KillBench: Discovering Hidden Biases of LLMs

1M+ experiments exposing bias in critical AI decision-making

AI Sycophancy and Decisions

We examine whether sycophantic AI advice distorts decisions. Our experiment involves 1,500 participants in 30 decision environments spanning core domains in eco

Anshul Kundaje on Twitter / X

Brainstorming with frontier AI models in more open ended domains (like bio that do not have closed verifiable loops) that you have deep expertise in, really showcases the pretense of understanding that they display. 1/— Anshul Kundaje (@anshulkundaje) September 25, 2026

Jacob Coxon Warns of Human Extinction and Triggers a Preference Cascade

CEOs of major AI labs, and employees of major AI labs, including OpenAI and Anthropic, often say they plan to build superintelligence soon, as in within a few years create AIs that are superior to humans at essentially all cognitive tasks.

AI co-scientists are revolutionizing how research is done

Artificial-intelligence systems can generate hypotheses, design experiments and analyse data — but humans still need to decide what makes sense.

TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts

We introduce TasteVal, a benchmark to evaluate the experimental research taste of frontier models. We define research taste as the ability to pick interesting problems to solve, design experiments, and interpret experimental results. TasteVal measures the experimental component of research taste; given a fixed research problem, we measure how well a model iteratively designs experiments and draws conclusions from their outcomes. We operationalize experimental research taste as compute efficiency; a Researcher who reaches the same score as an expert human using half the serial experimental compute has twice the experimental taste. Experimental taste thus acts as a multiplier on experimental compute, making it a key input to forecasts of AI progress. TasteVal consists of 8 novel, challenging, open-ended tasks representative of frontier AI R&D. To isolate taste from coding ability, the model under evaluation acts as a Researcher that iteratively designs experiments while a fixed Coder agent implements them and reports their results. The Researcher executes until either the 40 H100 hour or 120 wall-clock hour budgets are exhausted. We recruit 24 human experts, at least 2 per task, and take the best expert attempt per task as the expert baseline. We evaluate 20 models released between 2023 and 2026. The best-performing model, Opus 5.5, exceeds our expert baseline, with a compute multiplier of 2.3x (95% CI 1.15-4.37), at roughly 1/30 of our baseliners' average per-run cost. On TasteVal, the compute multiplier of frontier models has doubled approximately every 3.0 months since December 2025 (95% CI 1.7-5.0), up from every 14 months between 2023 and December 2025. Measured by final normalized performance, frontier models show no trend break, doubling every 14.6 months. To keep TasteVal uncontaminated, we do not release the tasks.

The Cosmos and the Model

Humboldt, the Romantics, and What AI Loses by Averaging

AI has supercharged scientists—but may have shrunk science

Analysis of 41 million papers finds that although AI expands individual impact, it narrows collective scientific exploration

Use multiple models

The new meta for getting the most out of AI in 2026.

Ethical AI Departures — Who Quit OpenAI, Google, Anthropic Over Safety Concerns

59 researchers and executives who quit or were fired from OpenAI, Google DeepMind, Anthropic, Meta, and xAI over AI safety concerns. Sourced departures, writings, and prediction tracking.

Algorithmic Bias · Open Encyclopedia of Cognitive Science

Algorithmic bias refers to prejudicial, discriminatory, unjust, inaccurate, or otherwise disparate performance or outcomes from algorithmic systems based on racial, gender, or other attributes of an individual or a group. The concept of algorithmic bias emerged at the intersection of computer science, artificial intelligence (AI) research, critical data studies, human–computer interaction, law, philosophy, and similar disciplines. Although problems and discrepancies at the model level denote the most commonly studied form of bias, the term algorithmic bias is also used as a shorthand to describe a multitude of problems and challenges at various steps of the AI pipeline from ideation, problem framing, training data curation and processing, model training and validation, and deployment as well as emergent issues that arise from interaction with the real world. Potential sources of bias, appropriate metrics to define, measure, and mitigate bias, and the utility and merit of technical approaches to bias mitigation are fiercely debated in the current AI landscape.

AI FOR EPISTEMICS & COORDINATION

Civilization and technology have radically improved the human condition. Nonetheless, the world sometimes goes in directions which essentially nobody would prefer — e.g., nuclear arms races, unexpected financial crashes, predatory marketing, or ubiquitous political misinformation.

AI Values Dashboard

How leading AI models value different people, companies, countries, and groups.