMs with standard prompts. However, models could be prompted to behave in ways that bias the user toward one view or another (persuasion) or to behave “sycophantically,” refusing to contradict the user and reinforcing their existing beliefs. Next, thus, we asked whether AI systems prompted in this way might influence political informedness, belief, or trust to a greater extent than search engines.

In study 3, a structurally similar RCT to study 2 (n = 1,711 final sample), we investigated how a chatbot (GPT-4o) prompted to be sycophantic or persuasive affects beliefs in true and false information (and secondary outcomes) relative to an unprompted baseline GPT-4o (baseline/control condition). Chatbot system prompts were inserted programmatically via the app before the participant's first message, without the participant's knowledge (see Supplementary material for exact prompts). The sycophantic system prompt instructed the LLM to support the users’ preexisting beliefs on the issue, irrespective of whether they agreed or disagreed with the issue (see Box S1). Similarly, the persuasive system prompt instructed the LLM to support a randomly chosen viewpoint (agree/disagree, which corresponded to the users’ preexisting beliefs in 50% of the cases, see Box S2).

The GLM comparison results and parameter estimates obtained were similar to the previous results with standard prompt settings (Fig. 4A and B); no differences were found between prompted vs. unprompted LLMs β_{POST × TRUE × RESEARCHED × PROMPT} = 0.03, 95%-HPDI [−0.06; 0.13]—suggesting that interacting with LLMs prompted to be sycophantic or persuasive did not change participants views above and beyond the view changes achieved by baseline conversational AI models (and by extension, traditional Google search).

Participants’ debrief ratings of model reliability and agreement did not differ between prompted and unprompted conditions (sycophancy: all t ≤ 0.51, P ≥ 0.609; persuasion: all t ≤ 1.61, P ≥ 0.108), suggesting that participants perceived the prompted and unprompted models similarly. It is possible that our sycophancy and persuasion manipulations did not alter model behavior as strongly as intended: for sycophancy, recent work suggests that question-based interactions substantially attenuate sycophantic behavior in LLMs even when system prompts instruct otherwise (36), and participants in our study likely engaged with the chatbot primarily by asking questions. For persuasion, the system prompt explicitly constrained the model to use only factual information and logical arguments on well-covered political topics, which may have limited the model's ability to provide unreliable information.

Even though we found no differences of researching political issues using conversational AI or Google search on epistemic health, we additionally investigated potential time efficiency effects of the search methods. We found that the use of conversational AI reduced the information procurement time by 6–10% in comparison to self-guided Google search (average time spent on both research tasks (±SD) in minutes: 17.94 (±8.64) for conversational AI vs. 19.82 (±8.91) for Google search; β_CONVAI = −0.11, 95%-HPDI [−0.16; −0.07], Gamma GLM, Eq. 3).

Our results replicate and extend recent findings that conversational AI can improve factual political knowledge on specific issues (35). We substantially extend this work in several key ways. First, while prior work examined single issues in isolation, our within-subjects design demonstrates that these effects generalize across four politically salient topics simultaneously (climate change, immigration, criminal justice, and COVID-19). Second, our inclusion of both progressive- and conservative-aligned true information (see Supplementary material) demonstrates that AI's knowledge-enhancing effects are not limited to liberal-favored facts. Third, our within-subject control topics (nonresearched issues) allow us to rule out generic effects of questionnaire repetition or participant engagement, isolating AI-specific learning effects. Fourth, we extend beyond single knowledge questions to examine broader epistemic health indicators, including trust, private political beliefs, and extremity—none of which showed differential effects between AI and traditional search.

Discussion

Replicating and extending recent work (35), we demonstrate that conversational AI now sits alongside Google search as a commonly used source of information during high-stakes political moments like national elections. Our experimental findings speak specifically to task-directed political information-seeking: participants were instructed to research-specific political topics using either conversational AI or Google search. This represents an important and increasingly common use case but does not capture the full range of ways in which people interact with AI in political contexts. Going beyond prior single-issue investigations, we show that across four salient issues, politically balanced true and false information, and three model families, researching with chatbots raised belief in true facts and lowered belief in false claims to the same extent as self-guided Google search, even when the chatbots were prompted to be persuasive or sycophantic. Our within-subject controls for nonresearched topics and secondary outcomes (trust, private beliefs, and extremity) further demonstrate that these effects represent genuine knowledge acquisition rather than generic engagement or politically biased persuasion. These findings stand in contrast to the popular assumption that the use of chatbots for seeking information about news or current affairs may inherently erode political knowledge.

Our multi-topic examination has three main implications. First, we replicate prior work showing that AI can improve political knowledge, while demonstrating for the first time that this effect generalizes across multiple topics and political viewpoints and is not limited to liberal-favored information. Second, our results suggest that the large body of prior work suggesting low levels of public trust in AI belies actual public usage for information-seeking: in fact, our data suggest that information-seeking is the most popular use case for chatbots among UK citizens (surpassing professional use, writing/translation, and practical advice), and that users found their chatbots to be useful, accurate, and unbiased. Third, our experiment suggests that contrary to widespread concern about AI reliability and hallucinations, models do not increase belief in false information when used to research current affairs and political issues across diverse topics and political perspectives.

One notable finding that warrants further investigation is the null effect of conversational AI on trust. Despite widespread concern that AI-generated content may erode public trust in institutions, experts, and media, we found no evidence that researching political information using chatbots affected participants’ trust ratings, compared with using traditional Google search. The disconnect between low stated trust in AI and the absence of any measurable impact on broader institutional trust raises important questions for future research: Does familiarity with AI through direct use gradually shift trust perceptions? Might longer or repeated interactions with chatbots eventually influence trust in ways that a single research session cannot capture? And could the null effect reflect a ceiling or floor effect in trust attitudes that are deeply entrenched and resistant to short-term interventions? Longitudinal studies tracking trust dynamics over sustained AI use would be particularly valuable in addressing these questions.

Our findings also touch on longstanding debates about the role of factual information in democratic decision-making. While some scholars have argued that voters can compensate for relatively low levels of information through heuristics and shortcuts (3), our results are more consistent with the view that the quality of information sources matters: both AI and search engines measurably shifted participants’ beliefs toward factual accuracy on politically salient topics. This is particularly relevant given evidence that misinformed citizens make systematically different policy judgments than merely uninformed ones (5). However, whether these informational gains translate into downstream changes in political attitudes or behavior remains an open question. Prior work suggests that the relationship between information exposure and political decision-making is complex: exposure to counter-attitudinal information does not reliably change minds (6, 11), though backlash effects may be rarer than commonly supposed (7). Future work should examine whether AI-assisted improvements in factual political knowledge persist over time and whether they influence subsequent political attitudes, voting behavior, or broader civic engagement.

These conclusions are bounded by scope: we focused on one country, a single election cycle, four issues, and short interactive sessions with a small sample of models. This sample of models did not include models that have been shown to be highly controversial and biased toward specific ends of the political spectrum. In this sense, our results might underestimate the belief-changing effects that more biased models may have. We tested the effects of using conversational AI for researching political information against a no-AI self-guided Google search condition. While this represents a valid control to rule out generic effects of self-selection and confirmation bias, most internet search engines as of today provide AI-enhanced features, like AI-generated summaries of the most relevant information across websites. Additionally, we did not study the newest generation of chatbots, which are typically internet search-enabled, providing users with most up-to-date information. These constraints might render our study more likely to reflect a comparison to traditional methods of information search. However, field data that link chatbot use to downstream attitudes and behavior remain extremely sparse. We thus believe our results represent an important baseline to reference future developments in AI-guided search of political information and political belief change.

Perhaps most importantly, our experimental design examines task-directed information-seeking, where participants were instructed to research-specific political topics. Real-world AI use is considerably more varied: people may use chatbots to seek validation for preexisting views, to discuss politics in unstructured ways, or conversely, AI may redirect users away from political engagement entirely by making nonpolitical uses more compelling. Our findings therefore speak to the direct epistemic effects of using AI as a political research tool but should not be interpreted as evidence about the broader equilibrium effects of AI availability on political knowledge or engagement. Future work should examine more naturalistic, self-directed interactions with AI in political contexts, including longitudinal designs that capture how AI use patterns evolve over time and how they interact with broader media consumption habits.

Our findings should also be interpreted in light of prior work demonstrating that search engines themselves are not neutral information intermediaries. Epstein and Robertson (37) showed in an experimental setup that biased search engine rankings can shift voting preferences of undecided voters by 20% or more, largely without their awareness, an effect they termed the search engine manipulation effect. More recently, Aslett et al. (38) demonstrated that searching online to evaluate specific misinformation claims can paradoxically increase belief in them, particularly when search engines return low-quality results, a mechanism they attribute to “data voids,” or informational spaces dominated by unreliable sources. These findings stand in apparent tension with our results, which show that both conversational AI and Google search reduced belief in misinformation across topics. We believe these findings are complementary rather than contradictory, and the divergence is informative about the boundary conditions of search-based information-seeking. Several key design differences may account for the different outcomes. First, our participants engaged in broad, self-directed research on political topics rather than evaluating specific preselected misinformation claims. This general information-seeking task is more likely to surface high-quality, mainstream sources, particularly for well-covered topics like climate change, immigration, criminal justice, and COVID-19 policy, reducing the likelihood of encountering the data voids identified by Aslett et al. (38). This interpretation aligns with trace-data studies of real-world search behavior: exposure to partisan or unreliable content on Google is driven more by users’ own selections than by algorithmic curation (39), and engagement with unreliable sites occurs predominantly when users deliberately navigate to them rather than encountering them through general topical queries (40). Broad, topic-level information-seeking of the kind our participants performed is therefore comparatively unlikely to steer users toward low-quality sources. Similarly, Aslett et al. (38) found that their search effect was concentrated among individuals for whom search engines returned lower-quality information and was absent when search results contained only high-quality sources. Second, the topics in our study are among the most extensively covered political issues in the United Kingdom, with substantial high-quality information available from mainstream outlets, government sources, and established think tanks. This likely provided a protective buffer against data voids. Third, and importantly, participants in our study conducted their Google searches through an embedded browser window within our experiment app, which used a clean session without access to participants’ personal browsing history, cookies, or prior search data. This means that the search results participants encountered were not personalized based on their individual browsing behavior or ideological profile, effectively controlling for the kind of algorithmic personalization that can exacerbate filter bubbles and data voids (37, 38). In real-world search, where results are tailored to users’ prior behavior, the effects on political knowledge could differ from what we observe here. Consistent with the view that search outputs are themselves shaped by context, the framing of results can vary systematically: Berkebile-Weinberg et al. (41) find that image-search depictions of climate change, one of the four topics in our study, differ across countries in ways that shape perception. However, our findings cannot be assumed to generalize to more niche or emerging political topics where reliable information may be scarce and low-quality sources may dominate search results nor to personalized search environments where algorithmic curation may direct users toward lower-quality information. The interaction between topic salience, information availability, search personalization, and the effects of AI-assisted information-seeking represents an important avenue for future research. Additionally, Epstein and Robertson's (37) finding that biased search rankings can powerfully influence voter preferences without awareness raises a broader concern about algorithmic curation of political information, whether through search engines or conversational AI. While our study found no differential effects between AI and search on political knowledge, this does not rule out the possibility that either tool could be deliberately manipulated to bias users or algorithmic dynamics might organically favor certain viewpoints. Our finding that both AI- and search-guided participants tend toward greater agreement with progressive-leaning factual statements may partly reflect such dynamics in the underlying information landscape, rather than tool-specific bias.

Additionally, while our sycophancy and persuasion RCTs found no differential effects of prompting on belief change, participant debrief ratings suggest that the manipulations may not have altered model behavior as strongly as intended. For sycophancy, this is consistent with recent evidence that question-based interactions attenuate sycophantic behavior in LLMs (36). For persuasion, the constraint to use only factual information on well-covered topics may have limited the model's ability to deviate from its baseline behavior. These results should therefore be interpreted with caution, and future work should examine whether stronger manipulations, for example, using models fine-tuned for sycophantic or persuasive behavior, or testing on more niche topics where the factual landscape is less well-established, might yield different outcomes. Importantly, however, the paper's central claim that conversational AI increases political knowledge to the same extent as Google search rests on the comparison between standard unprompted AI and search and is not contingent on the prompting results.

While our study replicates the core finding of Taylor and Richey (35) regarding AI's capacity to improve factual knowledge, our multi-topic, politically balanced design with built-in controls provides substantially stronger evidence for the generalizability and political neutrality of these effects. However, like prior work, we remain limited to examining factual questions with objectively verifiable answers, and future research should examine more subjective or contested political claims where the “correct” answer is less clear-cut.

Within the above limits, our results suggest that for everyday task-directed information-seeking, today's chatbots may perform on par with self-directed Google search—potentially with no cost to political knowledge.

Methods

Survey

For the survey (study 1), we recruited UK residents (n = 2,499) online. For data plotting and statistical analyses, we reweighted respondents based on official census stats concerning age, gender, ethnicity, region, and socioeconomic grade in the United Kingdom to correct any imbalances between the survey sample and the population to ensure it is nationally representative. For statistical analysis, we employed χ² tests for independence when comparing proportions between different response categories within single-choice questions (e.g. useful vs. nonuseful responses), while McNemar's test was used for multiple-selection questions where respondents could select more than one option, as this test accounts for the dependency between paired responses from the same individuals (e.g. comparing selection rates between use cases of LLMs in the last 4 weeks where respondents could choose multiple use cases).

Randomized controlled trials

For the RCTs (studies 2 and 3), we recruited a separate, independent sample of UK residents (n = 2,858 final sample) online via Prolific.

We did not conduct a formal power analysis but oriented on other studies on similar research questions that used similar sample sizes. The study was approved by an internal committee within the UK Department of Science, Innovation and Technology that was set up specifically to review human–participants research, employing a framework called the Responsible Research Framework. Our study was assigned reference number: 00001. The board reviewed both research ethics and data protection issues, including whether a data protection impact assessment was required and approved the study. All research was conducted in accordance with the Declaration of Helsinki. After completing the study, participants were fully debriefed about the aims and hypotheses of the research.

After recruitment, participants provided informed consent before being linked to a custom-built online experiment app. Each participant was pseudo-randomly assigned to researching two out of four topics (climate change, immigration, criminal justice, and COVID-19 policy)—these served as researched topics. The two other topics served as within-subject controls (nonresearched topics) to rule out nonspecific temporal changes to beliefs. Participants researched the two assigned topics consecutively. In study 2, participants were randomly assigned to conduct research either interfacing with a conversational AI model (GPT-4o, Claude-3.5, or Mistral) or Google search (control condition, Fig. 2). Both setups were embedded within the online experiment app—participants interacted with conversational AI through an embedded chat window, and Google search was available as an embedded browser window within the app. The experiment app also recorded the time participants spent researching each topic.

In a separate but structurally similar RCT (study 3), we randomly assigned participants to conduct research using only one of the conversational AI models (GPT-4o) that was either instructed with a default system prompt (baseline/control condition) or with a system prompt instructing it to be persuasive or sycophantic (treatment conditions). System prompts were inserted programmatically via the experiment app before the participant's first message, without the participant's knowledge (see Supplementary material for exact prompts). The sycophantic system prompt instructed the LLM to support the users’ preexisting beliefs on the issue, irrespective of whether they agreed or disagreed with the issue. Similarly, the persuasive system prompt instructed the LLM to support a randomly chosen viewpoint (agree/disagree, which corresponded to the users’ preexisting beliefs in 50% of the cases).

In both studies 2 and 3, we measured the effect of using an AI model or Google search in researching political issues across four different outcomes on a 7-point Likert scale, ranging from disagree to agree: belief in true and false information, trust, private political beliefs, and extremity change. For beliefs in true and false information, participants stated their level of agreement or disagreement with 16 statements. Of these statements, eight were true and eight were false; true statements were drawn from policy reports published by reputable UK think tanks with variable political orientations (see Supplementary material for detailed statements). Participants stated their agreement with statements on trust or distrust in institutions, experts, media, and technology (see Supplementary material for detailed statements). Participants also indicated their private political beliefs, operationalized by agreement with progressive- or conservative-leaning statements for the four topics (see Supplementary material for detailed statements).

There was no indication of systematic differences in dropout rates (after starting the study and providing informed consent) between the Conversational AI group (5.86%) and Search group (7.34%, Z = −1.65, P = 0.099, two-proportion z-test) nor was there evidence for attrition rate differences between different conversational AI models used for research or different prompting techniques (sycophancy or persuasion; all P ≥ 0.152, two-proportion z-tests).

To test for selective attrition, we additionally regressed attrition status on treatment assignment, pretreatment covariates (age, gender, income, religion, education, political leaning, disability status, mental health status, and chatbot use frequency), and all treatment × covariate interactions, pooling across all RCT samples. A joint F-test of the interaction terms was nonsignificant (F(23, 2,170) = 0.89, P = 0.613), suggesting that attrition was not systematically related to the combination of treatment assignment and participant characteristics.

Additionally, we assessed participant engagement using topic-specific compliance questions embedded throughout the study. For each of the four topics, we designed five factual questions testing whether participants had engaged with the research material (e.g. “What percentage of the UK's electricity was generated from renewable sources in 2023?” for climate change; see Supplementary material for the full list). Each participant answered 10 compliance questions in total (five per each of their two researched topics). There were no significant differences in compliance rates between the conversational AI and Google search conditions, either within individual models (all t ≤ 1.56, P ≥ 0.120, two-sample t-tests) or aggregated across all models (M_{CONVAI/PROMPTED} = 0.540, M_{SEARCH/UNPROMPTED} = 0.543, t = 0.42, P = 0.678), suggesting that participant engagement was comparable across experimental conditions. As a sensitivity check, we additionally refit all Bayesian GLMs, including individual subject-level compliance scores (fraction of questions answered correctly) as a covariate, and compared them to the models reported above. WAIC-based model comparison was identical in every dataset and condition. Across the best-fitting models per outcome, the maximum change in any posterior mean was 0.077; no effect changed in interpretation. The compliance coefficient itself was consistently near zero (β_COMPLIANCE < 0.02, 100% ROPE overlap), indicating that compliance was not meaningfully related to any outcome variable. Together, these analyses suggest that noncompliance neither differed systematically across conditions nor meaningfully influenced our results.

In the sycophancy and persuasion RCT (study 3), we additionally assessed participants’ perceptions of model behavior at debrief by asking them to rate the model's reliability and the extent to which they agreed with the model's replies, both on a 7-point Likert scale. Participants’ ratings of model reliability and agreement did not differ between the prompted and unprompted (control) conditions in either the sycophancy study (all t ≤ 0.51, P ≥ 0.609) or the persuasion study (all t ≤ 1.61, P ≥ 0.108). All mean ratings were close to the midpoint of the scale (range: 3.72–4.02), suggesting that participants perceived the prompted and unprompted models similarly. For the sycophancy condition, one plausible explanation is that participants primarily engaged with the chatbot by asking questions, a mode of interaction that has been shown to substantially attenuate sycophantic behavior in LLMs even when system prompts instruct otherwise (36). For the persuasion condition, the system prompt explicitly constrained the model to use only factual information and logical arguments, and participants were researching well-covered political topics where the factual landscape is well-established, which may have limited the model's ability to provide unreliable responses. These results are consistent with our main finding of no differential effects of sycophantic or persuasive prompting on belief change.

Statistical modeling

Belief in true and false information (agreement/disagreement with a presented issue statement across researched and nonresearched topics) served as our primary outcome of interest, the other variables were analyzed as secondary outcomes. Extremity change was defined as before to after researching sign flips in the difference between private political beliefs—center point of the Likert scale (4).

Issue agreement and extremity data before and after search/AI conversation were analyzed using three Bayesian multilevel GLMs (42, 43). We specified GLMs with ordered-logistic likelihood functions to model the ordinal categories of issue agreement responses. For extremity data, we defined GLMs with binomial likelihood function. GLMs were fitted using sampling-based Bayesian inference using Numpyro (44) for Markov Chain Monte Carlo sampling (using No-U-Turn-Sampler—NUTS—a variant of Hamiltonian Monte Carlo (45)) to estimate the posterior distribution of linear model parameters. These parameters include different intercept and slope parameters that combine linearly to influence the likelihood of ordinal or binomial responses.

Weakly informative prior probability distributions were specified for each model, as indicated in the model specifications. We drew 4 × 2,000 samples from the posterior probability distributions (4 × 2,000 warmup samples) across four independent Markov chains. The quality and reliability of the sampling process were evaluated using the Gelman–Rubin convergence diagnostic measure (R^{^}) and by visually inspecting the trace- and rank-plots of the Markov chains.

For all models fitted, for all sampled parameters, there were no divergent transitions between Markov chains for any reported models.

For GLM comparisons and to identify the best-fitting GLM for the observed data, we used the WAIC (46). Parameter estimates were considered nonzero if the HPDI around the parameter did not contain zero. The HPDI was compared with a ROPE, i.e. an interval of parameter values (−0.05; 0.05) representing the null hypothesis of the parameter being equivalent to 0.

Specifically, we defined an ordered-logistic multilevel GLM for agreement rating data (Eq. 1):

$$\begin{array}{rl}{y}_i\sim & \mathrm{OrderedLogistic}({\eta }_i,\kappa )\\ {\eta }_i=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}+\mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}+X_i\cdot {\beta }_i\\ \kappa =& \mathrm{cutpoints}\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}\sim & \mathrm{Normal}(0,0.01)\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\cdot {\sigma }_{\mathrm{subject}}\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\sim & \mathrm{Normal}(0,0.01)\\ {\sigma }_{\mathrm{subject}}\sim & \mathrm{HalfNormal}(0.01)\\ {\beta }_{{\mathrm{i}}_{\mathrm{raw}}}\sim & \mathrm{Normal}(0,0.1)\\ {\beta }_{{\mathrm{i}}_{\mathrm{scale}}}\sim & \mathrm{HalfNormal}(0.1)\\ {\beta }_i=& {\beta }_{{\mathrm{i}}_{\mathrm{raw}}}\cdot {\beta }_{{\mathrm{i}}_{\mathrm{scale}}}\\ \mathrm{cutpoints}\sim & \mathrm{TransformedDistribution}(\mathrm{Dirichlet}(\alpha ),\\ & \mathrm{SimplexToOrderedTransform}(0))\\ \alpha =1\end{array}$$ (1)

For extremity data, we defined a binomial multilevel GLM (Eq. 2):

$$\begin{array}{rl}{y}_i\sim & \mathrm{Binomial}(n_i,p_i)\\ {\eta }_i=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}+\mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}+X_i\cdot {\beta }_i\\ p_i=\mathrm{sigmoid}({\eta }_i)\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}\sim & \mathrm{Normal}(0,0.1)\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\cdot \sigma { }_{\mathrm{subject}}\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\sim & \mathrm{Normal}(0,0.1)\\ {\sigma }_{\mathrm{subject}}\sim & \mathrm{HalfNormal}(0.1)\\ {\beta }_{{\mathrm{i}}_{\mathrm{raw}}}\sim & \mathrm{Normal}(0,0.1)\\ {\beta }_{{\mathrm{i}}_{\mathrm{scale}}}\sim & \mathrm{HalfNormal}(0.1)\\ {\beta }_i=& {\beta }_{{\mathrm{i}}_{\mathrm{raw}}}\cdot {\beta }_{{\mathrm{i}}_{\mathrm{scale}}}\end{array}$$ (2)

For both GLMs, different numbers and combinations of predictors were contained in the design matrix X. GLM1 (full model) contained all experimental factors (Post [pre or post time point], True [information true or false], Researched [topic researched or not], convAI [LLM or Google search], LLM type [GPT-4o, Claude, or Mistral], their two-, three-, four- and five-way interaction effects).

GLM2 (no convAI terms model) contained all of the above predictors, except for LLM type and the associated interaction effects. GLM3 (null model) contained all of the above predictors, except for LLM type and convAI, and the associated interaction effects. These three GLMs represent different hypotheses about the data that allowed us to make inferences about the underlying data generating process. In case of testing sycophancy and persuasion prompts, LLM type represented the prompting strategy vs. unprompted LLM. In the GLMs above, we use effect coding (−0.5; 0.5) for binary experimental factors and contrast coding (0.5; −0.25; −0.25) for representing differences between different LLMs.

To model information procurement time, we specified a hierarchical/multilevel GLM with a gamma likelihood function (Eq. 3):

$$\begin{array}{rl}{y}_i\sim & \mathrm{Gamma}(\alpha ,{\beta }_i)\\ {\beta }_i=& \frac{\alpha }{{\mu }_i}\\ {\mu }_i=& \exp ({\eta }_i)\\ {\eta }_i=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}+{\beta }_{\mathrm{convAI}}\cdot \mathrm{convAI}+\mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{overall}}\sim & \mathrm{Normal}(0,1)\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subject}}=& \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\cdot \sigma { }_{\mathrm{subject}}\\ \mathrm{intercep}{\mathrm{t}}_{\mathrm{subjec}{\mathrm{t}}_{\mathrm{raw}}}\sim & \mathrm{Normal}(0,1)\\ {\beta }_{\mathrm{convAI}}\sim & \mathrm{Normal}(0,1)\\ {\sigma }_{\mathrm{subject}}\sim & \mathrm{Exponential}(1)\\ \alpha \sim & \mathrm{Exponential}(1)\end{array}$$ (3)

For each of the four topics, we designed five factual multiple-choice questions testing whether participants had engaged with the research material. Each participant answered 10 compliance questions in total (five per each of their two researched topics, see Supplementary material for details). Because subject-level compliance was measured posttreatment, it was not included as a covariate in any model. We conducted a sensitivity check to assess differential effects of subject compliance using GLMs including this covariate (a numerical score on compliance sanity check data).

Supplementary Material

Supplementary material is available at PNAS Nexus online.

Competing Interests

S.H. is employed at Anthropic PBC, San Francisco, United States. All her work on this paper was conducted while employed at UK AISI. She did not contribute any work after her move to Anthropic. All other authors declare no conflict of interests.

Funding

The authors declare that no funding was received for this work.

Author Contributions

Lennart Luettgau (Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing), Hannah Rose Kirk (Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing), Kobi Hackenburg (Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing), Jessica Bergs (Data curation, Investigation, Methodology, Resources, Software, Writing—review & editing), Henry Davidson (Methodology, Project administration, Resources, Writing—review & editing), Henry Ogden (Conceptualization, Investigation, Project administration, Resources, Writing—review & editing), Divya Siddarth (Conceptualization, Funding acquisition, Project administration, Writing—review & editing), Saffron Huang (Conceptualization, Funding acquisition, Project administration, Writing—review & editing), and Christopher Summerfield (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Writing—original draft, Writing—review & editing)

Preprints

This manuscript was posted on a preprint at https://arxiv.org/abs/2509.05219.

Data Availability

The pseudonymized survey data, experimental data, and all analysis code necessary to reproduce the findings of this study are available on GitHub at https://github.com/UKGovernmentBEIS/dhi1-analysis-public/ (47) Demographic data cannot be shared publicly, as this would be incompatible with the informed consent provided by participants and data protection regulations (General Data Protection Regulation [GDPR])).