I read the AI policies of 51 institutions. 44 drew the same line.

Wisdom that outlasts the algorithm.

──────────────────────────────

44 of 51 drew the same line.

Last week I told you about an attorney who lost ninety minutes to a statute that didn't exist, and I made you a promise: the room proved the pattern, and the data was about to test it. Here's the data.

Over the past few weeks I read the public generative-AI policies of 51 of the world's most-trusted knowledge institutions.

Nature. JAMA. The BBC. The New York Times. Oxford, Harvard, MIT. UNESCO. The Associated Press. Institutions that exist - in the most literal sense - to be believed.

I expected 51 different answers. Fifty-one committees, fifty-one cultures, fifty-one lawyers. What I found instead was one line, drawn over and over, by institutions that never talked to each other.

The line is this: every one of them lets AI do the work you can cheaply check.

Grammar. Translation. Formatting. Summarizing a document you can read yourself.
And nearly every one of them reserves for humans the work you can't cheaply check: authorship, accountability, original reporting, judgment.

44 of 51 drew that exact line. Six more drew it partway.

One reversed it entirely. I'll get to that one, because the dissent case is the most interesting record in the dataset.

What grabbed my attention wasn't just the pattern. It was hearing the institutions use almost the same sentence to explain it. Independently, Springer Nature and Wiley both say AI can't be an author because accountability "cannot be effectively applied" to a language model.

The APA goes further: AI can't consent to the duties of authorship, like issuing a correction or standing behind a retraction. Read enough of these and the logic underneath becomes visible.

Trust flows to whatever can be held answerable. A machine can show its work; only a human can own it.

The logic is older than the technology. An IBM training manual drew the same line in 1979: a computer can never be held accountable, therefore a computer must never make a management decision.

Notice what that line is not.
→ It's not about capability. Nobody's policy says the machine writes badly.
→ It's not really about ethics either, at least not as the deciding factor.

It comes down to the cost of proof. Where proof is cheap, the machine is welcome. Where proof is scarce, the human persists. The institutions never name that rule. They just keep drawing the same line.

Now the dissent, because demonstrating the pattern with only confirming cases doesn’t warrant a finding. That's a pitch.

EDUCAUSE, the higher-ed technology association, breaks the pattern on purpose. It studied AI-detection tools and concluded they don't reliably work: you cannot cheaply verify whether a student used AI.

So instead of trusting proof, it moved trust somewhere else entirely: into process and governance, into who's in the room when the policy gets written.

Where proof can't be made cheap, trust doesn't die. It relocates. That's the boundary of the pattern, and the boundary is where the science lives.

I'm calling the pattern Calibrated Authority: authority extended to machines exactly as far as their output can be verified, and no further. I've been arguing a version of it for more than two years. What I didn't expect was to find it already written into policy by 44 institutions that have never heard of me.

──────────────────────────────

→ Half-Baked.
[ COMPOUNDER ] A report dies the day you publish it. An index compounds.
This finding could have been a PDF one-pager. Instead it's a living instrument: the corpus re-scans every week, catches policies the moment they move, and logs every change with its receipt. It already caught ACM dropping its AI writing-disclosure requirement. Starting tomorrow it grows institution by institution, each addition published with a written reason it earned the slot. Curation, not volume.

→ Hot Take.
Your AI policy is a confession. It says almost nothing about AI. It says almost everything about which of your own outputs you can actually verify. Show me where you drew the permit line and I'll show you what you can prove.

→ Confession.
A week before launch, my own scanning engine re-scored part of the corpus while I slept. I rolled it back, because the video I'd already recorded said 51. That's the strange new shape of this work: the instrument now moves faster than the person who built it. Related: one of the 51 appears to have genuinely loosened its stance since I coded it. I'm verifying before I publish the re-score. If it holds, that story runs here next.

──────────────────────────────

Most orgs have the line but not the sentence. They permit some AI uses and prohibit others on instinct, and the instinct is usually right, because it's quietly tracking the same thing these 51 institutions track: what can we check, and what would we have to take on faith?

The gap between having the line and being able to say it is where policy fights, stalled pilots, and quiet non-compliance live.

Reply with where your org drew the line. I read every response, and the interesting ones become future issues.

──────────────────────────────

The Monday Move. Print your org's AI policy, or the closest thing you have to one. Mark every rule with one of two letters:
→ C if you could cheaply check the output it governs,
→ F if you'd have to take that output on faith.
If your permitted zone is full of F's, you've found your risk before it found you. That's a thirty-minute audit, and it's the same coding I ran on all 51.

The Asymmetric Move (this quarter). When it feels too risky to permit a certain AI practice, don't reach for a stricter rule. Instead, ask what would make the output checkable: sources that trace, versions that log, provenance built into the pipe. One university in the corpus took that road, building its own AI infrastructure so provenance is a property of the system rather than a policing problem. Make proof cheap and the permit line moves on its own.

The Decade Move (this decade). The institutions that thrive with AI won't be the strictest or the most permissive. They'll be the ones that industrialize verification, because every task where they make proof cheap is a task they can safely hand to machines ahead of their peers. Verification capacity is about to be a competitive asset the way compute is today.

──────────────────────────────

The whole instrument is public, and I built it to be checked, because an index about verification had better be verifiable.

I looked hard for someone else already doing this. Archivists version policy text but don't score it. Academics score policies but don't version them. One snapshot and done. As far as I can find, this is the only instrument that does both continuously, and the first to measure how institutions' division of work between humans and machines moves over time.

Every score traces to the institution's own published words. The dataset is open under CC-BY with a permanent DOI, so researchers can cite a version-stamped snapshot. And it's readable by machines as well as people: AI agents can query the corpus directly, which means when an AI system is asked how institutions govern AI, this index gives it something to check against.

→ Explore the index: https://calibrated-authority.chrishuberreitz.com

→ Cite or fork the dataset: https://github.com/chrishuberreitz/calibrated-authority-index

Borrow freely. If you run the coding on your own institution's policy, send me the result. I'm genuinely interested in where you land.

──────────────────────────────

This newsletter is called Signals from the Curve because there are two kinds of forecasting: the curve and the cliff. The cliff says everything changes at once. The curve says it's already changing, you just have to know where to look.

Last week the classroom. This week, 51 institutions that drew the same line without ever meeting. If something here changed how you're thinking, hit reply. I'll respond.

— Chris

Columbia Faculty · Chief of AI & Strategy at Essential Innovations · Founder, Attainable AI

chrishuberreitz.com · LinkedIn