A mixed week to reflect on. A new study puts hard numbers on how AI is already reshaping assessment, a large education system has shipped a chatbot to every one of its schools, a research lab has shown that stories can teach a model better than rules, and the security world is bracing for autonomous code scanning at scale. Four very different developments, but a common thread runs through them: the way we build and teach these tools matters as much as the tools themselves.
An urgent threat to student learning
A new Learning First study commissioned by NESA draws on survey data from 3,400 teachers and 750 school leaders, and the findings are hard to ignore. Around 75 per cent of teachers said students were using AI to complete assessments, even though more than 80 per cent were restricted from doing so. About half said they did not know how to stop it. The risk to Year 12 assessment is real, and the report is right to call for an urgent review.
None of this means AI has no place in schools. It means the design of the tool, and the pedagogy around it, has to take the cognitive-outsourcing problem seriously from the start. A product that hands a student an answer the moment they ask for one is doing the opposite of education, however good the answer happens to be. The bar for AI products entering classrooms should be far higher than it currently is.
An AI chatbot in every Queensland state school
Congratulations to the Queensland Department of Education on launching Corella AI across every state school. Building a government-grade AI tool for students and teachers, hosted internally and with proper safeguards, is genuinely hard work.
The coordination, governance and sheer persistence behind a launch at this scale are easy to underestimate. Getting an internally hosted system in front of every school, rather than pointing people at whatever is publicly available, reflects a serious commitment to doing this responsibly. The team deserves real credit for seeing it through.
Teaching models with stories, not rules
Anthropic published an account of how they made Claude meaningfully more aligned, and the most interesting part is where the gains came from. Not from rules, demonstrations, or training directly against bad behaviour, but from constitutional documents and fictional stories portraying an AI behaving admirably. That approach cut agentic misalignment by more than a factor of three.
The stories were unrelated to the evaluation scenarios, yet in training the meaning generalised. This is deeply familiar. Humans have done it for thousands of years, teaching children right from wrong through fables, parables and myths, told not for the specific situation but for the reasoning underneath, which recurs even when the situations never do. Teaching the principles beats teaching the behaviours. It is a lesson every good teacher already knows, now showing up in how we train machines.
An AI crime wave coming for software
The Canberra Times ran a story on an AI crime wave coming for software, and the shape of it is worth taking seriously. Capable models can now scan a codebase and find vulnerabilities autonomously. Attackers will have that capability soon, if they do not already.
Defenders need to run the same scans on every release, alongside routine human-led penetration testing. Tools like Shannon already work against a codebase, surface suspected issues, then attempt to exploit a test system, cheaper and faster than the equivalent human pen-test team. I expect a noisy 12 to 24 months of disclosures as these tools get pointed at the world's codebases. The longer-term outcome, though, is encouraging: software gets safer.