5 articles.
Real progress on open conjectures is impressive and mostly irrelevant to the problem set due Thursday. Separating the two is the skill students need right now.
METR ran a randomised controlled trial where 16 experienced open-source developers worked on issues in repositories they knew, with AI tools freely available in half the tasks. Implementation took 19% longer with AI. The developers predicted a 24% speedup before starting and still estimated a 20% speedup afterwards. Economics experts had forecast 39% and ML experts 38%. No confidence interval was published for the headline number, and the study has since been superseded.
A 2026 audit of 125 million references found fabricated citations rising more than twelvefold, and a benchmark of 13 models found citation hallucination rates spanning 14% to 95%. The references your AI hands you are the least trustworthy thing on the page.
Randomized trials from 2026 finally separate the two things people mean by AI tutoring, and they disagree. Unrestricted chatbot access produced negative retention effects in one trial. Coach-configured tutors embedded in structured practice produced measurable gains, and the gains were small.
I ran the same research question and set of papers through Elicit, Consensus, SciSpace and NotebookLM. Here’s what actually helped finish a literature review without inventing sources.