Roxane Gay AI Writing Advice Exposes a Management Trust Problem

Roxane Gay confronted a timely workplace conflict on September 13: a manager believes nearly every word an employee writes sounds generated by artificial intelligence. The Roxane Gay AI writing advice arrives after generative tools have made polished prose easy to produce and unusually difficult to attribute. The manager hears a synthetic voice, but suspicion alone does not establish who wrote the work.

The dilemma, presented through Gay’s Work Friend column in workplace advice, is larger than one strained reporting relationship. Managers now encounter sentences that feel competent but generic, complete but strangely detached from the work behind them. That feeling can identify a real quality problem. It cannot reliably identify its cause.

The central conflict is therefore not human writing versus machine writing. It is managerial intuition versus verifiable work. Employers still need accurate, original, secure, and useful output. Employees still need clear rules, fair evaluation, and an opportunity to explain their process before a stylistic impression becomes an accusation.

The New York Times AI Writing Question Is Really About Evidence

A manager’s unease can start a review, but it cannot serve as the verdict.

The headline of Gay’s column captures a familiar reaction to AI-assisted prose. Certain expressions now seem suspicious because chatbots repeat them across emails, reports, proposals, and social posts. Smooth transitions, balanced lists, tidy summaries, and impersonal enthusiasm can all trigger the same thought: a person did not write this.

That inference feels stronger than it is. Large language models learned from human language, so their output naturally resembles patterns that people already use. Human writers also absorb conventions from schools, employers, templates, style guides, and one another. Similarity to chatbot prose does not establish chatbot use.

A manager should first separate three questions that are often collapsed into one. Is the work bad? Does it violate an established rule? Is the employee misrepresenting how it was produced? Each question requires different evidence and a different response.

Poor work can be addressed directly. A report might contain unsupported claims, vague recommendations, repeated language, invented citations, or a tone that misses its audience. None of those defects requires the manager to prove AI use. The manager can identify the defect, explain its consequences, and require a revision.

A policy violation requires an actual policy. The organization must define which tools are permitted, what information may enter them, when disclosure is required, and who remains accountable. A manager cannot fairly enforce a private expectation that employees were never told to follow.

Misrepresentation is more serious, but it also needs stronger evidence. An employee who denies using a tool despite reliable logs or a documented admission creates an honesty problem. An employee whose sentences merely sound automated has not supplied comparable proof.

This distinction matters because authorship is no longer binary. One worker might ask a chatbot for an outline, write every paragraph personally, and use a grammar checker afterward. Another might generate a full draft, verify its claims, replace most of its language, and retain two sentences. A third might submit untouched output without reading it.

Calling all three practices “AI writing” hides the very differences a manager needs to evaluate. The relevant issues include judgment, disclosure, data handling, factual review, and ownership of the final result.

Process evidence can clarify those issues. Version history may show how a document developed. Source notes can establish where claims originated. A short conversation can reveal whether the employee understands the argument and can defend the recommendations. Earlier work can provide a baseline without pretending that a person’s style never changes.

None of these signals is perfect. Together, however, they produce a more defensible assessment than a manager’s reaction to prose. That is the first lesson behind the Roxane Gay AI writing advice: treat the output as work to examine, not as a personality test disguised as forensic analysis.

AI Detection Cannot Carry the Weight of a Workplace Accusation

Automated detection turns an uncertain stylistic judgment into a precise-looking score without eliminating the uncertainty.

Managers searching for confirmation will quickly find products that claim to distinguish human writing from generated text. These systems usually analyze statistical features rather than uncovering a hidden record of authorship. They estimate whether a passage resembles patterns associated with model output.

That distinction is critical. An AI detector is not equivalent to a plagiarism database matching copied sentences against a known source. It makes a probabilistic classification. Editing, translation, short passages, technical language, and changing model behavior can all affect the result.

Research has also documented a serious fairness problem. A Stanford-led study found that detectors frequently mislabeled writing by people who were not native English speakers. In the researchers’ test, the systems classified 61.22 percent of examined TOEFL essays as AI-generated.

At least one detector flagged 89 of the 91 essays. All seven detectors agreed on 18 of them. Stanford’s discussion of the detector bias warned that these tools were unreliable, easy to evade, and especially risky in evaluative settings.

That finding should change how employers handle employee AI writing. A polished but linguistically conventional document may attract suspicion because of the writer’s learned English style. A terse technical report may receive a different score than a narrative memo. Neither result establishes misconduct.

The problem becomes more severe when a detector influences performance reviews, promotions, discipline, or termination. At that point, a weak classification is no longer an editing aid. It becomes employment evidence with consequences for someone’s income and reputation.

Federal employment protections still apply when organizations use artificial intelligence in workplace decisions. The Equal Employment Opportunity Commission’s worker guidance states that anti-discrimination laws cover AI systems used in areas including hiring, promotion, and pay decisions. An employer cannot outsource responsibility to software.

Even a detector with better average performance would not answer every relevant question. It could not tell a manager whether an employee used an approved tool. It could not determine whether confidential information was exposed. It could not establish whether the employee checked the facts or contributed the underlying analysis.

A score also cannot measure whether the resulting work is useful. A human can write an inaccurate memo without assistance. A worker can use AI to reorganize a carefully researched draft while preserving full intellectual ownership. The production method affects risk, but it does not replace evaluation of the result.

Managers should therefore treat detector output, if they use it at all, as a weak signal requiring corroboration. It should never be the sole basis for an accusation. Employees should know when such tools are being used, what they measure, and how to challenge an incorrect result.

This is where the dispute becomes a governance issue. Organizations need documented standards for collecting evidence, reviewing contested findings, and preserving employee privacy. The moment a detector becomes part of personnel management, its limitations belong in policy rather than fine print.

The Primary Conflict Is Managerial Intuition Versus Verifiable Work

The strongest response to suspected AI use is a focused work review, not a confrontation about whether prose “reeks” of a machine.

A manager can begin by identifying the observable problem. “This section gives three recommendations without supporting evidence” is useful feedback. “This sounds like ChatGPT” is an allegation that gives the employee little guidance about what must improve.

Specific feedback changes the conversation. It asks the employee to explain sources, assumptions, decisions, and tradeoffs. Those questions test command of the work without forcing either side to litigate the style of every sentence.

Consider a client proposal that contains broad claims but few details. The manager can ask why one option was selected, which customer requirement shaped the recommendation, and what evidence supports the projected outcome. An employee who completed the analysis should be able to answer, regardless of which writing tools assisted the final wording.

The same approach works for internal reports. A manager can request source links, calculation notes, interview records, or a short verbal walkthrough. These materials expose shallow work more reliably than favorite phrases do. They also help identify training gaps that have nothing to do with AI.

Version history is useful when authorship genuinely matters. A sequence of notes, outlines, drafts, and revisions offers evidence of process. A single large paste does not prove prohibited assistance, but it can justify a neutral question. Employees sometimes draft in another application, work offline, or paste material supplied by a colleague.

The conversation should remain proportional. A manager might say that the document differs from prior submissions and ask how it was prepared. The employee can then describe the workflow, including any approved tools. That creates room for correction without assuming dishonesty at the outset.

Clear disclosure rules make this exchange easier. An organization might permit brainstorming and editing while requiring disclosure for generated passages. It might prohibit entering customer data into public models. It might require human verification for legal, financial, medical, or public-facing claims.

Those controls should follow risk rather than personal taste. A routine internal announcement does not present the same stakes as a regulatory filing. A marketing outline does not carry the same confidentiality concerns as a document containing employee medical information.

NIST’s AI risk framework offers a useful structure for this thinking. Its core functions are govern, map, measure, and manage. Applied to workplace writing, that means defining responsibility, mapping use cases, evaluating actual risks, and responding with appropriate controls.

The framework does not reduce governance to a ban. It asks organizations to connect technical use with context, impact, and accountability. A company can permit generative tools while still demanding disclosure, source verification, secure data handling, and meaningful human review.

Managers also need to examine their own contribution. If every assignment is vague, rushed, and repetitive, employees will naturally seek shortcuts. If leadership celebrates AI productivity but punishes any visible sign of AI use, workers receive incompatible instructions.

The New York Times AI writing dilemma surfaces that contradiction. Many organizations encourage experimentation without defining acceptable assistance. Managers then improvise standards after a document makes them uncomfortable. Employees learn that tool use is encouraged in theory but dangerous to acknowledge in practice.

A fair system resolves the contradiction before discipline becomes necessary. The organization names approved tools and restricted data. It defines when disclosure is needed. It clarifies that the employee remains responsible for accuracy, originality, judgment, and compliance.

That structure protects managers too. They no longer need to act as amateur linguistic detectives. They can evaluate work against agreed requirements and escalate only when evidence supports a policy concern.

AI Writing at Work Changes Voice, Responsibility, and Trust

The real cost of indiscriminate AI use is not an awkward phrase; it is the gradual separation of polished language from accountable thought.

Writing performs more than a presentation function inside an organization. A project brief records decisions. A customer email commits the company to a position. A strategy memo exposes what its author noticed, ignored, and prioritized.

When an employee delegates too much of that work, the text can become more fluent while revealing less understanding. The manager may recognize the gap without knowing how to name it. The prose feels generic because it does not contain the organization’s constraints, history, disagreements, or unresolved choices.

That does not make generic style proof of automation. It does explain why managers react strongly to it. They are often detecting a loss of context rather than a specific tool.

The remedy is to require work products that carry their reasoning. A recommendation should state the evidence behind it. A summary should distinguish source facts from interpretation. A plan should identify dependencies, owners, and open questions.

Employees need access to the knowledge required to do that. A fragmented workplace forces people to reconstruct context from chat threads, meetings, documents, and memory. Generative tools can then produce fluent drafts from incomplete inputs, hiding the gaps instead of repairing them.

A maintained personal knowledge base can help a worker preserve source material and retrieve the reasoning behind earlier decisions. The value comes from traceability, not automatic prose. A credible document should lead back to records that a reviewer can inspect.

This is also why an employee’s ability to discuss the work matters. A writer does not need to recite every sentence. The person should understand why the document takes its position, where its information came from, and what uncertainty remains.

Responsibility cannot sit with the tool because the tool does not hold the job. It does not owe duties to the customer, observe internal policy, or face consequences for an invented fact. The employee and the reviewing manager remain accountable.

Recent workplace research suggests that framing an AI system as an “employee” can complicate this responsibility. Boston University summarized research finding that managers overlooked more errors when work was attributed to an AI employee rather than a human or a tool. The researchers connected that framing with weaker perceptions of managerial accountability.

That result does not establish how every team will behave. It does highlight a dangerous linguistic shortcut. Calling software a colleague can make its output feel like someone else’s responsibility, even though no independent professional is standing behind it.

The same accountability gap can appear when managers demand AI use. If leadership requires faster production, employees may assume that leadership has accepted the resulting risks. Managers may assume workers will catch every error. Both sides then believe the other owns verification.

A useful AI writing policy must assign each control to a person. The employee verifies facts and citations. The manager reviews high-impact conclusions. Security teams define permitted data handling. Legal or compliance teams review regulated claims.

The policy should also protect voice without romanticizing inefficiency. Employees should not have to preserve clumsy sentences merely to prove they are human. They should be able to use spelling, grammar, transcription, retrieval, and drafting tools within clear boundaries.

What matters is whether the work still reflects human judgment. The final document should incorporate the author’s knowledge of the audience, operating constraints, and consequences. It should not simply sound polished.

Managers can reinforce that standard through assignments. Ask for a short decision memo with sources and rejected alternatives. Request the assumptions behind a forecast. Include a section for uncertainties or missing information.

These formats make reasoning visible. They also reduce the value of unreviewed generation because a generic answer cannot satisfy the assignment. Employees remain free to use permitted assistance, but they must supply the context and judgment that make the output useful.

Trust grows when this expectation applies consistently. It collapses when a manager suspects one employee while ignoring equally weak work from favored colleagues. It also collapses when workers hide routine assistance because disclosure feels professionally fatal.

The Roxane Gay AI writing advice points toward a difficult middle position. Managers should neither ignore obvious quality changes nor convert discomfort into certainty. They need a process that can examine both the work and its production without treating every unusual sentence as evidence of deception.

What the Suspicion Still Does Not Prove

A writing-style change can support a question, but several plausible explanations remain until the employee answers it.

The employee might be using generative AI extensively. That possibility should not be minimized. If the person submits unverified output, conceals prohibited use, or places confidential information into an unapproved service, the employer has legitimate concerns.

The employee might also be using an approved editing tool. Many applications now offer rewriting, tone adjustment, summarization, and completion features without presenting themselves as standalone chatbots. Workers may not know which functions their employer considers generative AI.

A third possibility is deliberate adaptation. Employees often change their writing after criticism, training, promotion, or exposure to a manager’s preferred format. The resulting prose can become more formal and less distinctive.

A fourth explanation involves language background or accessibility. A worker writing in an additional language may rely on translation or grammar assistance. A disabled employee may use tools as an accommodation. Treating a standardized tone as misconduct can create both fairness and legal concerns.

The manager therefore needs to ask what changed. Was there a new assignment template, writing course, software feature, or feedback cycle? Did the organization recently encourage employees to use AI? Was the worker trying to correct earlier complaints about grammar or tone?

The employee’s answer should then be evaluated against evidence. If tool use occurred, the next questions concern scope and policy. What information entered the system? Which parts were generated? How were claims checked? Did the worker understand the final document?

The manager should avoid demanding a confession to an undefined offense. A productive conversation distinguishes curiosity from accusation. It also gives the employee a chance to provide records, explain the workflow, or identify an unclear policy.

This restraint does not require managerial passivity. If a document contains fabricated citations, the manager can require correction immediately. If protected information entered an unauthorized service, security procedures should begin. If the employee cannot explain critical recommendations, the work may need reassignment or closer review.

The response must match the demonstrated problem. A factual error calls for quality control. Unauthorized data sharing calls for a security response. Dishonesty calls for an employment process. Generic prose calls for editorial feedback.

Conflating those categories produces poor decisions. It also teaches employees to conceal their methods rather than discuss them. A company then loses visibility into how AI is actually being used.

Leaders should be equally careful about overclaiming the benefits of AI writing. Faster first drafts do not guarantee faster completed work. Time saved during generation can return as source checking, rewriting, correction, and managerial review.

The opposite claim also deserves skepticism. AI assistance does not automatically erase expertise or originality. A subject-matter expert can use a model to reorganize material while retaining control of every substantive decision.

The unresolved variable is the workflow. Who supplied the knowledge, who evaluated the result, and who accepted responsibility? Those questions reveal more than stylistic guesses about whether a machine was involved.

Three Signals Will Show Whether Workplace AI Rules Are Maturing

The next stage of workplace AI writing will be measured through disclosed workflows, review quality, and fair dispute procedures.

The first signal is whether employers replace broad prohibitions with task-specific rules. A mature policy will distinguish brainstorming, editing, translation, summarization, generation, and autonomous action. It will also identify restricted information and approved services.

If that change occurs, the central argument here becomes stronger. Employees and managers will have a shared basis for discussing conduct. If employers keep issuing vague instructions to “use AI responsibly,” disputes will continue to depend on personal interpretation.

The second signal is whether companies measure downstream correction work. Leaders often track how quickly a draft appears, but not how long people spend verifying sources, repairing tone, or correcting confident mistakes. That creates an incomplete productivity story.

Teams should compare completed, accepted work rather than raw generation speed. They can examine revision cycles, factual error rates, reviewer time, customer complaints, and rework. These measures reveal whether AI assistance reduces total effort or merely moves it to someone else.

Evidence of lower total review time without declining quality would strengthen the case for managed AI writing. Rising correction costs would weaken claims that broad adoption automatically improves productivity.

The third signal is whether employees receive a fair way to contest AI-related accusations. That process should disclose the evidence, allow a response, and avoid treating detector scores as conclusive. High-impact decisions should involve qualified human review.

A credible process will also examine whether a rule was clear when the work was created. Retroactive enforcement damages trust and encourages silence. Employees should know what records are retained and whether AI usage data affects performance evaluation.

These signals matter beyond one advice column. AI writing has moved from an individual productivity choice into the relationship between workers and managers. It now affects authorship, privacy, evaluation, security, and professional identity.

Organizations should ask a practical question before the next suspicious memo arrives: could a manager explain the company’s standard without referring to how the prose feels? If the answer is no, the policy is not ready.

The Roxane Gay AI writing advice ultimately identifies a management problem that software cannot solve. A manager must evaluate the work, state the concern, hear the employee, and connect any response to evidence. No detector can perform that responsibility on the manager’s behalf.

Start with the document’s concrete weaknesses. Ask the employee to explain the process. Check sources, drafts, and relevant policy. Then decide whether the issue is quality, security, disclosure, or honesty.

That sequence leaves room for both accountability and fairness. It also gives organizations something more valuable than a guess about authorship: a repeatable way to decide whether AI-assisted work deserves trust.