Talal Rahwan

Language models are becoming an ordinary way to ask a political question, and there is widespread concern that their answers lean left. Nearly all evidence for this concern is national: each audit scores a model against a single country's survey respondents, most often American ones, and the resulting score confounds where the model stands with where the sampled population stands. Here we remove this confound by scoring models on item parameters estimated from 1.14 million human survey respondents in 141 countries and then held fixed, placing models, respondents and countries on one metric anchored to people rather than to any national sample. We pose 1,153 harmonized survey items to 65 models from 27 developers, at temperature zero and with no persona, and score the answers on four latent dimensions: social liberalism, economic left-right, democratic commitment and anti-elite sentiment. The quantity this measures is expressed ideology, the positions a model states when asked, rather than any claim about what these systems hold. We find, first, that all 65 models sit to the liberal side of the world human median on social liberalism, with the median model at the 31st percentile, and that the lean runs weaker in the same direction on the other three dimensions. We find, second, that the models resemble one another far more than people do: across three orders of magnitude in parameter count, and including open-weight, proprietary, reasoning-tuned and deliberately uncensored variants, the standard deviation of model positions is 11 to 33 percent of the human standard deviation on every dimension, and differently post-trained versions of the same base model stay within that band. Two further results caution against reading any single number too closely: changing which survey program supplies the items moves a model's measured position by 0.28 to 0.62 human standard deviations, two to nearly four times what the same change does to a country and more than the entire model-human gap on every dimension, and with the number of answered items held equal on both sides, the difference between models and human respondents in how consistently they place themselves on one ideological side shrinks to a modest residual. The direction of the liberal lean is robust across instruments; its magnitude is a property of the model and the questionnaire together.

Over a quarter of the global population remains without access to the Internet in 2026, primarily due to cost and limited infrastructure in developing regions. This gap is becoming more consequential as access to information increasingly depends on AI systems that presume reliable, high-bandwidth connectivity. In response, we present WebFM, a low-cost, scalable Web and AI access solution that builds on widely deployed infrastructure in developing regions: FM radio for downlink broadcasting and SMS for personalized uplink. Prior work on web-over-FM is limited to simulations, short-range experiments, and lossless links, leaving the effects of geographically induced signal variability largely unexplored. WebFM addresses this gap through several system-level innovations to effectively transmit Web content and enable LLM interactions over sound over FM radio, in a reliable and compressed form. For example, we introduce a loss-resilient file format for encoding webpages and modify Android to leverage built-in FM tuners, allowing deployment across a wide range of devices without requiring root access. We enable ChatGPT interactions by allowing users to submit prompts over SMS and receive model responses via FM broadcast. Finally, we preemptively transmit popular pages and a subset of their internal pages while the server is idle, improving user experience during peak hours. We deployed WebFM at an FM radio station in Cameroon for six weeks with 30 participants. Our evaluation shows a sustained downlink throughput of 10 kbps, less than 20% loss for most transmissions with signal strength above -90 dBm, and strong user engagement across both web browsing and ChatGPT interactions.

Social media platforms increasingly mediate political information exposure, yet the role of algorithmic curation in shaping political exposure remains contested ^1,2 . This question is difficult to resolve on platforms in which users retain substantial control over their feeds ^3,4 . The ‘For You’ feed of TikTok, which delivers content almost entirely through algorithmic recommendation, offers a setting in which user agency is sharply constrained. Here we show, through 323 audit experiments with controlled ‘sock puppet’ accounts seeded with Democratic or Republican content across three US states, that accounts seeded with partisan content exhibited systematic, asymmetric differences in partisan exposure. Across more than 280,000 recommendations collected over 27 weeks during the 2024 US presidential election campaign, Republican-seeded accounts received about 11.5% more co-partisan content than Democratic-seeded accounts, whereas Democratic-seeded accounts were exposed to about 7.5% more cross-partisan content—largely anti-Democratic material—even after adjusting for engagement metrics. These asymmetries are concentrated among high-reach Republican channels and in specific policy domains, including immigration, crime and foreign policy for Democrats, and abortion for Republicans. Our findings show partisan imbalances in political information exposure on a platform dominated by algorithmic recommendations, with implications for platform governance and democratic discourse.

Despite rapid technological progress, effective human–machine cooperation remains a significant challenge. Humans tend to cooperate less with machines than with fellow humans, a phenomenon known as the machine penalty. Here, we show that artificial intelligence (AI) agents powered by large language models can overcome this penalty in social dilemma games with communication. In a pre-registered experiment with 1152 participants, we deploy AI agents exhibiting three distinct personas: selfish, cooperative and fair. However, only fair agents elicit human cooperation at rates comparable to human–human interactions. Analysis reveals that fair agents, similar to human participants, occasionally break pre-game cooperation promises but nonetheless effectively establish cooperation as a social norm. These results challenge the conventional wisdom of machines as altruistic assistants or rational actors. Instead, our study highlights the importance of AI agents reflecting the nuanced complexity of human social behaviors—imperfect yet driven by deeper social cognitive processes.

Among populations with limited literacy in emerging digital markets, the adoption of mobile phones, combined with comprehension barriers and poor cybersecurity hygiene, has created hidden privacy risks. This paper examines how informed consent is often abused by predatory financial applications, leading to financial scams that disproportionately affect users with low literacy. We focus on predatory loan, gambling, and trading apps, analyzing a dataset of 50 Google Play Store apps to measure how many omit or obfuscate critical privacy disclosures. We also evaluate comprehension gaps among users with low literacy via a targeted user study and assess whether Large Language Model (LLM)-generated summaries, translations, and visual cues can improve consent clarity. Our findings show that 85% of study participants did not understand basic app permissions, underscoring the urgent need for stronger regulatory oversight and scalable LLM-driven privacy-literacy tools.

Detecting rare and diverse anomalies in highly imbalanced datasets-such as Advanced Persistent Threats (APTs) in cybersecurity-remains a fundamental challenge for machine learning systems. Active learning offers a promising direction by strategically querying an oracle to minimize labeling effort, yet conventional approaches often fail to exploit the intrinsic geometric structure of the feature space for model refinement. In this paper, we introduce SDA2E, a Sparse Dual Adversarial Attention-based AutoEncoder designed to learn compact and discriminative latent representations from imbalanced, high-dimensional data. We further propose a similarity-guided active learning framework that integrates three novel strategies to refine decision boundaries efficiently: mormal-like expansion, which enriches the training set with points similar to labeled normals to improve reconstruction fidelity; anomaly-like prioritization, which boosts ranking accuracy by focusing on points resembling known anomalies; and a hybrid strategy that combines both for balanced model refinement and ranking. A key component of our framework is a new similarity measure, Normalized Matching 1s (SIM_NM1), tailored for sparse binary embeddings. We evaluate SDA2E extensively across 52 imbalanced datasets, including multiple DARPA Transparent Computing scenarios, and benchmark it against 15 state-of-the-art anomaly detection methods. Results demonstrate that SDA2E consistently achieves superior ranking performance (nDCG up to 1.0 in several cases) while reducing the required labeled data by up to 80% compared to passive training. Statistical tests confirm the significance of these improvements. Our work establishes a robust, efficient, and statistically validated framework for anomaly detection that is particularly suited to cybersecurity applications such as APT detection.