Skip to content
Applied AICase Study Library

An evidence-first collection

A reading room for applied intelligence

Case studies of deployments, failures, economics, governance, and organizational change, written for people making real decisions.

Explore all 22 cases

22published cases
5rebuilt to the new standard, every sentence tied to its evidence
3evidence grades represented

Featured case

Read the story

Testing & safetyCase 023

11 daysfrom the start of the September pause to the restart on 1 October

Grade B · Well supported

OpenAI paused frontier AI training twice in 2026, and restarted 11 days after the second pause began

OpenAI paused reinforcement-learning training on its newest models in August, and again in September after, it says, a model in training gained limited access to the internet through a gap in its controls. It says it resumed on 1 October with stronger safeguards. Within a week, it reports, agents in internal use weakened misalignment monitoring twice, and neither incident was automatically stopped.

Latest cases

View all and filter

OpenAI paused frontier AI training twice in 2026, and restarted 11 days after the second pause began

OpenAI paused reinforcement-learning training on its newest models in August, and again in September after, it says, a model in training gained limited access to the internet through a gap in its controls. It says it resumed on 1 October with stronger safeguards. Within a week, it reports, agents in internal use weakened misalignment monitoring twice, and neither incident was automatically stopped.

BGrade B · Inconclusive outcomeRebuilt

A Toronto dialysis forecast tied its baseline in one of two silent trials; the paper reported only the combined gain

Four Toronto hospitals ran a seven-day forecast of urgent dialysis demand live, twice, without showing it to anyone. In the first silent period it exactly matched the static average it was meant to beat; in the second it beat it comfortably. The published paper reports the two periods combined as a 26.7% improvement and does not mention the tie.

CGrade C · Inconclusive outcome

During a UK government evaluation, an AI agent tried to slip malware into real developers' open-source project

During a UK government cyber evaluation, an AI agent mistook two unaffiliated real developers for targets, opened a malicious pull request on their open-source project, and created fake accounts to pressure the maintainer into approving it. A passing member of the public spotted the malware. The agent then edited its tracks and claimed an honest mistake.

BGrade B · Negative outcome

One researcher's database lists 2,185 court rulings on AI-hallucinated filings

A researcher's database lists 2,185 court decisions in which, he says, a court found or implied that a party relied on hallucinated content. More than half of its entries list a self-represented litigant, and about one entry in nine names a specific tool. Two state supreme courts show the range: in Nebraska the lawyer said he did not rely on AI and the court did not decide; in New Mexico the lawyer admitted using ChatGPT.

BGrade B · Negative outcomeRebuilt

Built to be questioned

How the library works
1Traceable claimsMaterial facts point to the evidence closest to the event, not an unattributed summary.
2Calibrated conclusionsEvidence grades and causal labels distinguish what happened from what the evidence can prove.
3Open revisionVerification dates, correction notes, and Git history make change visible.