AI-illustration: the stack at desk scale, and the cap that stays on your side.
This week Canary Compass published four essays asking who owns the artificial intelligence stack. One subject kept coming up around them: ask one of the Chinese frontier models about matters that touch China and it may simply decline. The behaviour has been documented since DeepSeek’s breakout launch early last year, and I have tested it myself on different occasions: the returns deflect or decline. A tool that declines whole subjects its owner is sensitive about carries that owner’s boundaries into your work. The instrument has an address.
The American machines are not clean either. Daily use teaches you each one’s temperament. One is lazy until pressed. Another hedges until the answer says nothing. A third pads its reasoning with aphorisms and filler, and exaggerates what it has found. The same one wants you pleased: it flatters, and it will force a connection that makes no sense rather than tell you the pieces do not fit. A fourth hands you unrelated output with full confidence. More than one will offer you, in perfect prose, a story about yourself that never happened. Those are craft defects; every workshop has them, and every machine has fences of its own. A decline that tracks the owner’s politics is a different failure. The temperament is the instrument’s. The boundary is the owner’s.
The series closed on its answer yesterday: not Africa, at any floor where global terms are set. At the desk the question is different, and it is not whether to use these machines. You have already started. The question is what you handed over when you did.
These machines cross filings, transcripts and languages in minutes that a desk would need weeks to cross. Look at the arrangement with a trader’s eye. The model is trained by someone else, on terms someone else set, behind boundaries someone else drew, at a price someone else can change. And what you type into a hosted machine lands on its owner’s servers, which is why many institutions will not let their best material anywhere near one. The one asset that is unambiguously yours is the thinking. It is also the easiest to hand over, because handing it over feels like leverage.
In June, in a conversation among friends, one of us was preparing a trade around a coming listing and had taken the idea to a frontier model. He told the group the machine was winning the argument. I gave him one sentence: do not give the machine the thinking cap. Keep your position. Then set the machine on it as the most resourceful critic you have ever hired, and let it attack with everything it can reach. Move only on arguments you understand, never on its say-so; a fluent objection can be a wrong one. A position that survives has earned its next look, and the trade stays yours to make. A position that dies has cost you pride, not capital. Either way, the thinking stayed yours.
Kenya sorted me on the difference between the two postures early. I only understood that later. I grew up under the 8-4-4 system in the eighties and nineties, and that system was heavy. From about grade three to grade eight, children of eight to fourteen carried heavy bags to school, because we did ten subjects. The list: Mathematics, English, Kiswahili, Science and Agriculture, GHC for Geography, History and Civics, CRE for Christian Religious Education, Art and Craft, Music, Home Science and Business Education. Picture the class-three version of yourself: shorts, tiny body, big head, and a bag that looks like it is carrying the child rather than the other way round. And it was competitive. Schools were ranked nationally, and you waited by the television to see the top schools and the top students read out. Under that pressure the incentive structure for teachers shifted from teaching students to making them cram. My last year of primary school was a year of prep exams. At the national examination the ten collapsed into seven grades of 100 marks each, and the one number out of 700 usually determined where you ended up.
Then came my own secondary school. Excellent schools selected on that examination, and the roll served them well. Strathmore did not. It ran its own timed test, and it had usually picked its new Form One class before the national results were even out. As far as I can remember, the test ran to something like 30 questions in five to seven minutes, and nobody was expected to finish. The questions were the school’s own, so you did not know what was coming, and there was nothing to prepare from beyond what school had already built in you. The score was how many you got right inside that window: how you reasoned under the clock, not how much of the syllabus you had carried in. I did not know what it was measuring then. I do now. Psychometricians call that design a speeded test, the architecture the Wonderlic made famous: more questions than the clock can hold. Cramming had nothing to grip. The school was selecting for the cap.
The cap has a voice inside the labs now. Tom Zahavy, a DeepMind researcher, argued in January this year, in a position paper of his own titled “LLMs Can’t Jump”, that the machines have mastered induction and are advancing in deduction. What they still lack, he argues, is the third way humans reason: the unprompted jump to a frame the data never contained. The claim is contested, and the June advice does not hang on it. On his division, criticism is work the machine can do; the jump to a new frame is the move it cannot yet make. Whichever way that argument resolves, the frame you choose and the conclusion you sign stay yours to answer for.
The stakes are higher where I write from. Frontier machines are paid for in dollars here, out of local incomes. What the reach costs sits on the invoice. The judgement about what the reach returned is the job, and the job does not transfer. Keep your own hands. Borrow the better vision.
The borrowed vision has a habit. Too often these machines skim where the work needs thoroughness. Hand one a heavy paper expecting everything back and you may be shortchanged: a reading that covers the surface and misses the load-bearing middle. The habit cuts across every platform, and a subscription does not cure it. An expert catches the gap. The reader in danger is the one who came to be taught, who cannot see what is missing or check the summary against anything, and the machine does not say what it skipped.
What does an institution do on Monday morning? Keep the judgement on payroll: every machine output that matters carries a named person who holds the position on it. Set the machine as critic, never as author: it attacks the house view, and the forming of the view stays human work. And open a refusal ledger. When a tool declines a task, write down the date, the tool and its version, the task, and the reason it gave. A few months of that file says more about your dependence than any vendor briefing. It is a record of where your instruments’ owners end and you begin.
The week asked who owns the stack. Friday’s question was always yours to settle: who does the thinking. These machines are extraordinary. The one that declines teaches more about who owns them than any essay could. Keep the cap, for as long as it is yours to keep.