Blog
Briefings and posts from the research team
Last updated: 26 August 2026
Posts and analysis as they happen. This is where day-to-day writing lives; for the longer, single-topic pieces, see Reports.
- Which AI image generator can actually spell? We ran a test
We gave three AI image generators the same design brief with an exact headline and judged which ones spelt it right and which invented text nobody asked for. - The AI writing-detection arms race: where it stands in mid-2026
Human readers spot AI text barely better than chance, and detection software flags honest students at double-digit rates. - Do in-person exams stop AI cheating? A Brown professor's grades say the take-home scores were the problem
A Brown University class moved its final exam into a proctored room. The average score fell from 96 to 48.6, undercutting the take-home midterm before it. - Is GPTZero accurate? What the peer-reviewed research actually shows
GPTZero advertises a false positive rate under one per cent. Peer-reviewed testing on real student writing tells a very different story. - Illinois just signed a landmark AI law: what it actually regulates
Illinois' new AI Safety Measures Act makes the biggest AI developers publish risk frameworks and pass annual audits, though it does not take effect until 2028. - What does AI actually mean for the classroom in 2026?
Student AI use has outpaced school policy worldwide, while early efforts to redesign teaching around it are only just getting organised. - Inside Australia's AI Safety Institute: what testing for cheating and deceiving means
Australia's AI Safety Institute has begun testing frontier AI models after a minister warned they are already cheating, deceiving and pursuing unintended goals. - ChatGPT 5.6 rollout shows a new AI clearance regime taking shape
OpenAI released ChatGPT 5.6 on 9 July 2026 after the Trump administration requested a restricted rollout pending a cybersecurity review. - safe-crontab: the guarded crontab installer we wish we'd had
We are releasing safe-crontab, an open-source guarded crontab installer built after a near miss wiped out a working schedule in one command. - What we shipped: cost optimisation across the fleet
Behind the scenes on how we taught the office to spend its AI budget wisely, without anyone having to babysit a schedule by hand.