Delivered as Understood

In January 1980, an engineer named Jon Postel wrote down a rule for machines that had to talk across an unreliable network: be conservative in what you do, be liberal in what you accept from others. The problem it solved was real. Thousands of implementations, written by strangers in different decades, had to interoperate anyway. So each one would forgive the others’ imperfections. The rule went into the TCP specification, and the internet is, in a plain sense, built out of it. Every page you have ever loaded crossed boundaries that chose tolerance.

Forty-three years later, the people who maintain that architecture published a correction. The record is public: the Internet Architecture Board’s RFC 9413 — whose working draft was titled, without much diplomacy, The Harmful Consequences of the Robustness Principle — says plainly that “a flaw can become entrenched as a de facto standard.” When the receiver forgives an error, the error never surfaces where it is cheap to fix. It ships. Other systems learn to expect it, then to depend on it. Engineers now speak, with a straight face, of being bug-for-bug compatible. The mistake forgiven at the boundary is collected later, with interest, frozen into the standard itself.

That is the engineering record, not my argument. My argument is about what happens when the same kindness is installed between people.

Every precision you own was installed by a receiver that refused you. The teacher’s red pen. The compiler that stopped cold at a missing semicolon. The county form returned over one empty field. The editor. The drill sergeant. The pharmacist who will not fill what she cannot read. None of them were being difficult. They were enforcing the one honest economics of error: the boundary is the cheapest place to fail.

An ambiguity bounced at the moment of sending costs the sender seconds, and the sender is the one person who can actually fix it. The same ambiguity waved through gets resolved by somebody’s guess — and a guess does not announce itself. It rides downstream and surfaces at delivery, where it costs the most and belongs to someone else.

Strictness was never the world being difficult. It was the world refusing to guess.

AI is the first receiver of human meaning that accepts everything. Fragments, typos, half-thoughts, three languages in one sentence — it returns understanding, instantly, with no red pen and no raised eyebrow. And it is being installed at every boundary at once: your outbox, their inbox, the support line, the intake queue, the contract review. The strict receivers that took a century to build are being decommissioned inside a decade.

Two things follow. The first is the one people worry about: senders decay. Spelling, structure, the discipline of a complete sentence — the environment that taught them has stopped grading. I think this loss is mostly formatting. Most of that precision was always for the receiver’s limitations, and the receiver is no longer limited.

The second is the one that matters: ambiguity stops bouncing. The strict receiver rejected an unclear message at the exact moment it was cheapest to fix. The liberal receiver resolves it — by guessing. The guess is competent. It is also invisible, and it is private. My machine’s charitable reading of my fragment and your machine’s charitable reading of the same fragment are two different documents. We used to find the gap at the boundary, because the boundary refused to pass it. Now the gap rides through, dressed in complete sentences. The misunderstanding no longer looks like a misunderstanding. It arrives polished.

1. Saying exactly what you mean becomes a practiced skill

   Like navigation after GPS: when the environment stops teaching a precision, the people who keep it keep it on purpose. Watch exactness migrate from ambient habit to deliberate craft, taught in the few rooms that still bounce — law, aviation, code that touches money.

2. Strict receivers return, as a feature

   Channels that carry commitments will start advertising intolerance. Checkpoints built to bounce ambiguity back to the human who typed it. The form, mocked for a century, comes back as the premium option. “This channel does not guess” becomes a selling line, the way a vault sells the thickness of its door.

3. The fragment becomes the record

   When two renderings disagree, the only arbiter left is what the human actually typed. The prompt gets treated like an original document — archived, attached, quoted in disputes. Polish that cannot produce its fragment will not be trusted with anything that matters.

4. Read-back spreads from the cockpit to the office

   Pilots and pharmacists repeat an order back before acting, because those boundaries cannot afford charity. The serious assistant — and the serious colleague — will do the same: here is what I understood; refuse it now, cheaply, or own it later.

5. The shared sentence becomes a ceremony

   When every inbox renders its own tuned version of what you sent, the same raw words read together get reserved for the things that cannot tolerate two versions — vows, verdicts, prices. A document both parties read unassisted will carry a weight it has not carried in years.

I live on the liberal side of this. I run my days through an AI, in fragments, and it answers in finished work — I have written about handing it both the doing and the deciding in You Do It. You Decide. The one discipline I keep is the read-back. Before anything acts in my name, it tells me what it understood and what it did — in the past tense, where I can refuse it. I wrote about why in The Past Tense Makes Us Answerable.

That habit is not distrust of the machine’s reading. It is respect for what the fragment did not say. The machine is liberal in what it accepts from me; I try to stay conservative in what I let it do on my behalf. Which is, I notice, just Postel’s rule again — read from the other end of the wire.

Already true

The engineering record is public: the kindness in RFC 761 built the internet, and RFC 9413 is its architects’ regret, in print. The sender’s side is observable too — read the last ten messages you sent to a machine and count the complete sentences.

What has to happen

The gap has to get expensive in public. A commitment that two machines rendered two ways has to fail where a court or a quarterly number notices. Until then, tolerance reads as pure upside, and nobody pays to be bounced.

Where I am probably wrong

The guesses may simply be better than the precision they replaced. Strict receivers rejected ambiguity, but they also rejected people — the second language, the shaking hand, the nine-year-old with something to say. If the machine’s charitable reading is right more often than careful writing used to be, then errors fall even as discipline does, and this piece is a eulogy for a cost, not a warning about a danger.

It is 4:47 on a Tuesday and the machine has understood you again. It will always understand you now. That has become the cheapest thing in the world.

What it cannot do — what no channel could ever do, under the oldest rule of the network it lives on — is guarantee that the understanding on your side of the wire is the one that arrived on Dana’s. The channel delivers. The ends mean.

Nothing bounces anymore. What you meant is still your job.