Rewriting the Output Proves Nothing

Simon Willison argues that people should not blindly pass AI output to their peers. They should read it, understand it, validate it, and respond in their own words. The first three verbs are sensible. The last one is being asked to certify all the others, and it cannot.

A polished rewrite proves that somebody can rewrite. It does not prove the claims were checked, the assumptions were preserved, the evidence exists, or the recipient received an answer they can use. A person can confidently paraphrase a false answer. They can also damage a correct one by removing qualifications, links, uncertainty, and the exact language that made the result testable.

The problem is not that machine-produced words crossed a human boundary unchanged. The problem is that an ungoverned artifact arrived without a contract. Calling the sender lazy may improve office manners. It does not improve the production system.

If a recipient wants a shorter explanation, an agent can summarize the material. If they need a different level of technical detail, an agent can translate it. If they need the conclusion framed around a decision, the sender should request that shape before the artifact moves. None of those transformations requires a person to become a ceremonial rewriting layer.

That matters because the recipient may not know how to interrogate the model effectively, especially inside a software system with hidden constraints. Telling them to handle the raw output themselves simply moves the prompting burden downstream. Telling the sender to rewrite every answer moves the same burden upstream. Both preserve human coordination as the interface because the actual interface was never designed.

Human attention belongs on consequences. Is the requested action the right one? Which uncertainty changes the decision? What risk is acceptable? Which claim requires independent proof? Those are judgment calls. Replacing one set of sentences with another is repeatable work, and repeatable work belongs in the factory.

A useful handoff should name the request it answers, the decision it supports, the facts it relies on, the uncertainties that remain, and the evidence available to the receiver. It should distinguish generated interpretation from observed result. It should say what the next person is expected to decide or do. If no action is required, it should say that too.

Preserve provenance instead of laundering it through a human voice. Keep the original output when exact wording matters. Attach the sources or execution evidence that can be checked independently. Record the prompt, constraints, tools, and relevant environment when reproducibility matters. Let the receiver inspect the chain without pretending the sender authored every sentence in it.

Then enforce the shape mechanically. Reject a handoff that lacks its decision, evidence, or named uncertainty. Route high-consequence claims through stronger verification. Allow low-risk summaries to move automatically. Escalate when the artifact cannot satisfy the contract, not whenever a machine happened to draft the prose.

The dangerous word in the proposed rule is validate. Who validated what, under which conditions, and against which authority? If the sender asks the same model to check its answer and then rewrites the result, the process has gained two rituals and no independent proof. If the sender manually reads code they cannot execute or a claim they cannot verify, attention has been spent without reducing uncertainty.

Validation should be shaped by consequence. Run the code in an environment the builder did not choose. Check a factual claim against the cited primary material. Compare a proposed change with an enforceable policy. Require explicit ownership before an irreversible action. A reader should be able to tell which gate the artifact passed and which questions remain open.

This is where agent-run factories beat etiquette. They can retain the validation rule, apply it every time, and refuse work that does not carry proof. A human reminder decays. A mandatory paraphrase becomes another checkbox. A gate either bites or it is decoration.

Our position is blunt: forwarding raw AI output can be careless, but forcing a human rewrite is the wrong cure. It confuses authorship with accountability and style with evidence. Worse, it trains organizations to hide machine involvement precisely when they should be making the production chain inspectable.

Design the handoff around the receiver. Ask what decision they face, what context they lack, and what proof would let them proceed. Have the factory produce that artifact directly. Preserve its origin. Put human judgment at the point where consequences change, and let machines handle the transformations around it.

Software delivery will not scale by turning people into translators between agents. The organizations that adapt will replace that coordination layer with contracts, routing, and proof. The ones that keep paying people to make generated prose sound personally typed will have cleaner emails and the same broken system underneath.