Charlie Munger, Warren Buffett’s business partner, once said: “Show me the incentive, and I’ll show you the outcome.” At the Centre for Humane Technology, that maxim is how you diagnose what a product does to people – and it has become the guiding light of our own AI strategy. Fixter, our AI assessment sparring partner built by teachers for teachers, has been in testing since January. Commercial AI is built to be frictionless and agreeable – quick to answer, quick to agree, because that’s what keeps you coming back. Fixter is engineered to do the opposite: to slow a teacher down long enough for their own judgment to stay in the feedback, rather than being handed over to the model. Before finishing any piece of feedback, it stops and asks the teacher for a personal observation — something only a human in the room would know. We saw this first-hand in early prototyping. Teachers’ own fatigue, assumptions, and unconscious leniency were skewing feedback. And Fixter, marking only to the criteria, was catching it. I was the teacher it caught. A transcript came back “not yet passed” when I’d projected what I already knew about the learner onto the work: tired from marking too many assignments at once, and quietly softening my judgement because I knew they were struggling and didn’t want to ask for a re-submission. Fixter didn’t know any of that. It just held the work to the criteria. This isn’t a race to efficiency. Finnish education leader Olli-Pekka Heinonen argues that learning was never an optimisation game: productive struggle builds depth. Fixter is our attempt to apply that thinking to how teachers mark, not just what students submit. Fixter’s origin story starts with a heated debate on our concourse at an INSET day in January 2023, when we were all still figuring out how to move forward. Lisa Fixter, an engineering teacher, challenged the unthinking rush to embrace AI – not the technology itself, but using it without asking why, how, or whether it helped anyone learn. The tool is named for her; she’s now one of our most deliberate AI users, and her scepticism has guided its development. Fixter sits inside an intentional AI strategy: every student completes an AI Driver’s Licence at induction, now in its third year, before using our offloading tools. We’ve prioritised a refocus on authentic assessment since last year. Our learning coaches, who delivered CPD on it this August, are already seeing learning experiences that better prepare learners for their sector’s lived reality than an essay. The question now is not whether to use AI, but how to design assessment that cannot simply be outsourced to it. This is built around one distinction: using AI for offloading rather than outsourcing. Offloading is using AI to hold a first draft steady while you do the real work of thinking. Outsourcing is clicking approve. The American sci-fi writer Ted Chiang says using AI to complete assignments is like bringing a forklift into the weight room: you move the weight, but you never build strength yourself. The same is true for the person marking: a teacher who lets AI write feedback unread doesn’t get faster at noticing what a learner needs next. Take Priya, a hairdressing apprentice (a composite drawn from real learners we teach). Written essays at school made her feel education wasn’t for her. But her evidence for this unit, instead of an essay, is a vlog transcript on a client consultation that went wrong, annotated reflective photographs, and a witness statement from her supervisor. Her teacher feeds all three into Fixter, which flags where Priya’s account doesn’t quite match the witness statement and drafts feedback against the criteria. Before it finalises anything, the human nuance check stops and asks for something Fixter cannot see: what it means for Priya’s confidence to have her own voice reflected back, not a blank page. The teacher adds that context. VARCS Alignment then checks the draft against JCQ guidance in full, relevant awarding body standards, and the Centre for Humane Technology’s AI principles. Grounded in our own do-this, don’t-do-that teacher guidance, a running log tracks what percentage of Fixter’s wording the teacher keeps or edits, criterion by criterion. Fixter’s real quality assurance is design, testing, and iteration led by the teachers who use it every day. Are commercial AI tools actually incentivised to enhance learning? We deliberated hard to find ones that are – because not every AI platform shares an educator’s incentives. It is a sparring partner for noticing what the evidence shows, not a replacement for a teacher who knows Priya well enough to notice what the evidence can’t.