A few threads caught my attention this week, and they sit oddly well together: progress on assessing the skills we have always struggled to measure, fresh evidence about how young people are really using AI, and a sobering reminder that the same capabilities can be turned against us. Taken together they map the terrain we are all now working across.
Assessing the skills that are hard to assess
We have said for years that collaboration, critical thinking and creativity matter most, and then quietly assessed the things that were easiest to mark instead. New work from Google Research makes a genuine dent in that gap. Students interact with AI avatars in simulated team scenarios, and the clever part is the orchestration: an "executive" model steers the conversation to surface a specific skill, introducing pushback or conflict so there is actually something to assess, while a separate evaluator scores the transcript against a rubric.
Validated with NYU across 188 testers, the steering approach produced significantly more scoreable information than uncoordinated avatars. More telling still, the evaluator's agreement with human expert raters was comparable to the agreement the human raters reached with each other. A separate creativity study reported a 0.88 correlation between AI and human expert scores. Assessing the things we claim to value has always been the hard part. This looks like real progress.
What NSW young people say about AI
The NSW Office for Youth has released its 2026 Youth Week poll of 2,308 young people aged 10 to 24, and it is worth reading closely. 71% have used generative AI in the past year; among them, 63% use it at least a few times a week and 26% multiple times a day. The headline uses are predictable enough: study and homework help, information searches, self-learning and assignments.
What I did not expect was the personal dimension. 27% use AI for conversation and personal advice, 24% for mental-health information, and 29% said they had used it as a mental-health support strategy in the past year, rising to 37% among 18 to 24 year olds. That deserves care, not alarm. Notably, young people themselves named tool dependence and the erosion of critical thinking as their top concern. And there is an equity gap worth naming plainly: Aboriginal and Torres Strait Islander young people were far less likely to have used generative AI, 53% against 71% overall.
Bracing for a vulnpocalypse
A thoughtful article lays out the case for a "vulnpocalypse" this year. If the claims about the latest models hold, a capable AI can scan a codebase and find vulnerabilities autonomously. The uncomfortable corollary is that the same capability is equally available to attackers.
I expect something like a coordinated event across enterprises, on the scale of Y2K, with technical staff conscripted from every corner of the organisation. It sounds dramatic, but the defensive response is not novel; it is simply the discipline we already know, done faster and more often:
- Run the scans yourself, on every release.
- Assume attackers have the same tools, and are patient.
- Do it before someone else does it for you.
None of that is glamorous, and that is rather the point. The advantage goes to whoever treats this as routine practice rather than a one-off scramble.