A curriculum-aligned AI tutor is only as accurate as the curriculum it can read. A mixed week of lawsuits, interpretability, labour-market data and the unglamorous foundations that decide whether classroom AI is accurate or just plausible. Also a reminder that a third of workers are already using these tools without telling their employer, and that most of them learned how on social media.
What we mean by an AI tutor is set out on the AI tutor for schools page.
Apple sues OpenAI
I am not a lawyer, but I do watch a lot of TikTok. Apple sued OpenAI on Friday, and here is my take.
Apple alleges two former employees carried iPhone and Apple Watch hardware secrets into OpenAI's device business. One is accused of exploiting an authentication bug to download over a thousand pages of files, then coaching a colleague still at Apple on avoiding the security team. The other is accused of using internal code names in interviews and asking candidates to bring parts for show and tell.
The download and hacking allegations are concrete and forensic, and courts will weigh that heavily. The interview and recruiting claims are weaker, because California treats poaching as normal and rejects knowledge alone as theft.
Apple's hardest task is corporate liability. Proving individuals took secrets is one thing. Pinning it on OpenAI is another, and Apple lost that exact argument in its 2022 Rivos case.
Reading the model's workspace
It took me a while to get to it, but I have now read Anthropic's global workspace paper. At minimum you should watch the video. What is immediately noticeable is the uncanny similarity with how we think. I am most excited about what this means for trust and transparency.
For the first time we have a way to read what a model is silently reasoning with, before it produces a single word. The researchers found a small set of internal representations they call the J-space. The model uses it to hold concepts in mind, chain reasoning steps, and plan what it might say. In their alignment audits the technique surfaced strategic deliberation, emotional reactions, and even the model's awareness that it was being tested.
Every evaluation regime we have today judges models by their behaviour. But a model that knows it is being evaluated can behave accordingly. This research showed that removing the model's internal recognition of the test surfaced propensities that were otherwise concealed.
The full technique is too heavy for live traffic, but the paper shows you can check for a single concept with one cheap operation. A curated panel of safety probes could be performed at every step of the model's internal processing at negligible cost. This is a monitor with known blind spots, not a lie detector. The authors say so themselves. But it moves interpretability from explaining outputs after the fact to observing intent as it forms.
Not the job-pocalypse
"This is not definitive evidence of job loss caused by AI." "Should therefore be read as justification for ongoing monitoring rather than clear evidence that AI has reduced employment."
The Department of Employment and Workplace Relations released AI and employment in Australia. It finds no broad AI-driven disruption. Unemployment is at 4.2 per cent, below the pre-COVID decade. Youth outcomes are holding and there is no acceleration in occupational churn.
The fifth most at-risk occupations grew 5.6 per cent since November 2022, versus 9.5 per cent for the least at risk. The core model estimates employment about 2 per cent below the pre-ChatGPT trend for the most exposed roles. Perhaps early signals, but definitely not the job-pocalypse.
The foundations of a curriculum-aligned AI tutor
Australian and New Zealand education systems should read the evidence Oak National Academy gave to the Education Select Committee this week. A machine-readable national curriculum, an open content ontology, and quality-assured source material are the unglamorous foundations that decide whether AI in classrooms is accurate and curriculum-aligned or just plausible-sounding text. Get those right and every tool built on top gets better. Get them wrong and small errors scale across the whole system.
Shadow AI is shadow engagement
Everyone is going to be quoting the 1 in 3 stat: workers using AI without telling their employer. You would be tempted to think the solution is technical, but I think the stat underneath it significantly contributes to the problem. Fifty-one per cent learned their AI skills from social media. This is a training problem, and the training is happening on TikTok.
Employment Hero's research is a useful prompt for anyone still treating this as a tooling issue. Modern corporate training needs to catch up. It needs to be engaging and short-form, reframed around what is allowed versus disallowed, highly personalised and relevant. We have known for a while that long-form is not working any more. Nobody chooses a corporate LMS module over a three-minute video that is actually good.
You cannot fix this with another PDF on the intranet. You have to make AI use legitimate, and you have to build relevant content people would actually pick over TikTok. Maybe we should think of this less as "shadow AI" and more as "shadow engagement".
A big week of models, briefly
Cursor and SpaceXAI shipped their first jointly trained model, built on Cursor's real developer workflow data. Google led the 3.5 family with Flash instead of Pro. OpenAI's GPT-5.6 lineup is now in the wild, with Sol at the top, Terra matching the previous generation at half the cost, and Luna at the fast end, powering ChatGPT Work. And full-duplex voice that backchannels with "mhmm" and stays quiet when you need to think. The demo with older women chatting away was genuinely charming. We need more of this: warm, human, accessible AI.