01
01The Meaningful Override RateHuman oversight of AI is either real, or it is theater. This is the instrument that tells them apart.
02Human Oversight Is Mostly TheaterThe EU AI Act made human oversight the law. Almost no one is measuring whether it's real. Here's the metric that would, and the bet I'm making on it.
03Accountability InversionWhen a person and an AI system share a decision, the blame does not split evenly. The authority drains to the machine. The accountability stays with the human, who often had the least real control. That gap is the quiet defect under most AI oversight.
04Who sets the rules AI runs on?The fight that decides how AI shapes your life is not about how capable the models get. It is about who writes the rules underneath them, and whether you are ever allowed to see those rules at all.
05The Game of Life, explainedIt is not a game you play. It is a game you watch. A grid, four rules, and a world you did not design. Here is what it is, why it is mesmerizing, and what it has to do with who sets the rules AI runs on.
Start here
502
06A constitution for AIThe values a model holds are real whether or not anyone wrote them down. A constitution, in the loose sense, is just the decision to write them down: to make them explicit, public, and amendable, instead of leaving them implicit and beyond reach.
07A crumple zone with a job titleA car's crumple zone is engineered to take the impact so the passenger does not. Put a human in that role inside an AI system, a person designed to absorb the blame, and you have built something with the same shape and a worse purpose.
08A decision too important to automateThe question is not whether a machine can make the decision. Often it can. The question is whether it should, and some decisions belong to a person for reasons that have nothing to do with capability.
09A portal for human judgementIf an AI agent needs a person's judgment, where does it go to get it? Most systems have no answer, just an informal favor from whoever happens to be around. Naming the destination, a portal for human judgment, is what makes it something you can rely on.
10A recall for a decision systemA faulty car gets recalled. A faulty toaster gets recalled. A faulty model that made thousands of wrong decisions usually gets a quiet update, while the decisions it already made stay exactly where they landed. That asymmetry is strange, and worth fixing.
11Accountability CoverageFor how many of your AI system's consequential actions could you name the specific person answerable for it? If the honest answer is "not many," the oversight is thinner than the org chart suggests, and this is the figure that shows it.
12Audit trails for AI decisionsMost systems log plenty. Almost none log the thing that matters after a bad decision: enough for an outsider to reconstruct what happened, why, and who was answerable. A trail only the vendor can read is not an audit trail.
13Auditability is the new uptimeA generation ago, software earned trust by publishing uptime, a number you could check instead of a promise you had to take. AI decisions need the same move, and the number is auditability.
14Building AI oversight in AmsterdamIt can look like a handicap to build under the strictest AI law in the world. I think it is the opposite. The hardest environment to satisfy is the best place to learn what real oversight requires, and to build it before everyone else has to.
15Capability is downstreamThe models keep getting better, and that is the story everyone tells. It is the wrong story to obsess over. Capability is an input. The rules that decide what it is used for, and who answers when it is wrong, are the output that actually matters.
16Contestation LatencyHow much time and context does a human actually have to disagree with the machine? I call that contestation latency. When it is three seconds, the oversight is a turnstile, not a judgment.
17Designing the meaningful overrideAn override is only real if the human can actually reach it, has what they need to use it, and the system honors the result. Miss any of the three and you have built a button that changes nothing but the audit log.
18Dignity as a layer, not a sloganDignity appears in almost every AI principles page and almost no actual systems. The difference between the two is whether dignity is a line the system cannot cross by design, or a word it prints while crossing it.
19Equal say on AI valuesExpertise earns extra weight on questions that have a right answer. Values are not that kind of question. Confusing the two is how a small group ends up quietly deciding what everyone else's AI should care about.
20Escalation FitIt is not enough that an AI system escalates to a human. It has to escalate the right things. Send too little and the agent decides hard cases alone; send too much and the human is buried and starts approving on reflex.
21Every model is an opinionWe talk about models as if they report the world. They argue about it. Every dataset is a selection, every objective a preference, every guardrail a value. Neutral is the one thing a model cannot be.
22Four places missing oversight hits the balance sheetRemoving the human looks like a saving on the day you do it. The bill arrives later, on different pages, attributed to other causes. Here is where to look for it before it finds you.
23From data labeling to live judgmentFor years, "human in the loop" meant a person labeling examples so a model could learn. That human mattered, and then left before the system ever made a real decision. For agents that act, the human has to be in a different place entirely.
24Giving the governed a voteA system can be accurate and still have no right to decide. Legitimacy is the missing word in most AI conversations: not is it correct, but on whose authority does it impose its values, and do the people affected have any say.
25Governance is a productMost AI governance lives as a PDF: written once, filed, and quietly contradicted by what the system actually does. Governance that works is not a document. It is a product, something you ship, measure, and improve like any other part of the system.
26How an AI agent asks a humanStrip away the jargon and it is simple. At a decision it should not make alone, the agent stops, hands the question to a person with enough context to judge, waits for the answer, and only then acts. The simplicity is the point.
27How do you measure whether oversight is real?You cannot manage what you have not named, and human oversight of AI has gone unnamed for too long. Here is the small set of numbers that, taken together, tell command apart from theater.
28How I got hereThe kid in a 1990s computer lab, the two games that marked me, and the one idea I have been circling ever since.
29How to build oversight that actually holdsOversight that holds is not a warning label or a person cc'd on the outcome. It is a handful of design choices, each boring on its own, that together keep a human genuinely in command when it counts.
30Human response-time metricsReliability teams never settled for one timing number, and neither should oversight. How long until a human is reached, how long until they respond, and how long they actually spend deciding are three different questions with three different answers.
31Layered oversight for AI decisionsNo serious security team trusts a single wall. They assume any one control will fail and layer the next behind it. Human oversight of AI deserves the same humility, because a single point of oversight is a single point of failure.
32Liability launderingSome human sign-offs do not add judgment to a decision. They add a name to blame. That is not oversight. It is liability laundering, and once you see it you cannot unsee it.
33Model size is not trustA bigger model is more capable. It is not, by that fact, more trustworthy. Capability and trust are different axes, and confusing them lets a very smart system be a very unaccountable one.
34Override ValidityIt is not enough that a human overrules the machine. The override has to be right. A reviewer who confidently overturns correct decisions is not oversight, just a different source of error wearing the costume of judgment.
35Oversight BudgetHuman attention is fixed. The number of decisions a machine makes is not. Your oversight budget is what is left when you divide one by the other, and past a point, it rounds to nothing.
36Provable versus performed oversightThere is a real difference between proving a human was present and proving a human was in command. Most oversight, even the kind with a cryptographic receipt, proves the first and quietly skips the second.
37Refusal as a featureA system that will attempt anything is not powerful, it is unbounded. The decisions an AI deliberately will not make, the calls it hands to a person or declines outright, are a designed feature, and often the most important one.
38Rubber-Stamp RateCount the times a human was asked to approve a machine's decision and changed nothing. Divide by the times they were asked. The closer that fraction sits to everything, the more your oversight is a signature, not a judgment.
39The accountable humanWhen an AI agent makes a consequential call, there should be a person, a specific one with a name, who is answerable for it and could have changed it. A queue is not accountable. A department is not accountable. A person is.
40The AI trust deficitSay "responsible AI" to an experienced buyer now and watch the small flicker of doubt. The phrase has been used too often by systems that were not, and the discount is now applied automatically.
41The Automation Bias IndexAutomation bias is not a new discovery. The tendency to lean on reliable automation and stop checking it has been documented for decades. What we lack is a way to watch it grow inside a specific system, before it has hollowed the oversight out.
42The confirm button is not oversightMost human oversight of AI comes down to a button. A button that is only ever pressed measures nothing. If the person approves every time and changes nothing, you have not added oversight. You have added a stamp.
43The cost of a wrong noCompanies watch the wrong yes, the fraud that got through, the loan that defaulted. The wrong no is quieter and often costlier: the good customer denied, with no person to appeal to, who becomes a complaint, a headline, or a regulator's example.
44The expertise layer for AI agentsModels have a knowledge layer. Agents need something it does not provide: a path that takes the few decisions a machine should not make alone, and puts them in front of a qualified person who can be answerable for them.
45The human SLASoftware has long sold service levels: we will be up this much, respond this fast. Human oversight needs the same move. A human-in-the-loop SLA turns "a person can step in" from a hope into a commitment with a number behind it.
46The kill switch nobody wants to needEvery serious system needs a way to stop. The trouble is that the stop control is the one feature nobody exercises, so it tends to exist on the diagram and fail in the moment, which is the only moment it matters.
47The next decade of humans and AII will make a falsifiable bet. The next ten years of AI will be decided less by how capable the models become and more by whether we build the unglamorous infrastructure that keeps a human genuinely in command of them.
48The oversight illusionA power nobody uses looks exactly like a power that was never there. That is why so much AI oversight passes inspection: from the outside, the real thing and the empty thing are the same picture.
49The regulator is not your oversight strategyLaws like the EU AI Act are starting to require a human over high-risk AI. That is a floor, and a welcome one. It is not a strategy. Treating compliance as the whole of your oversight is how you pass an audit and still fail the people the system decides about.
50The right to a humanWhen a machine makes a decision that changes your life, you should be able to ask for a person. Not a different machine, not a form, a human who can look again and decide. That claim is starting to have the force of law behind it.
51The rubber-stamp problemA better model should make oversight easier. Instead it often makes it emptier. The more reliably the machine is right, the more reflexive the human's approval becomes, until the sign-off means nothing.
52The thought economyWhen machines can do the work and recall the knowledge, the scarce thing left is judgment: deciding what matters, what is acceptable, and who answers for it. That scarcity reshapes where value sits, and who gets to think.
53The trust layer for AIReliability engineering gave us uptime, a number you can check instead of a promise you have to take. AI needs the same move for trust: a layer that produces evidence, so trust is something you can verify rather than something you are asked to feel.
54The trust taxEvery decision a customer cannot appeal to a human carries a hidden surcharge. They do not send you an invoice. They pay you a little less attention, a little less loyalty, and a little more suspicion, and it compounds.
55The values layer for AIEvery model applies values. The values layer is the part of the system where those values are written down where you can see them, argue with them, and change them, instead of leaving them buried where you can only discover them by being harmed.
56The vantage pointI did not arrive at this work from one place or one discipline. The thread that runs through it, a wariness of any system that decides for people without answering to them, came from seeing how differently that plays out depending on where you stand.
57The whose half of alignmentAlignment research mostly asks how to make AI follow human values. It quietly assumes there is one set to follow. There is not, and the assumption hides the harder, more political question: whose.
58Time-to-HumanOperations teams already live by one number: how long until a problem reaches a person who can fix it. Point that same number at AI, and you can finally measure the thing everyone claims and almost no one checks.
59Trust is something you engineer, not advertiseYou cannot label your way to trust. The word "responsible" on a page is not a property of the system, it is a hope about it. Trust is something you build in and then prove, or it is nothing.
60Value pluralism for AIA single set of values, applied to everyone, is either bland enough to be useless or specific enough to be unfair to someone. The honest design stance is pluralism: a values layer that can hold real differences instead of pretending they do not exist.
61What is human-in-the-loop for AI agents, really?The phrase is old and the meaning has shifted under it. For models that only predicted, a human in the loop was a labeler. For agents that act, it has to mean a person who can step into a live decision and be answerable for it.
62What it costs when no human is really in the loopThe bill for missing oversight does not arrive as a fine. It arrives quietly, as liability you did not price, work you have to redo, customers who stop trusting you, and a regulator who arrives later than your problem did.
63What trustworthy AI actually requiresTrustworthy AI is one of the most used and least specified phrases in the field. It is worth asking what it would take to deserve the word, rather than just print it.
64When an algorithm decides, and no one owns the ruleThe danger is rarely a dramatic robot. It is a quiet rule, set once, far upstream, that then decides at scale while the people who wrote it are nowhere near the outcome. By the time the harm is visible, no one can say who chose it.
65When expertise should outweigh consensusCounting heads is the right way to settle some questions and the wrong way to settle others. The trick is knowing which is which, because using the wrong rule quietly hands the answer to the wrong people.
66When should an AI agent escalate to a human?Too much escalation and the agent is pointless and the humans are buried. Too little and it acts alone on calls it had no business making. The art is in the line between, and the line can be drawn on purpose.
67When the AI agent acted aloneA model that predicts can be wrong on a screen. An agent that acts can be wrong in the world: the message sent, the account closed, the order placed. The new failure mode is not a bad prediction. It is an action no human ever checked.
68Whose values should AI hold?There is no view from nowhere. Every model that helps decide anything carries values, picked by someone, somewhere, usually without being asked out loud. The only honest question is whose, and how they were chosen.
69Why my logo is a flipped glider, and a question markA small thing, explained once. What five pixels say about how I think, and why the glider is turned over.
70Why the human goes quietThe better an automated system gets, the less the human checks it. That is not laziness. It is a rational habit, and it is exactly why the rare error slips through unguarded.