Wisdom that outlasts the algorithm.
THE CURVE [insight]
Somewhere on a dashboard you trust, a zero is standing in for "no one looked." A month ago I was reading one of those. I had built it myself.
It tracks a small content operation. How much gets written and how much actually ships, one row per week. I built it because I wanted to track what shipped, and I had been running for weeks on the writing number alone.
The shipping column read zero. Week after week after week. Seven straight weeks of a team producing nothing.
That is not what happened. The team shipped. What had happened is that the field the table counts, the date a thing went out, was not being filled in during those weeks. There was no data. And the table, doing exactly what tables do, rendered no data as 0.
Same column. Same font. Same alignment. Sitting in a row next to numbers that were real. Nothing on the screen distinguished "we measured and it was none" from "we never measured."
I want to be clear about who is at fault here, because the answer is no one. The person who did not set the date was not hiding anything. That field had no purpose yet. The table simply summed an empty set and got the correct answer to a question I had not meant to ask. I built the thing and I read the thing and for a while I believed it.
Nobody lied. The interface did.
The fix took one character. Those weeks now render as a dash, with a line underneath the table that exists solely to say so: not instrumented that week. Not a zero. And a standing gate above every output number: at last check, 23 of 79 items carried the date at all, so the number is a floor, never a count, and the coverage percentage gets quoted alongside it every single time.
One character and one sentence. What it bought was the difference between a number that is red and a number that is blind. This is a useful distinction in measurement. A red number tells you something is wrong. A blind number tells you nothing at all while looking exactly like a red one.
Blind is worse than red, and it is the one no one escalates.
This is a source of false confidence, and it is not where we usually go looking. We go looking for a person who overstated. Sometimes there is one. But far more often the number is wrong and everyone in the chain was honest, because the system made the misleading answer the cheap one and the accurate answer the expensive one.
There are two versions of this and they are the same failure wearing different clothes.
The first is rendering. Absence and zero are different facts and most tools display them identically. "We measured and found none" and "we never measured" are very different, yet they render as nothing at all on the screen. Same for withheld versus empty, stale versus current, estimated versus counted. In every one of those pairs, one member is a finding and the other is a gap, and the default rendering flattens them into the same confident-looking digit.
Ask why, and the answer is not accuracy. A blank cell looks like a bug. A zero looks like a finished product. The default was chosen to make the software review well, not to make the business decide well. Different goals.
Somewhere along the way a false read stops being information and it becomes furniture. A warning that cannot be satisfied and a warning that never fires trains a reader to stop looking.
The second is the moment.
Sonar's State of Code survey (January 2026, more than 1,100 developers) found that 96% do not fully trust AI-generated code, and 48% always verify it before committing.
Read those two numbers together, because the gap between them is the whole point. This is not a population that needs convincing to be skeptical. They are already skeptical. Professionals. Paid to be skeptical. And half of them still ship without checking.
That is not a knowledge problem and it is not a character problem. It is a scheduling problem. Verification that requires you to stop, open a second window, remember to be suspicious, and go perform a separate task loses to a deadline every single time.
Verification does not fail because it is hard. It fails because nothing ever asks you to do it.
When the honest answer costs more than the misleading one, you get misleading answers, produced by honest people, at scale. The fix is not more integrity. It is to make the honest claim the easy one, and the over-claim impossible to type.
Week over week I got a little smarter about the instrument, and the machine producing the instrument stayed wrong. The honest answer has to land at the place the work actually happens.
That is not a documentation problem. It is the same failure as the zero. The honest thing existed as a description instead of as a mechanism. A correction you wrote down is a correction you are hoping someone reads. Making it so that the system cannot skip the correction is a solution.
Almost everything written about trustworthy AI right now is addressed to the person. This is addressed to the surface.
──────────────────────────────
THE SIGNALS [three reads: unfinished, sharp, costly]
→ Half-Baked.
I have been treating the zero as a defect. But a zero ends a meeting. A dash starts one. Someone has to say what they do not know and why. I keep wondering whether organizations render absence as zero by accident, or because a confident wrong number is cheaper to sit with than an honest gap. If it is the second, the fix is not a better interface.
→ Hot Take.
Every AI literacy framework I’ve read this year lists "verify outputs" as a competency, in roughly the way a health curriculum lists "eat vegetables." True, universally endorsed, behaviorally inert.
The frameworks are written at the altitude of principle because principle survives committee review. The thing that would actually change behavior is small, specific, slightly undignified, and difficult to put into a strategy document.
I would rather have it get done.
→ Confession.
I built four separate verification reps into a single five-hour teaching day. Four. Not because the idea is complicated. It takes one sentence. But because I did not believe one telling would be sufficient to be retained. This raises the obvious question about what I am doing right now in this newsletter. If I needed four reinforcements for a room I had in front of me for five hours, one paragraph is not going to install this in a reader I have for two minutes. I am aware of the tension. I am not going to pretend I have resolved it.
──────────────────────────────
THE NEXUS [your turn]
What is the last number you trusted mainly because it was rendered like a number?
Reply with yours. I read every response, and the interesting ones become future issues.
──────────────────────────────
THE MOVE [this week, this quarter, this decade]
THE MONDAY MOVE
Next time you ask an AI for a count of anything, rows, items, totals, categories, add: Show your counts.
Then pick any one number it gives you and check it against the actual source, by hand.
That’s the whole thing. It takes 11 seconds.
If the number is wrong, you found out before anyone downstream did. If it is right, you now know something narrow and real: that this answer is trustworthy. Not the tool. This answer. The distinction is the point, and it is the only kind of trust that holds up, because it is the only kind you actually earned.
THE ASYMMETRIC MOVE
Two steps.
First: Whatever prompt scaffold, SOP, or checklist your team already reuses, add "show your counts" as a standing line. The 11 seconds should be a property of the workflow. Put the line in the template, not in your memory.
Next: run an audit. Take your most-used dashboard or recurring report and ask seven questions of it:
1. Does absence render differently from zero?
2. Is "withheld" distinguishable from "measured and empty"?
3. Does a stale number look different from a current one?
4. Is an estimate labeled as an estimate at the point of reading, not in a footnote?
5. When the underlying data is missing, does the surface say so, or does it fill in?
6. Does every caution on this surface have a state in which it switches off, and has anyone ever seen it off?
7. When someone corrects this report, does the correction land in the thing that generates it, or only in the thing you read?
Find one cell that reads zero and ask the person who owns it whether that is a real zero or a "we do not know" dressed up as a zero.
THE DECADE MOVE
Stop trying to install skepticism and start engineering the moment.
Seatbelts did not work as a poster. They worked as a click built into the act of sitting down, and then as a chime that made not doing it more annoying than doing it. The behavior did not change because people became more careful. It changed because the honest action became the path of least resistance.
The organizations that come out of the next 10 years with their judgment intact will not be the ones with the best verification policy on the intranet. They will be the ones that engineered a checking moment into every workflow that had room for one, and refused to let a surface render a gap as a fact.
That is a design job. In most organizations it belongs to no one.
──────────────────────────────
THE COMPOUNDING ASSET [the part you keep]
The Rendering Audit. The seven questions above as a one-page checklist, built to be printed, run against a real dashboard in a team meeting, and forwarded to whoever owns the report. It is tool agnostic on purpose. It is not about your vendor. It is about whether the surface in front of you has a way to tell you it does not know, whether its cautions can ever be switched off, and whether a correction reaches the thing that builds it.
It is the first entry in something I intend to keep adding to: small instruments that make the honest answer cheaper to give than the confident one. The Show Your Counts line needs no page. It is three sentences, and they are above, ready to paste into a team channel.
──────────────────────────────
THE GROUNDING [return trip]
I write this newsletter on a simple bet: that most of what matters about this technology is happening on a curve, not at a cliff, and that the people who read the curve early get to make different decisions than the people who wait for the drop.
In the last issue I argued that the market is hiring builders, and that the work, once you are in it, is actually supervision. Supervision runs on readouts. So this one goes a level deeper and asks a harder question: even when you have the record in front of you, what makes a number in it true. It’s usually not anyone's integrity; it’s the design of the thing it was written on.
If something here changed how you are thinking, hit reply. I will respond.
Thank you for reading.
— Chris
Columbia Faculty · Chief of AI & Strategy at Essential Innovations · Founder, Attainable AI