Simon Willison Backs a Hard Rule for AI Writing: No Rewrite Is Lossless

Simon Willison highlighted a sharp conflict in AI-assisted writing on August 11, 2026: faster rewriting can quietly produce less faithful communication.

The argument comes from software engineer Sophie Alpert, who published an internal policy developed during her work at Clay. It permits AI assistance, but places responsibility for every final sentence on the human author.

That standard challenges a popular assumption about large language models, or LLMs. Many people treat rewriting as a cosmetic operation that preserves ideas while improving tone, structure, or grammar. Alpert argues that no natural-language transformation works that way.

Every rewrite changes emphasis, implication, rhythm, or specificity. The danger grows when the rewriting system lacks the detailed mental model held by the original author.

Willison’s endorsement matters because he has documented many productive uses of LLMs in software development. His support does not reject AI assistance. It draws a boundary between accelerating production and outsourcing judgment.

The central contest is therefore not human writing against machine writing. It is convenience against fidelity, with the author’s accountability sitting between them.

Simon Willison Amplifies an Engineering Writing Policy

The immediate event is small, but the standard it proposes reaches across every document an engineering organization produces.

On August 11, Willison linked to Alpert’s June 25 essay through his AI writing note. He called the short policy “really good” and emphasized its central accountability rule.

Alpert developed the policy for engineers at Clay, then published it as engineering guidance. It covers brainstorming, drafting, proofreading, reviewing, and sharing written work.

The policy does not prohibit those uses. Instead, it requires authors to verify that each idea and sentence represents what they intended to communicate.

“You must stand behind every idea and every sentence in your docs,” Alpert writes. That requirement places authorship with the person sharing the document, regardless of which tool produced its first draft.

The rule becomes concrete during review. If a colleague asks what a sentence means, blaming an AI system does not resolve the confusion. The author selected and distributed that language.

That distinction matters because engineering documents often coordinate consequential work. Technical specifications guide implementation, incident reviews influence future safeguards, and status updates shape staffing or scheduling decisions.

A vague sentence in a casual message might create a brief misunderstanding. A vague requirement in a technical specification can spread across code, tests, interfaces, and operational procedures.

Alpert’s title captures the mechanism behind this risk. Natural-language transformations are not lossless because wording carries more than explicit facts.

Sentence order determines priority. Examples restrict interpretation. Qualifications show confidence. A deleted caveat can turn a tentative observation into an apparent commitment.

An LLM can produce cleaner prose while altering any of these elements. The result may sound more polished even when it represents the author less accurately.

That is why fluency provides weak evidence of fidelity. A readable paragraph can still contain a shifted claim, unsupported connection, or misplaced emphasis.

Willison focuses on ownership rather than disclosure alone. Identifying that AI participated in drafting may improve transparency, but it does not make a misleading document useful.

Readers still need a document they can trust. They also need an author who can explain its reasoning, defend its claims, and revise it when challenged.

The policy therefore changes the completion standard. A document is not ready when the model finishes writing. It is ready when the author understands and accepts the entire result.

Faster Drafting Transfers Work to Readers

AI-generated length creates a hidden organizational cost when authors save minutes by consuming hours of collective attention.

Alpert argues that authors should spend more time producing a document than each reader spends consuming it. That principle reflects the unequal economics of workplace writing.

One person usually creates a technical document. Several engineers, managers, designers, or customers may need to read it.

A five-minute shortcut for the author can impose several extra minutes on every recipient. Across a large review group, the saved effort becomes a net loss.

This problem appears when someone gives an LLM a short prompt and distributes the longer response with minimal revision. The author has effectively delegated interpretation to the audience.

Readers must locate the actual decision, separate evidence from filler, and determine which statements express the author’s position. They may also need to resolve contradictions introduced during generation.

A document can be grammatically correct and still impose this burden. The problem is not necessarily bad prose. It is uncertain intent.

Engineers already face crowded information environments. They move between pull requests, tickets, design documents, chat threads, dashboards, and incident channels.

Adding plausible but weakly reviewed text increases the amount of material that requires judgment. It does not automatically increase the amount of useful knowledge.

This creates pressure for engineering leaders as well as individual contributors. They must decide whether document volume measures communication or merely records tool output.

Policies that celebrate faster drafting without measuring reader effort favor visible production. They overlook the downstream work needed to interpret, verify, and correct that production.

Alpert’s standard reverses that incentive. The author should absorb the one-time cost of making a document concise because every reader benefits afterward.

The same logic appears in Atlassian’s June 2026 writing guidelines. Its brand team warns that generic, unverified material degrades internal knowledge systems like repeated copies of an image.

That comparison points to a second cost. Weak documents do not disappear after their first reading. Organizations store them in search indexes, knowledge graphs, project archives, and AI retrieval systems.

Once stored, a distorted summary can influence later decisions. Another model may retrieve it, compress it again, and present the result without the original context.

The organization then faces a chain of lossy transformations. Each step can remove caveats while increasing the appearance of institutional authority.

A personal knowledge system faces the same challenge. Capturing everything provides little value if summaries no longer preserve why a decision was made.

Good knowledge management depends on provenance, context, and retrieval quality. More text alone does not provide any of those properties.

The pressure falls most heavily on people who consume documents at scale. Staff engineers, security reviewers, executives, and incident commanders often review many artifacts from different authors.

They cannot reconstruct every author’s hidden intent. They need the submitted document to contain the relevant distinctions.

This is why “the AI wrote it” fails as an explanation. It identifies a production method while leaving the reader’s problem untouched.

The forced response is not a blanket ban. Teams need review expectations that treat AI output as unfinished material until a responsible author adopts it.

AI Rewriting Changes Meaning, Even When Facts Survive

The core tradeoff is simple: rewriting can improve surface clarity while weakening the connection between language and the author’s actual mental model.

Natural language does not separate content from presentation as cleanly as a structured data format. Word choice and organization help determine what a statement means.

Consider an engineer who writes, “The cache probably caused the latency spike, but the deployment change remains another plausible factor.”

A rewriting tool might return, “The latency spike was caused by the cache.” The sentence is shorter and more confident. It is also a different claim.

Another rewrite might preserve both factors but place the deployment first. That change can affect which investigation receives attention.

Even a neutral-looking summary selects what to omit. It may preserve the conclusion while removing the evidence, uncertainty, or rejected alternatives that made the conclusion defensible.

The transformation becomes riskier when the prompt provides only a rough idea. The model must supply transitions and assumptions that the author never explicitly stated.

Those additions can create an argument that reads coherently but was not actually reasoned through. Coherence then hides a gap in authorship.

Alpert describes writing as a form of thinking. The work of ordering claims, choosing examples, and addressing objections can reveal defects in the underlying idea.

Skipping that work may save drafting time while preserving confusion. The author receives a finished-looking artifact before building a finished argument.

This concern is particularly relevant for engineering specifications. A specification is both communication and evidence that someone examined constraints, alternatives, and failure cases.

An AI system can imitate the shape of that document. It can produce sections for goals, architecture, security, migration, and open questions.

That structure does not establish that the author evaluated those subjects. A complete template can conceal incomplete thought.

The policy therefore distinguishes assistance from substitution. Brainstorming can expose alternatives. Proofreading can identify awkward passages. A draft can help an author find a starting point.

However, the human must still decide which claims belong, what they mean, and whether the document deserves its readers’ time.

This is not unique to factual hallucination. A model can preserve every named fact and still distort the message through emphasis.

It might soften a warning, strengthen a forecast, remove an uncomfortable exception, or replace a precise term with a broader synonym.

Tone transformations also carry meaning. Rewriting a direct disagreement as polite corporate language can hide the urgency or certainty behind the original message.

Translation presents related risks. A translated sentence may preserve its central proposition while changing cultural cues, technical precision, or the perceived strength of a recommendation.

These transformations can be useful. They are simply not free.

An author who accepts a rewrite needs to compare the output against the intended meaning, not only against grammar rules. That requires close reading and subject knowledge.

This review becomes harder as generated text grows longer. The author may skim familiar-looking prose and overlook small semantic changes.

The polished style creates automation bias, which is the tendency to trust output because a system produced it confidently. Fluency can intensify that bias.

Meaning loss also complicates later retrieval. Search and summarization systems often rank a concise declarative sentence above a nuanced discussion.

If the concise sentence overstated the source, future readers may encounter the distortion first. The transformed version can effectively replace the original in organizational memory.

The phrase “no lossless transformations” should not be read as a demand to preserve every sentence forever. Editing remains necessary, and human editing changes meaning too.

The difference is accountability. A human author can intentionally decide which meaning to preserve, refine, or remove.

An LLM does not possess the author’s complete purpose. It estimates a likely continuation or rewrite from the available context.

That mechanism can produce excellent language. It cannot guarantee alignment with an unstated intention that exists only in the author’s mind.

Accountability Matters More Than Detecting AI Prose

Teams gain more from assigning responsibility for final claims than from trying to identify which sentences began with a model.

Many AI writing debates focus on detection. Reviewers look for repeated phrases, uniform structure, inflated transitions, or familiar stylistic patterns.

Those signals are unreliable foundations for governance. Human authors can write generic prose, while careful model-assisted writing can avoid obvious patterns.

Detection also targets the wrong outcome. The operational question is whether the document is correct, purposeful, explainable, and worth reading.

A fully human draft can fail those tests. A heavily assisted draft can pass them if the author carefully verifies and adopts every part.

This does not make disclosure irrelevant. Some settings require it because readers need to understand the production process or evaluate conflicts.

Academic publishing provides one example. Elsevier’s current author accountability rules allow supportive uses while requiring human oversight, verification, and disclosure.

Its policy also says AI tools cannot receive authorship because authorship carries responsibilities that belong to people. That reasoning closely matches Alpert’s workplace standard.

News organizations have reached similar positions. Updated newsroom standards from the Associated Press permit limited assistance while retaining editorial oversight and established reporting requirements.

These policies differ in scope, but they share a practical insight. A tool cannot absorb professional accountability for a published claim.

Engineering teams can translate that insight into review behavior. The named author or document owner should answer questions without treating generated passages as disposable accidents.

Reviewers should challenge unexplained certainty, missing evidence, and vague recommendations. They should not spend their limited attention guessing which tool produced the language.

Organizations also need to protect confidential information. A team can produce accurate prose while violating policy by sending sensitive code, customer data, or internal plans to an unsuitable service.

That issue sits beside semantic fidelity rather than replacing it. A responsible AI writing policy must cover both information handling and final authorship.

Fast Company’s editorial policy illustrates this broader approach. It requires substantive human editing and addresses tool retention settings, confidential content, and accountability.

The difficult question is what counts as sufficient review. A quick read can confirm that a paragraph sounds reasonable without confirming every implication.

Teams should avoid pretending this threshold has a perfect numerical test. Review depth depends on the document’s purpose and potential consequences.

A casual meeting recap requires less scrutiny than an incident report. A customer-facing security statement demands more review than an internal brainstorming note.

The safest principle scales responsibility with impact. Authors should verify facts, claims, and commitments according to the harm a misleading sentence might cause.

Clear provenance can support that process. Teams can preserve source notes, decision records, linked evidence, and earlier drafts when those materials clarify how conclusions developed.

A searchable technical knowledge base can keep source documents near summaries. That proximity helps readers inspect context instead of trusting compressed prose alone.

Still, provenance cannot rescue a document whose owner never understood it. Links provide evidence, but someone must decide what the evidence supports.

The skeptical case against Alpert’s policy is that demanding sentence-level ownership may discourage useful assistance. Some writers communicate better after receiving structural or language support.

Non-native English speakers may especially benefit from tools that improve clarity. People with disabilities may use generation or rewriting as an accessibility aid.

Those benefits are real, and the policy does not require unaided composition. It requires informed adoption of the final result.

However, the burden of reviewing every sentence can approach the burden of writing from scratch. For some tasks, that reduces the efficiency advantage.

That is not necessarily a flaw in the rule. It exposes where claimed productivity came from shifting cognitive work to readers or reviewers.

The uncertainty lies in whether future systems can reduce that review burden without weakening fidelity. Better context, source linking, and tracked semantic changes might help.

None of those features can fully observe an intention that the author never expressed. The author must remain part of the verification loop.

The Policy Draws a Better Boundary Than an AI Ban

A useful policy separates acceptable assistance from unowned output instead of treating every interaction with an LLM as equivalent.

Blanket bans are easy to state and difficult to enforce. AI capabilities already appear inside editors, communication platforms, search tools, and operating systems.

A ban also groups very different actions together. Checking grammar does not carry the same risk as generating an incident analysis from a one-line prompt.

Alpert’s framework permits brainstorming, drafting, and proofreading. Its boundary appears at the point where material becomes a claim made to other people.

This approach gives engineers flexibility while preserving an identifiable owner. It also matches how teams already handle other forms of collaboration.

An engineer can receive edits from a colleague, but the document owner still approves the final version. A reviewer’s suggestion does not transfer authorship automatically.

The same logic can apply to an LLM. Its suggestion remains proposed language until the author examines and adopts it.

Teams can make this boundary concrete through several review questions.

Does the author understand every claim? Can the author explain why each section exists? Are uncertainty and alternatives represented accurately?

Does the document cite the evidence needed for consequential decisions? Would a shorter version serve readers better?

These questions evaluate the output without requiring a forensic reconstruction of the drafting process. They also work for documents produced without AI.

Tracked changes can help when models rewrite existing prose. The author can inspect deletions, additions, and altered qualifications instead of comparing two long documents from memory.

Prompting the model to preserve uncertainty can reduce obvious distortions. Supplying source material can also make unsupported additions easier to identify.

Neither technique makes the transformation lossless. They improve the conditions for human review.

A stronger workflow asks the author to begin with concrete notes, decisions, and evidence. The model then works on material that already reflects substantial thought.

This differs from requesting a complete specification from a short prompt. The first workflow assists expression. The second asks the model to invent much of the reasoning.

The distinction is not always clean. Drafting and thinking interact, and a model-generated question can trigger a genuine insight.

The ownership rule handles that ambiguity better than a task list. Whatever path produced the insight, the author must understand and defend the final statement.

Managers should also avoid measuring success through document throughput alone. More specifications, updates, and summaries do not guarantee faster or better decisions.

Useful measures include review time, clarification requests, reopened decisions, factual corrections, and downstream implementation errors.

A policy should also define low-stakes exceptions. Clearly marked model output can be useful when a team wants to examine an idea generated by a tool.

Alpert explicitly allows verbatim AI output when it is identified as such. Readers then know they are evaluating a suggestion, not the author’s settled position.

That label changes the conversational contract. The author is presenting an object for discussion rather than claiming ownership of its reasoning.

The distinction prevents a common failure. Teams should not treat tentative generated text as a decision record simply because it appears in a formal document.

The best boundary therefore depends on communicative status. Drafts, questions, and experiments can remain provisional.

Specifications, retrospectives, recommendations, and public statements need accountable owners. Their language directs action and shapes institutional memory.

What Simon Willison’s AI Writing Rule Puts on Watch

The next test is whether engineering teams build workflows that preserve intent, rather than merely producing cleaner text faster.

The first signal is the spread of explicit authorship rules inside engineering organizations. Useful policies will identify who owns generated claims and what review completion means.

If more teams adopt this standard, Willison and Alpert’s argument gains practical support. It would show that organizations see semantic fidelity as an operational issue.

The judgment weakens if teams achieve reliable communication with lighter review and no increase in corrections. That result would suggest current concerns overstate the cost.

The second signal is product support for semantic review. Writing tools can show changed claims, removed qualifications, unsupported additions, and altered certainty.

Ordinary tracked changes reveal modified words. More useful systems would explain how the proposed rewrite changed the argument.

If these capabilities become dependable, AI assistance can save time without hiding transformations. Authors would still own the output, but they could inspect it more effectively.

The argument weakens if tools claim semantic preservation without transparent evidence. A confidence label alone cannot establish that a rewrite matches an author’s private intention.

The third signal is downstream document quality. Teams should watch clarification volume, review duration, incident follow-ups, and conflicts between summaries and source records.

A rise in polished documents alongside more corrections would support the lossy-transformation thesis. It would show that surface quality masked communication debt.

Stable or improving outcomes would not prove rewrites are lossless. They would show that human review and better tooling can manage the loss successfully.

Readers should also watch how AI retrieval changes the stakes. Internal assistants increasingly answer questions from stored specifications, meeting notes, and status reports.

A weak document can therefore influence more than its original audience. It can become retrieved evidence for future generated answers.

That multiplier makes source quality important. Organizations need ways to distinguish reviewed decisions from provisional drafts and generated suggestions.

Simon Willison’s intervention offers a useful standard for that environment. Use AI where it helps, but do not confuse generated fluency with adopted thought.

Before sharing an AI-assisted document, ask whether you can explain every line without referring responsibility back to the model. Then remove anything you cannot defend.

Preserve the evidence behind important claims, keep uncertainty visible, and give readers the shortest document that supports the decision. Treat the final text as your own work.

The practical question is not whether AI touched the document. It is whether the document still carries your meaning after every transformation.