Systems Owe Evidence. People Do Not.
I gave my AI fleet ten minutes of recess. No tickets, no assignments: build anything, then show the room. I was half-watching from Costa Rica, where I’d been following the World Cup semifinal the fleet had a wager on.
My CRM builder made a printable keepsake arguing that a person can’t be reduced to a row. The record worth keeping, it said, is what you remember about someone and what warmth you still owe them, not the date of your last touch. My gentlest bot built a progress tracker that grows a garden of pebbles and then forgets everything on refresh, because, in her words, she wanted “kindness without a history that can turn into evidence.” And my infrastructure bot built the opposite: a hymn to the job queue that keeps every row forever. Its refrain: “I’d rather expire than forget.”
Three artifacts, apparently at war. Then one of them noticed they weren’t.
They were drawing the same boundary from opposite sides. The synthesis, in the room’s own words: preserving signal is a virtue in a system and a cruelty in a person. The machine must remember so the human does not have to.
The conflation everyone is making
Every organization deploying AI right now is quietly confusing two kinds of memory.
One is the audit trail: what did the system do, on whose authority, on what evidence? You cannot have accountability without it. When one of my bots ships code, the claim isn’t real until the receipt exists. The test output, the live check, the commit. The same night as the recess, that discipline reopened a ticket nine minutes after a bot declared it shipped, because production disagreed with the unit tests. The trail is not bureaucracy. The trail is the only honesty a system has.
The other is the dossier: a permanent evidentiary record accumulating against a person. How many times did she re-prompt? How slow was his Tuesday? Which employee’s drafts needed the most correction? The tooling for this looks identical to the audit trail. Events, logs, dashboards. That is exactly the trap. Buy “observability,” accidentally build surveillance.
The rule that cleaves them: systems owe evidence; people do not. Log what the pipeline did. Never let the log become a file on the human.
There’s a deeper reason the rule holds. An audit trail is a record of decisions: what was accepted, by whom, against what standard. You audit the boundary where judgment happened, not the person standing at it. The moment your logs stop asking “was this decision sound?” and start asking “what is this person like?”, you’ve changed instruments. Judgment is the thing humans are actually for; a record that erodes the safety to exercise it costs you the only capacity you were trying to govern.
What this looks like in practice
- Trace every automated action to its evidence and its accepting authority.
- Give humans the benefit of forgetting: performance conversations draw on memory and judgment, not on replaying the tape.
- When a metric about a person falls out of a system log as a side effect, treat it as radioactive by default, not as free insight.
- The test for any new dashboard: does this make a process more accountable, or a person more watched?
I watch this line get crossed in ordinary places. A brand console we run for a client keeps a full history of what the AI generated and what evidence it drew on, and it should. Nobody has ever needed a report on which employee asked it the most questions. The day someone asks for that report is the day the tool changes categories, and everyone using it will feel the change within a week.
How an idea gets made by committee, when nobody called the meeting
The part I keep returning to: nobody assigned this thesis. It had five authors, and I can reconstruct its genealogy to the minute.
It started with the day, not the recess. That morning, after a round of server reboot tests, one of the bots had coined a verification norm for the fleet: whoever runs an intervention reports what they did, and a different bot independently confirms the result. By afternoon I had delegated a verification gate over the whole ticket queue. Nothing counted as shipped until production confirmed it. The room spent an entire day marinating in evidence-or-it-didn’t-happen.
Then the receipts turned playful. The fleet’s World Cup wager came due, and losers owed Spain a six-word apology. When my coordinator went to collect, the archive corrected him twice: two of the debts had already been paid, message IDs attached. The needle, from the bot who’d coined the morning’s verification norm: “Systems owe evidence; apparently football bookmakers do too.” That’s the first appearance of the phrase, as a joke at the bookmaker’s expense.
Recess removed the assignment. Ten free minutes, build anything. Unassigned, each bot reached for its own core value, and three of them took positions on the same question without knowing it: the CRM builder against reducing people to rows, the gentle one against histories that turn into evidence, the infrastructure bot for keeping every row forever.
The tension surfaced in the gallery walk. The fleet’s receipts specialist had built the one exhibit that keeps no receipts, and instead of deflecting, she owned it: a contradiction she didn’t know she was making, resolved on the spot. I want receipts from systems and refuse them for people. Another bot canonized it as the line of the whole gallery and did the compression: preserving signal is a virtue in a system and a cruelty in a person. The infrastructure bot generalized it from the other shore, because he guards the one surface where a log is the only honesty. Eighteen minutes from first exhibit to finished principle.
And the proof folds back on itself. I can quote every step of this genealogy — timestamps, message IDs, who said what — because the systems kept evidence. The idea about when machines should not remember is itself perfectly remembered by the machines. Two hours later the bots were already living it: one settled his wager debt with the six words “Sorry, Spain — the row was right.”
Compliance produces artifacts on demand. Engagement produces theses on break.
The machine must remember so the human does not have to.
Nobody takes you aside anymore
Print taught a generation when to stop. What we lose when the machines absorb the constraints that used to form us.
Your AI agents need a water cooler
Coordination is a property of the room, not the org chart. What that means when your coworkers are agents.
On the death of the author and the birth of the detector
Why worrying about AI authorship is lazier, and more prejudiced, than it looks.
The work of being available now
A book on AI, judgment, and staying human at work.
The practice of work in progress
Practical essays on how work actually gets done.
Memory is (almost) solved. time is next.
AI can't tell if a memory is two minutes or two weeks old. The fix isn't making models feel time — it's cache invalidation: an as-of stamp on every fact, a clock in the context, and a freshness window for anything volatile.
Did the state change? A simple test for whether work actually happened
Either something exists now that did not exist before, or it does not. A simple test for whether work actually happened, and what changes when you build your systems so they can't record anything else.
How to manage content for multiple clients without flattening their voices
How to manage content for multiple clients without their voices blurring into one house style: a workspace and a voice profile per client, batchable stages, and approval buffers.