There is substantial concern about the ability of advanced artificial intelligence to influence people's behaviour. A rapidly growing body of research has found that AI can produce large persuasive effects on people's attitudes, but whether AI can persuade people to take consequential real-world actions has remained unclear. In two large preregistered experiments N=17,950 responses from 14,779 people), we used conversational AI models to persuade participants on a range of attitudinal and behavioural outcomes, including signing real petitions and donating money to charity. We found sizable AI persuasion effects on these behavioural outcomes (e.g. +19.7 percentage points on petition signing). However, we observed no evidence of a correlation between AI persuasion effects on attitudes and behaviour. Moreover, we replicated prior findings that information provision drove effects on attitudes, but found no such evidence for our behavioural outcomes. In a test of eight behavioural persuasion strategies, all outperformed the most effective attitudinal persuasion strategy, but differences among the eight were small. Taken together, these results suggest that previous findings relying on attitudinal outcomes may generalize poorly to behaviour, and therefore risk substantially mischaracterizing the real-world behavioural impact of AI persuasion.

Advances in artificial intelligence pose risks to humanity's collective capacity to form accurate beliefs, reason well, and maintain a healthy information environment—risks we term epistemic risks. These risks arise from AI's integration into the infrastructure through which individuals and societies think, form beliefs, and make sense of the world together. We highlight three primary mechanisms by which AI can cause systemic epistemic decline: persuasion and manipulation, cognitive offloading, and feedback loops. First, AI can persuade and manipulate. It can be misused for economic or political manipulation, inciting crime or radicalization, and escalating conflict. It can also create unintentional harms, e.g. sycophancy and mental health risks. Second, AI's unique features enable cognitive offloading in a qualitatively different way from previous technologies, risking cognitive decline. Third, human-AI and AI-AI feedback loops can narrow the epistemic space which humans and AI systems draw from. This already drives homogenization and may potentially lead to fragmentation. We record the emerging empirical evidence for each mechanism and analyze how they likely amplify one another, creating compounding systemic risk. This evidence suggests that AI could be an unprecedented lever for improving epistemics, but this will not happen by default. Finally, we outline cross-cutting solutions which span AI system design, human-AI interaction, institutional reform, and wider incentive structures. Our paper ends with urgency: humanity's ability to navigate AI and non-AI risks alike may depend on how epistemically resilient our institutions are. It may be critical to act before that capacity is eroded.

Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has remained unclear. Here, in a series of four preregistered experiments (n = 18,978 conversations from 6,923 people), we pitted AI systems against a range of human persuaders, including laypeople, winners of a separately preregistered four-round online persuasion tournament, professional canvassers, and world championship debaters. We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses. In a follow-up study, AI's advantage persisted after experts received a coaching tool that let them practice against the AI that beat them, review their performance history, and see what AI would have said at key moments. We found converging evidence that AI's advantage stemmed from rapidly deploying larger quantities of information: after coaching, expert humans could tie an AI constrained to respond at human speeds and with human-length messages. In a final study, we show that AI's advantage extends to consequential real-world behavior: AI was nearly 3x more effective than professional canvassers from a UK fundraising firm at raising real-money donations to Save the Children. Together, these results establish that frontier AI systems out-persuade expert humans in conversation, with significant implications for political communication.

There is growing interest in how large language models (LLMs) can advance social and behavioral science [1–5]. Prior work has assessed LLMs’ ability to predict survey responses [6–9], but less is known about whether they can predict the outcomes of social science experiments [10], particularly those absent from training data. Here, we built an archive of 70 pre-registered, nationally representative, U.S. survey experiments, involving 469 experimental effects and 119,330 participants. We prompted an LLM to simulate how representative samples of Americans would respond to experimental stimuli, then inferred treatment effects by comparing simulated responses across conditions. Predictions derived from GPT-4, whose training-data cutoff predated the publication of many studies in our archive, were strongly correlated with actual treatment effects, achieving accuracy similar to pooled human forecasts. Correlations remained high for studies not published or publicly posted by the model’s training-data cutoff date, and for predictions from prominent open-weight models. Despite high correlations, predictions systematically overestimated effect sizes. In a secondary archive of 15 megastudies featuring 606 effects, correlations were lower but comparable to pooled expert forecasters. To assess implications for scientific practice, we surveyed 460 social scientists about likely uses and perceived risks, and used our archives to assess several applications (pilot testing, intervention selection, identifying effects needing replication) and risks (bias, misuse). Together, these results suggest LLMs can augment experimental methods in science and practice while raising important considerations for responsible use.

There are widespread fears that conversational AI could soon exert unprecedented influence over human beliefs. Here, in three large-scale experiments (N=76,977), we deployed 19 LLMs-including some post-trained explicitly for persuasion-to evaluate their persuasiveness on 707 political issues. We then checked the factual accuracy of 466,769 resulting LLM claims. Contrary to popular concerns, we show that the persuasive power of current and near-future AI is likely to stem more from post-training and prompting methods-which boosted persuasiveness by as much as 51

There are widespread fears that conversational artificial intelligence (AI) could soon exert unprecedented influence over human beliefs. In this work, in three large-scale experiments (N = 76,977 participants), we deployed 19 large language models (LLMs)-including some post-trained explicitly for persuasion-to evaluate their persuasiveness on 707 political issues. We then checked the factual accuracy of 466,769 resulting LLM claims. We show that the persuasive power of current and near-future AI is likely to stem more from post-training and prompting methods-which boosted persuasiveness by as much as 51 and 27%, respectively-than from personalization or increasing model scale, which had smaller effects. We further show that these methods increased persuasion by exploiting LLMs' ability to rapidly access and strategically deploy information and that, notably, where they increased AI persuasiveness, they also systematically decreased factual accuracy.

Large Language Models (LLMs) have been shown to be highly persuasive, but when and why they outperform humans is still an open question. We compare the persuasiveness of two LLMs (Claude 3.5 Sonnet and DeepSeek v3) against humans who had incentives to persuade, using an interactive, real-time conversational setting. We demonstrate that LLMs persuasive superiority is context-dependent: it depends on whether the persuasion attempt is truthful (towards the right answer) or deceptive (towards the wrong answer) and on the LLM model, and wanes over repeated interactions (unlike human persuasiveness). In our first large-scale experiment, humans vs LLMs (Claude 3.5 Sonnet) interacted with other humans who were completing an online quiz for a reward, attempting to persuade them toward a given (either correct or incorrect) answer. Claude was more persuasive than incentivized human persuaders both in truthful and deceptive contexts and it significantly increased accuracy if persuasion was truthful, but decreased it if persuasion was deceptive. In a follow-up experiment with Deepseek v3, we replicated the findings about accuracy but found greater LLM persuasiveness only if the persuasion was deceptive. Linguistic analyses of the persuaders texts suggest that these effects may be due to LLMs expressing higher conviction than humans.

Advanced artificial intelligence (AI) systems capable of generating humanlike text and multimodal content are now widely available. Here we ask what impact this will have on the democratic process. We consider the consequences of AI for citizens' ability to make educated and competent choices about political representatives and issues (epistemic impacts). We explore how AI might be used to destabilize or support the mechanisms, including elections, by which democracy is implemented (material impacts). Finally, we discuss whether AI will strengthen or weaken the principles on which democracy is based (foundational impacts). The arrival of new AI systems clearly poses substantial challenges for democracy. However, we argue that AI systems also offer new opportunities to educate and learn from citizens, strengthen public discourse, help people to find common ground, and reimagine how democracies might work better.

The world could witness another pandemic on the scale of COVID-19 in the future, prompting calls for research into how social and behavioral science can better contribute to pandemic response, especially regarding public engagement and communication. Here, we conduct a cost-effectiveness analysis of a familiar tool from social and behavioral science that could potentially increase the impact of public communication: survey experiments. Specifically, we analyze whether a public health campaign that pays for a survey experiment to pretest and choose between different messages for its public outreach has greater impact in expectation than an otherwise-identical campaign that does not. The main results of our analysis are 3-fold. First, we show that the benefit of such pretesting depends heavily on the values of several key parameters. Second, via simulations and an evidence review, we find that a campaign that allocates some of its budget to pretesting could plausibly increase its expected impact; that is, we estimate that pretesting is cost-effective. Third, we find pretesting has potentially powerful returns to scale; for well-resourced campaigns, we estimate pretesting is robustly cost-effective, a finding that emphasizes the benefit of public health campaigns sharing resources and findings. Our results suggest survey experiment pretesting could cost-effectively increase the impact of public health campaigns in a pandemic, have implications for practice, and establish a research agenda to advance knowledge in this space.

Perception has long been envisioned to use an internal model of the world to explain the causes of sensory signals. However, such accounts have historically not been testable, typically requiring intractable search through the space of possible explanations. Using auditory scenes as a case study, we leveraged contemporary computational tools to infer explanations of sounds in a candidate internal generative model of the auditory world (ecologically inspired audio synthesizers). Model inferences accounted for many classic illusions. Unlike traditional accounts of auditory illusions, the model is applicable to any sound, and exhibited human-like perceptual organization for real-world sound mixtures. The combination of stimulus-computability and interpretable model structure enabled ‘rich falsification’, revealing additional assumptions about sound generation needed to account for perception. The results show how generative models can account for the perception of both classic illusions and everyday sensory signals, and illustrate the opportunities and challenges involved in incorporating them into theories of perception.

Political campaigns increasingly conduct experiments to learn how to persuade voters. Little research has considered the implications of this trend for elections or democracy. To probe these implications, we analyze a unique archive of 146 advertising experiments conducted by US campaigns in 2018 and 2020 using the platform Swayable. This archive includes 617 advertisements produced by 51 campaigns and tested with over 500,000 respondents. Importantly, we analyze the complete archive, avoiding publication bias. We find small but meaningful variation in the persuasive effects of advertisements. In addition, we find that common theories about what makes advertising persuasive have limited and context-dependent power to predict persuasiveness. These findings indicate that experiments can compound money's influence in elections: it is difficult to predict ex ante which ads persuade, experiments help campaigns do so, but the gains from these findings principally accrue to campaigns well-financed enough to deploy these ads at scale.

GPT-4o is an autoregressive omni model that accepts as input any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs. It's trained end-to-end across text, vision, and audio, meaning all inputs and outputs are processed by the same neural network. GPT-4o can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response time in conversation. It matches GPT-4 Turbo performance on text in English and code, with significant improvement on text in non-English languages, while also being much faster and 50% cheaper in the API. GPT-4o is especially better at vision and audio understanding compared to existing models. In line with our commitment to building AI safely and consistent with our voluntary commitments to the White House, we are sharing the GPT-4o System Card, which includes our Preparedness Framework evaluations. In this System Card, we provide a detailed look at GPT-4o's capabilities, limitations, and safety evaluations across multiple categories, focusing on speech-to-speech while also evaluating text and image capabilities, and measures we've implemented to ensure the model is safe and aligned. We also include third-party assessments on dangerous capabilities, as well as discussion of potential societal impacts of GPT-4o's text and vision capabilities.

Advanced AI systems capable of generating humanlike text and multimodal content are now widely available. In this paper, we discuss the impacts that generative artificial intelligence may have on democratic processes. We consider the consequences of AI for citizens' ability to make informed choices about political representatives and issues (epistemic impacts). We ask how AI might be used to destabilise or support democratic mechanisms like elections (material impacts). Finally, we discuss whether AI will strengthen or weaken democratic principles (foundational impacts). It is widely acknowledged that new AI systems could pose significant challenges for democracy. However, it has also been argued that generative AI offers new opportunities to educate and learn from citizens, strengthen public discourse, help people find common ground, and to reimagine how democracies might work better.

Perhaps hundreds of survey experiments have shown that political party cues influence people’s policy opinions. However, we know little about the persistence of this influence: is it a transient priming effect, dissipating moments after the survey is over; or does influence persist for longer, indicating learning? We report the results of a panel survey experiment in which U.S. adults were randomly exposed to party cues on five contemporary U.S. policy issues in an initial survey, and gave their opinions. A follow-up survey three days later polled their opinions again. We find that the influence of the party cues persists at ~50% its original magnitude at follow-up. Notably, our design rules out that people simply remembered how they previously answered. Our findings have implications for understanding the scope and mechanism of party cue influence as it occurs in the real world, and provide a benchmark for future research on this topic.

Abstract Perception has long been envisioned to use an internal model of the world to explain the causes of sensory signals. However, such accounts have historically not been testable, typically requiring intractable search through the space of possible explanations. Using auditory scenes as a case study, we leveraged contemporary computational tools to infer explanations of sounds in a candidate internal model of the auditory world (ecologically inspired audio synthesizers). Model inferences accounted for many classic illusions. Unlike traditional accounts of auditory illusions, the model is applicable to any sound, and exhibited human-like perceptual organization for real world sound mixtures. The combination of stimulus-computability and interpretable model structure enabled ‘rich falsification’, revealing additional assumptions about sound generation needed to account for perception. The results show how generative models can account for the perception of both classic illusions and everyday sensory signals, and provide the basis on which to build theories of perception.

Much concern has been raised about the power of political microtargeting to sway voters’ opinions, influence elections, and undermine democracy. Yet little research has directly estimated the persuasive advantage of microtargeting over alternative campaign strategies. Here, we do so using two studies focused on U.S. policy issue advertising. To implement a microtargeting strategy, we combined machine learning with message pretesting to determine which advertisements to show to which individuals to maximize persuasive impact. Using survey experiments, we then compared the performance of this microtargeting strategy against two other messaging strategies. Overall, we estimate that our microtargeting strategy outperformed these strategies by an average of 70% or more in a context where all of the messages aimed to influence the same policy attitude (Study 1). Notably, however, we found no evidence that targeting messages by more than one covariate yielded additional persuasive gains, and the performance advantage of microtargeting was primarily visible for one of the two policy issues under study. Moreover, when microtargeting was used instead to identify which policy attitudes to target with messaging (Study 2), its advantage was more limited. Taken together, these results suggest that the use of microtargeting—combining message pretesting with machine learning—can potentially increase campaigns’ persuasive influence and may not require the collection of vast amounts of personal data to uncover complex interactions between audience characteristics and political messaging. However, the extent to which this approach confers a persuasive advantage over alternative strategies likely depends heavily on context.

From sparse descriptions of events, observers can make systematic and nuanced predictions of what emotions the people involved will experience. We propose a formal model of emotion prediction in the context of a public high-stakes social dilemma. This model uses inverse planning to infer a person’s beliefs and preferences, including social preferences for equity and for maintaining a good reputation. The model then combines these inferred mental contents with the event to compute ‘appraisals’: whether the situation conformed to the expectations and fulfilled the preferences. We learn functions mapping computed appraisals to emotion labels, allowing the model to match human observers’ quantitative predictions of 20 emotions, including joy, relief, guilt and envy. Model comparison indicates that inferred monetary preferences are not sufficient to explain observers’ emotion predictions; inferred social preferences are factored into predictions for nearly every emotion. Human observers and the model both use minimal individualizing information to adjust predictions of how different people will respond to the same event. Thus, our framework integrates inverse planning, event appraisals and emotion concepts in a single computational model to reverse-engineer people’s intuitive theory of emotions. This article is part of a discussion meeting issue ‘Cognitive artificial intelligence’.

Expert problem-solving is driven by powerful languages for thinking about problems and their solutions. Acquiring expertise means learning these languages-systems of concepts, alongside the skills to use them. We present DreamCoder, a system that learns to solve problems by writing programs. It builds expertise by creating domain-specific programming languages for expressing domain concepts, together with neural networks to guide the search for programs within these languages. A 'wake-sleep' learning algorithm alternately extends the language with new symbolic abstractions and trains the neural network on imagined and replayed problems. DreamCoder solves both classic inductive programming tasks and creative tasks such as drawing pictures and building scenes. It rediscovers the basics of modern functional programming, vector algebra and classical physics, including Newton's and Coulomb's laws. Concepts are built compositionally from those learned earlier, yielding multilayered symbolic representations that are interpretable and transferrable to new tasks, while still growing scalably and flexibly with experience. This article is part of a discussion meeting issue 'Cognitive artificial intelligence'.

Modeling complex phenomena typically involves the use of both discrete and continuous variables. Such a setting applies across a wide range of problems, from identifying trends in time-series data to performing effective compositional scene understanding in images. Here, we propose Hybrid Memoised Wake-Sleep (HMWS), an algorithm for effective inference in such hybrid discrete-continuous models. Prior approaches to learning suffer as they need to perform repeated expensive inner-loop discrete inference. We build on a recent approach, Memoised Wake-Sleep (MWS), which alleviates part of the problem by memoising discrete variables, and extend it to allow for a principled and effective way to handle continuous variables by learning a separate recognition model used for importance-sampling based approximate inference and marginalization. We evaluate HMWS in the GP-kernel learning and 3D scene understanding domains, and show that it outperforms current state-of-the-art inference methods.

Central to theories of political persuasion is treatment effect heterogeneity—the idea that people respond to political messages in different ways—so persuasion is easier when different messages are targeted to different audiences. The standard approach to testing for heterogeneity is to examine whether the effect of an individual message differs between subgroups of people (such as liberals versus conservatives). We describe the shortcomings of this approach, and propose an alternative: jointly examining many messages on the same political issue, and assessing whether the rank-order of their effects differs between subgroups (which we call “rank-heterogeneity”). Implementing this approach, we conduct two large-scale survey experiments spanning two policy issues, 59 message treatments, and over 40,000 American adults. Across experiments we find mixed evidence of rank-heterogeneity, suggesting that it depends upon the particular issue in question. However, in the case where we do observe strong evidence of rank-heterogeneity, its primary cause is consistent with the predictions of moral reframing theory, an influential account of heterogeneity in political persuasion. Alongside these implications for theory, our results have implications for political persuasion in practice.