Arvind Narayanan

How does generative artificial intelligence (AI) work? Where does it work well, and where does it falter? In this case study, we go over the mechanics underlying the successes of deep learning in generating images and text. We then turn to the harms of generative AI, including the lack of factuality, misuses such as non-consensual deepfakes, and labor exploitation. We conclude with a discussion of how we can mitigate the harms of generative AI through technical and structural changes.

AI agents have the potential to aid users on a variety of consequential tasks, including conducting scientific research. To spur the development of useful agents, we need benchmarks that are challenging, but more crucially, directly correspond to real-world tasks of interest. This paper introduces such a benchmark, designed to measure the accuracy of AI agents in tackling a crucial yet surprisingly challenging aspect of scientific research: computational reproducibility. This task, fundamental to the scientific process, involves reproducing the results of a study using the provided code and data. We introduce CORE-Bench (Computational Reproducibility Agent Benchmark), a benchmark consisting of 270 tasks based on 90 scientific papers across three disciplines (computer science, social science, and medicine). Tasks in CORE-Bench consist of three difficulty levels and include both language-only and vision-language tasks. We provide an evaluation system to measure the accuracy of agents in a fast and parallelizable way, saving days of evaluation time for each run compared to a sequential implementation. We evaluated two baseline agents: the general-purpose AutoGPT and a task-specific agent called CORE-Agent. We tested both variants using two underlying language models: GPT-4o and GPT-4o-mini. The best agent achieved an accuracy of 21% on the hardest task, showing the vast scope for improvement in automating routine scientific tasks. Having agents that can reproduce existing work is a necessary step towards building agents that can conduct novel research and could verify and improve the performance of other research agents. We hope that CORE-Bench can improve the state of reproducibility and spur the development of future research agents.

Machine learning (ML) methods are proliferating in scientific research. However, the adoption of these methods has been accompanied by failures of validity, reproducibility, and generalizability. These failures can hinder scientific progress, lead to false consensus around invalid claims, and undermine the credibility of ML-based science. ML methods are often applied and fail in similar ways across disciplines. Motivated by this observation, our goal is to provide clear recommendations for conducting and reporting ML-based science. Drawing from an extensive review of past literature, we present the REFORMS checklist (recommendations for machine-learning-based science). It consists of 32 questions and a paired set of guidelines. REFORMS was developed on the basis of a consensus of 19 researchers across computer science, data science, mathematics, social sciences, and biomedical sciences. REFORMS can serve as a resource for researchers when designing and implementing a study, for referees when reviewing papers, and for journals when enforcing standards for transparency and reproducibility.

We formalize predictive optimization, a category ofdecision-making algorithmsthatuse machine learning (ML)topredict future outcomesof interest aboutindividuals. For example, pre-trial risk prediction algorithms such as COMPAS use ML to predict whether an individual will re-offend in the future. Our thesis is that predictive optimization raises a distinctive and serious set of normative concerns that cause it to fail on its own terms. To test this, we review 387 reports, articles, and web pages from academia, industry, non-profits, governments, and data science contests, and we find many real-world examples of predictive optimization. We select eight particularly consequential examples as case studies. Simultaneously, we develop a set of normative and technical critiques that challenge the claims made by the developers of these applications—in particular, claims of increased accuracy, efficiency, and fairness. Our key finding is that these critiques apply to each of the applications, are not easily evaded by redesigning the systems, and thus challenge whether these applications should be deployed. We argue that the burden of evidence for justifying why the deployment of predictive optimization is not harmful should rest with the developers of the tools. Based on our analysis, we provide a rubric of critical questions that can be used to deliberate or contest specific predictive optimization applications.¹

Is AI set to redefine the legal profession? We argue that this claim is not supported by the current evidence. We dive into AI's increasingly prevalent roles in three types of legal tasks: information processing; tasks involving creativity, reasoning, or judgment; and predictions about the future. We find that the ease of evaluating legal applications varies greatly across legal tasks, based on the ease of identifying correct answers and the observability of information relevant to the task at hand. Tasks that would lead to the most significant changes to the legal profession are also the ones most prone to overoptimism about AI capabilities, as they are harder to evaluate. We make recommendations for better evaluation and deployment of AI in legal contexts.

We formalize predictive optimization, a category of decision-making algorithms that use machine learning (ML) to predict future outcomes of interest about \ individuals}. For example, pre-trial risk prediction algorithms such as COMPAS use ML to predict whether an individual will re-offend in the future. Our thesis is that predictive optimization raises a distinctive and serious set of normative concerns that cause it to fail on its own terms. To test this, we review 387 reports, articles, and web pages from academia, industry, non-profits, governments, and modeling contests, and find many real-world examples of predictive optimization. We select eight particularly consequential examples as case studies. Simultaneously, we develop a set of normative and technical critiques that challenge the claims made by the developers of these applications—in particular, claims of increased accuracy, efficiency, and fairness. Our key finding is that these critiques apply to each of the applications, are not easily evaded by redesigning the systems, and thus challenge whether these applications should be deployed. We argue that the burden of evidence for justifying why the deployment of predictive optimization is not harmful should rest with the developers of the tools. Based on our analysis, we provide a rubric of critical questions that can be used to deliberate or contest specific predictive optimization applications.

We describe a series of security policy audits that we conducted, exposing policy flaws affecting billions of users that are often exploited by even low-tech attackers. We argue that a systematic study of security policies and processes is sorely needed, and present a research agenda.

We collect and analyze a corpus of more than 300,000 political emails sent during the 2020 US election cycle. These emails were sent by over 3000 political campaigns and organizations including federal and state level candidates as well as Political Action Committees. We find that in this corpus, manipulative tactics—techniques using some level of deception or clickbait—are the norm, not the exception. We measure six specific tactics senders use to nudge recipients to open emails. Three of these tactics—“dark patterns”—actively deceive recipients through the email user interface, for example, by formatting “from:” fields so that they create the false impression the message is a continuation of an ongoing conversation. The median active sender uses such tactics 5% of the time. The other three tactics, like sensationalistic clickbait—used by the median active sender 37% of the time—are not directly deceptive, but instead, exploit recipients’ curiosity gap and impose pressure to open emails. This can further expose recipients to deception in the email body, such as misleading claims of matching donations. Furthermore, by collecting emails from different locations in the US, we show that senders refine these tactics through A/B testing. Finally, we document disclosures of email addresses between senders in violation of privacy policies and recipients’ expectations. Cumulatively, these tactics undermine voters’ autonomy and welfare, exacting a particularly acute cost for those with low digital literacy. We offer the complete corpus of emails at https://electionemails2020.org for journalists and academics, which we hope will support future work.