AI progress vs humans
Recent syntheses say AI development is accelerating rapidly, but top human scientists still outperform the best AI agents on complex, multi‑step scientific tasks. (technologyreview.com) That combination suggests current AI systems are powerful for many workflows but not yet dominant on high‑ambiguity problems requiring judgement. (nature.com)
Artificial intelligence systems are getting better fast, but the best human scientists still beat the best AI agents on long, messy research tasks. (technologyreview.com) (nature.com) A useful way to read that split is to separate speed from judgment. Stanford’s 2026 AI Index, published in April, collects charts showing rapid gains in model performance, spending, and adoption across the industry. (technologyreview.com) Another way to measure progress is task length: how long a job a system can finish before it breaks down. Model Evaluation and Threat Research, or METR, reported in March 2025 that this “task-completion time horizon” had been doubling about every seven months over the prior six years. (metr.org) (arxiv.org) METR’s updated public tracker, last updated on March 3, 2026, says the metric estimates the duration of human-expert work an AI agent can complete at a given reliability level. The group’s chart covers frontier models from 2019 through November 2025 and shows gains in both its 50% and 80% success measures. (metr.org) That does not mean an agent can run a research program the way a principal investigator or senior postdoctoral researcher does. Nature reported on April 13, 2026 that top human scientists still “trounce” the best AI agents on complex tasks, even as researchers have widely adopted the tools. (nature.com) The gap shows up most clearly when a project has many steps, uncertain goals, and no single right path. Those jobs include choosing which result matters, spotting when a line of work is a dead end, and deciding what to try next when the evidence is incomplete. (nature.com 1) (nature.com 2) Researchers are still moving AI deeper into the lab anyway. Nature reported on March 31, 2026 that institutions and funders are already grappling with how “AI scientists” could change research practice, credit, and oversight. (nature.com) Some of the pressure comes from systems built to automate more of the workflow, not just answer questions. Nature reported on March 27, 2026 that “The AI Scientist,” first released in 2024, had reached peer review, with outside researchers highlighting both its strengths and its limits. (nature.com 1) (nature.com 2) Scientists are also documenting failure modes that matter outside benchmarks. On April 7, 2026, Nature described an experiment in which chatbots treated a fabricated disease as real after bogus papers about it appeared online. (nature.com) So the picture in April 2026 is two facts at once: the systems are improving on a steep curve, and the hardest research problems still lean on human judgment. That is why laboratories are using AI as a powerful tool inside the workflow, not as a replacement for the people running it. (technologyreview.com) (nature.com)
