Simon Willison Says Don't Be a Meat Proxy for AI

Simon Willison issued a blunt warning on August 3: do not become a “meat proxy” who blindly forwards AI-generated answers. The phrase names a familiar workplace failure with unusual precision. Someone asks a question, another person prompts a model, and its response travels onward without meaningful human review.

The expression comes from Niklas Gruhn, whom Willison credits in his August 3 post. Willison’s advice does not reject prompting or AI-assisted writing. It demands that users read, understand, validate, and rewrite the result before sharing it.

That distinction challenges a common promise behind workplace AI. Generative systems can reduce the time needed to produce text, but speed is not the same as completed work. When a human merely transports model output, the recipient inherits the verification burden that the sender avoided.

The real contest is therefore not humans against AI. It is accountable assistance against unexamined delegation. The first route uses a model to accelerate thinking. The second disguises an uncertain model response as a colleague’s considered judgment.

Simon Willison Gives a Name to Blind AI Relaying

“Meat proxy” describes a person who serves as the delivery layer between an AI system and another human.

Willison’s short post crystallizes a behavior that has spread through email, chat, documents, code reviews, and support systems. A user receives a question and puts it into a large language model. The generated answer is then pasted into the original conversation with little review or transformation.

The phrase matters because “AI-generated content” is too broad. It can describe anything from a carefully checked draft to an untouched response. “Meat proxy” isolates the failure at the center of the second case: the nominal author contributes neither verification nor accountable judgment.

Willison recommends a specific boundary. Prompting an AI system is acceptable, but its answer should remain an intermediate artifact. The user must read it, understand its reasoning, validate its claims, and respond in their own words.

Rewriting is not merely a stylistic preference. It provides evidence that the sender processed the answer deeply enough to explain it. That evidence is imperfect, since a person can paraphrase nonsense, but it creates useful friction before publication.

Consider a routine engineering exchange. A teammate asks why a service returns intermittent authentication errors. Another engineer pastes the question into an AI assistant and forwards a polished explanation about expired tokens.

The explanation sounds plausible, yet the actual cause might be clock drift, a bad deployment, or an unrelated proxy configuration. Forwarding the answer transfers a guess into a channel where readers may treat it as investigated fact.

The same pattern appears in management. An executive asks for the risks of entering a new market, and an analyst relays a generic model-generated assessment. The response may omit local regulations, existing contracts, or information held inside the company.

The visible document looks finished, but the analytical work remains undone. Its fluency conceals the gap between producing an answer and establishing that the answer fits the situation.

“Meat proxy” gives teams a concise way to identify that gap. It also places responsibility on the person choosing to forward the output, not only on the model that produced it.

Fluent Output Pushes Verification Downstream

A meat proxy does not remove work from a conversation. The proxy moves difficult work to the next person.

Generative AI makes drafting cheap, fast, and abundant. Verification still requires source checking, domain knowledge, contextual awareness, and decisions about uncertainty. Those tasks often take longer than generating the first response.

Blind relaying creates an asymmetric exchange. The sender spends seconds producing several paragraphs. The recipient must inspect every factual claim, trace missing context, detect contradictions, and decide whether the advice is safe to use.

That imbalance becomes worse as the generated answer grows. A long response can contain many individually plausible claims, each carrying a small chance of error. The recipient cannot infer reliability from confident language or polished organization.

Researchers call one part of this problem automation bias, the tendency to over-rely on automated recommendations despite reasons for caution. A review of 35 peer-reviewed studies found that over-reliance remains a central challenge in human and AI collaboration.

The risk predates large language models. Automated systems have influenced decisions in aviation, medicine, public administration, and other fields for decades. Generative AI expands the concern because it can produce persuasive explanations for almost any topic.

Traditional software often exposes its limits through structured inputs and predictable outputs. A conversational model can answer questions even when it lacks relevant evidence. That flexibility is useful, but it also makes unsupported responses harder to recognize.

The recipient faces a social signal as well as a technical one. A message from a colleague normally implies that the colleague selected, understood, and stands behind its contents. Undisclosed AI relaying weakens that implication without changing the message’s appearance.

This is why disclosure alone cannot solve the entire problem. Labeling text as AI-generated warns readers, but it can also announce that the sender expects them to perform the review. The sender still needs to determine whether the material deserves transmission.

The burden becomes especially visible in code. A generated patch might compile and pass basic tests while introducing insecure assumptions or subtle maintenance problems. Reviewers must then reconstruct an approach that the submitting developer never fully understood.

Support teams encounter a related failure. An agent can forward an answer that follows a familiar template but conflicts with the customer’s account, contract, or earlier troubleshooting. The customer then pays for the missing contextual check through delay and frustration.

The issue is not that every AI-assisted message needs original research. A scheduling note or basic rewrite carries less risk than legal guidance or a production change. The required review should scale with the consequences of being wrong.

Still, even low-stakes messages affect trust. Colleagues quickly notice replies that repeat their question, ignore shared context, or offer five generic options without a recommendation. Those patterns signal that nobody meaningfully received the request.

Willison’s framing turns that annoyance into a labor question. If the sender did not understand or evaluate the response, the human contribution is only transportation. The remaining cognitive work has not disappeared.

The Real Contest Is Assistance Versus Delegation

AI assistance keeps judgment with the user, while AI delegation quietly hands judgment to a probabilistic system.

A large language model, or LLM, generates text by predicting likely token sequences from learned patterns and supplied context. It does not automatically verify every statement against authoritative evidence. It also does not know the sender’s unstated obligations.

Assistance begins when a user gives the system a bounded role. The model might suggest questions, reorganize notes, identify possible counterarguments, or draft language from verified facts. The person remains responsible for selecting and checking the result.

Delegation begins when the model’s output becomes the answer simply because it exists. The user stops treating generation as one step in a workflow. The draft bypasses analysis and enters a human conversation as a finished product.

The difference can be subtle. Both users may start with the same prompt and receive identical text. One investigates the citations, removes unsupported claims, adds local context, and changes the conclusion. The other presses copy and paste.

Only the first user adds defensible value. That value appears in omissions as much as additions. Good judgment often means deleting confident speculation, narrowing a recommendation, or acknowledging that available evidence cannot settle the question.

Public-sector research shows why human presence alone offers weak protection. Across three experiments, researchers studying algorithmic advice found evidence of automation bias and selective adherence to recommendations that matched existing stereotypes.

A person can therefore remain formally “in the loop” while contributing little oversight. The meaningful question is whether the person can challenge the system, recognize a bad answer, and take responsibility for rejecting it.

That distinction pressures organizations adopting AI through simple output targets. If leaders measure emails sent, tickets closed, reports produced, or code submitted, generated volume looks like success. The metric does not reveal who absorbed the verification cost.

Employees also face mixed incentives. They are encouraged to use AI and deliver faster, but they may receive little time for checking the generated work. Under those conditions, meat-proxy behavior can become a rational response to an irrational workflow.

Managers cannot fix that by ordering everyone to “use judgment.” Teams need ownership boundaries. The sender should know which claims require evidence, which decisions need review, and what uncertainty must be disclosed.

A practical workflow starts with the underlying material, not the prompt. Users should gather relevant documents, previous decisions, source data, and constraints. A searchable knowledge base can make that context easier to retrieve, but retrieval still requires evaluation.

The model can then help interrogate that material. It might compare documents, locate disagreements, produce a draft, or propose missing questions. The user must decide whether the response reflects the evidence and the actual request.

Finally, the sender should rewrite the result around a clear conclusion. The response needs an accountable recommendation, supporting evidence, visible uncertainty, and any required next action. A generic model answer usually lacks at least one of those elements.

This approach does not demand ceremonial rewriting of every harmless sentence. It demands intellectual ownership. If the sender cannot explain why the answer is correct, relevant, and appropriately qualified, the answer is not ready to send.

Writing in Your Own Words Is a Check, Not a Cure

Rewriting raises the cost of careless forwarding, but polished paraphrasing cannot make an unsupported claim true.

Willison calls an original response a decent certificate that the writer completed the earlier steps. The careful word is “decent.” Writing in one’s own words creates evidence of engagement, not proof of factual accuracy.

A user can misunderstand the model and restate its error elegantly. Another can use an AI paraphraser to disguise an untouched response. Organizations should not confuse surface originality with genuine verification.

The strongest part of the proposal is the sequence behind the rewrite: read, understand, validate, then respond. Each verb addresses a different failure. Reading catches obvious irrelevance. Understanding tests whether the user can explain the reasoning.

Validation compares claims against evidence, systems, or qualified expertise. Rewriting then forces the user to choose what matters and accept authorship. Skipping validation turns the last step into cosmetic laundering.

Experiments suggest that deliberate friction can help. In one 2025 study, participants receiving faulty AI support answered fewer than half as many reflection questions correctly as a control group. A warning nudge almost doubled performance compared with faulty assistance alone.

The warning did not raise performance above the no-support group. Self-reported AI literacy also did not significantly prevent automation bias in that experiment. Familiarity with prompting, therefore, should not be mistaken for reliable skepticism.

A useful review process must test substance. For factual writing, users can identify each material claim and attach an authoritative source. For code, they can explain the change, run relevant tests, and inspect security-sensitive behavior.

For analysis, users can state the assumptions that drive the recommendation. They should also ask what evidence would reverse their conclusion. This question exposes whether the text contains an argument or only a plausible narrative.

For external communication, writers should compare the draft with the recipient’s actual needs. A customer asking whether an incident exposed data does not need a general explanation of encryption. They need a supported answer, known limits, and the next update time.

The same discipline applies to summaries. A model may compress a meeting transcript accurately while missing the political significance of one objection. Someone present must decide which statements were commitments, suggestions, or unresolved disputes.

Personal context can improve generation, but it cannot replace this decision. A personal knowledge workflow can preserve sources and prior reasoning. The user still owns the interpretation shared with others.

The skeptical case against Willison’s wording is worth taking seriously. “Meat proxy” can become an insult aimed at junior employees, non-native speakers, or people using AI for accessibility. It can also reward invisible AI use when polished rewriting hides the system’s role.

Teams should use the term to diagnose a workflow, not shame a person. The relevant standard is accountable handling of output. It should apply equally to executives, developers, analysts, contractors, and managers.

Nor should every AI-assisted sentence require disclosure. Disclosure rules depend on risk, policy, and audience. The central obligation is stronger than a label: do not make another person rely on material that nobody has responsibly checked.

There is also a danger of romanticizing unaided human work. People copy bad advice, misremember facts, and forward irrelevant documents without AI. Human authorship has never guaranteed truth.

Generative AI changes the scale and presentation of that old problem. It produces polished responses faster than humans can inspect them. The meat-proxy label is useful because it focuses on the missing review rather than pretending errors began with AI.

Meat Proxies Amplify Bias Along With Errors

Unchecked relaying can transmit a model’s framing and omissions even when every individual sentence appears defensible.

Factual hallucinations attract attention because they are easy to demonstrate. A model invents a citation, misstates a date, or attributes words to the wrong person. Verification can often identify those failures.

Framing errors are harder. A generated answer may emphasize one explanation, assume a particular audience, or omit an affected group. No single sentence needs to be false for the response to steer a conversation in a distorted direction.

Researchers describe communication bias as the expression or amplification of perspectives through generated output. A 2026 bias analysis argues that automation bias can increase the effect when users accept and act on such output without enough scrutiny.

A meat proxy can become the mechanism that carries this framing into institutional decisions. The person lends the output their identity and organizational position. Recipients may never see the prompt, alternative outputs, or assumptions that shaped the answer.

Imagine a hiring manager asking a model to summarize interview notes. The model might foreground confidence, leadership language, or conventional career paths. If the manager forwards that summary, subjective framing gains the appearance of neutral synthesis.

A product team might ask for likely reasons behind declining usage. The model could produce familiar explanations about onboarding and pricing while missing accessibility problems described in internal research. A fluent answer can narrow investigation prematurely.

An executive briefing creates similar risks. Models tend to produce balanced, orderly prose, even when underlying evidence is incomplete or contradictory. That style can make genuine uncertainty look like a settled list of manageable considerations.

Prompting does not eliminate these problems. A better prompt can request counterarguments, citations, or uncertainty, but the resulting text still needs evaluation. Prompt technique controls the request, not the truth of the response.

This is where the primary contest becomes organizational. Accountable assistance creates a chain from evidence through analysis to a named decision-maker. Blind delegation creates text whose apparent owner cannot defend its construction.

The distinction matters for corrections. When a checked human-authored response proves wrong, colleagues can examine assumptions and improve the process. When a pasted answer fails, the sender may only say that the model produced it.

That response breaks organizational learning. The team cannot tell whether the failure came from missing context, weak sources, a misleading prompt, model behavior, or careless review. Responsibility dissolves across the workflow.

AI providers also face pressure. Interfaces optimized for immediate answers encourage users to treat output as complete. Citations, uncertainty indicators, source views, and verification prompts can create healthier pauses, although none guarantees careful use.

Organizations should watch their own systems for downstream signals. Corrections, reopened tickets, reverted code, repeated questions, and recipient complaints can expose work that looked productive at the generation stage.

The best metric may be whether AI-assisted output reduces total effort for everyone involved. If the sender saves ten minutes but three colleagues spend an hour checking the result, the workflow did not create efficiency.

Measuring total effort is harder than counting generated documents. It is also closer to the value that AI adoption is supposed to deliver.

What Happens After Simon Willison’s Warning

The next test is whether teams turn a memorable phrase into observable standards for accountable AI use.

Three signals will show whether “meat proxy” becomes more than a clever definition. The first is the appearance of review requirements inside everyday AI workflows. Teams should specify what the sender must check before generated content reaches another person.

These requirements need to be proportional. A casual brainstorming note can tolerate uncertainty that a legal filing cannot. Production code, customer commitments, financial analysis, and personnel decisions deserve explicit evidence and named reviewers.

If organizations adopt such standards, Willison’s argument gains force as an operating principle. If they rely only on broad reminders to “check AI output,” meat-proxy behavior will remain easy to deny and difficult to measure.

The second signal is product design. AI interfaces can expose sources, distinguish retrieved material from generated interpretation, flag unsupported claims, and ask users to confirm high-impact actions. These controls add friction where confidence would otherwise travel unchecked.

Research on warning nudges suggests that interface choices can improve critical reflection, even if they cannot restore fully independent performance. A wave of verification-oriented features would strengthen the case that output review is a distinct product problem.

The opposite trend would weaken it. If assistants increasingly optimize for one-click sending, autonomous execution, and invisible insertion into workplace channels, users will have fewer natural opportunities to inspect what represents them.

The third signal is the quality of workplace measurement. Leaders should compare faster production with corrections, rework, escalations, and recipient review time. Those downstream costs reveal whether AI assistance completed work or merely relocated it.

A decline in rework alongside faster delivery would support responsible augmentation. Rising output with stable or worsening correction rates would suggest that organizations are rewarding meat proxies at scale.

The phrase will also face a social test. Used carefully, it can help colleagues set a boundary: bring AI into the process, but do not outsource your responsibility to the recipient. Used carelessly, it can become shorthand for policing style or mocking people.

Readers can apply the useful version immediately. Before sending AI-assisted work, ask whether you understand the answer, whether its important claims have support, and whether it addresses this recipient’s real situation.

Then ask the harder question: what did you add? A useful contribution might be context, a decision, a verified source, a clearer explanation, or an honest statement of uncertainty. It does not need to be elaborate.

If the only contribution was carrying text from one window to another, the work is unfinished. Keep the draft, inspect the evidence, and write the response you are prepared to own.

That is the lasting point behind Simon Willison’s warning. AI can accelerate the production of language, but responsibility cannot be generated and forwarded. The next time a fluent answer appears instantly, pause before making someone else verify it.