MANAGEMENT SIGNAL | HYPE AND REALITY CHECK

A polished proposal can be accurate, useful and properly reviewed, yet be judged more harshly once AI involvement is suspected. Organisations need to separate provenance, quality and appropriate use before stylistic stereotypes become governance.

A recent paid social advert tells a compact horror story. A senior consultant sends a proposal drafted in ChatGPT. The firm learns this only when the client remarks on it. The imagined warning is not that the numbers are wrong, the argument is weak or confidential information has leaked. The embarrassment is that the client thinks it can see the machine.

That is why the advert works. It sells the fear of being found out.

The scene is psychologically astute. In a 2025 study spanning 13 experiments, Oliver Schilke and Martin Reimann found that people trusted actors less when AI use was disclosed. The penalty was stronger when AI use was exposed by a third party rather than volunteered by the user. The advert compresses that transparency dilemma into a single scene: the client knows before management does.¹

Yet the scene also collapses three different questions. Was AI involved? Is the work any good? Was AI used appropriately? They are related, but they are not interchangeable.

A professional writer told me that she used em dashes naturally. After more than one telephone conversation in which someone accused her of using ChatGPT for work she had written laboriously herself, she stopped using them.

She had not changed the authorship of her work. She changed her punctuation to avoid an accusation about authorship.

This is a small but revealing consequence of widespread AI use. The suspicion of AI can make human writers write less like themselves. An authentic habit becomes reputationally dangerous because it has been absorbed into a folk checklist of AI tells.

That checklist grows quickly. It may include em dashes, symmetrical paragraphs, three-part lists, unusually clean grammar, neat headings, a construction such as ‘this is not X, it is Y’, or prose that feels too tidy. Some cues may correlate with particular model habits at a particular time, especially when a model is used with weak prompting and little editing. That still does not make a cue a signature.

Research on human detection gives a mixed and useful picture. A 2023 study in the Proceedings of the National Academy of Sciences found that ordinary readers’ judgements remained close to chance because they relied on a mixture of valid and misleading cues. People sometimes recognised repetition or nonsense, but they also treated characteristics that were more common in human writing as signs of AI.²

A later ACL study found that frequent users of AI writing tools could perform very well on a controlled set of 300 long-form articles. A majority vote among five selected high-performing annotators misclassified only one article. Their explanations, however, drew on clusters of features across whole texts, including vocabulary, structure, originality, clarity, formality and factuality. They were not relying on one magic punctuation mark.³

The difference between those findings is useful rather than inconvenient. Detection depends on the model, the task, the text length, the reader’s experience and whether several judgements are combined. It does not justify workplace accusations based on a familiar stylistic trope.

Automated detection does not remove the judgement problem. OpenAI withdrew its own AI text classifier in 2023 because of its low rate of accuracy.⁴ Turnitin continues to offer AI writing detection, but its current guidance says the model can misidentify human, AI-generated and AI-paraphrased text. It warns that the report should not be the sole basis for adverse action and suppresses scores below 20 per cent because false positives are more likely in that range.⁵

A peer-reviewed Stanford-led study also found that several widely used detectors disproportionately misclassified writing by non-native English speakers. The authors showed that simple changes to word choice could reduce the misclassification, which is another warning that a style pattern can be mistaken for provenance.⁶

These tools may be useful as prompts for review. They do not convert a probability into authorship, and they do not assess whether the work is accurate, useful or appropriately produced.

This distinction matters in organisations as much as it does in education. A false accusation can damage a professional relationship and place the writer in the difficult position of having to prove a negative. The accuser may feel certain because a cue is familiar, even when the cue is weak.

In a 2024 experiment, participants read a GPT-generated science journalism article that was labelled either as AI-written or human-written. The words did not change. The label did. When the text was labelled as AI-written, participants rated its message and source credibility lower.⁷

That is the provenance trap in its purest form. Information about origin changes the evaluation even when the output remains constant.

Sometimes that reaction is justified. Provenance may reveal that a task was delegated inappropriately, that disclosure was required, or that confidential information may have entered an unapproved system. The error comes when provenance becomes a shortcut for quality and professional legitimacy rather than a trigger for the right questions.

The distinction is becoming more important because provenance technology is improving. In August 2026, Anthropic announced that future Claude models will generate text containing a statistical watermark. Its own explanation sets a careful boundary: a detected watermark can indicate that Claude was likely involved at some point, but it cannot distinguish between Claude writing the text and Claude heavily editing it. Anthropic also says the watermark does not determine ownership or authorship.⁸

The same boundary appears in the C2PA provenance standard for digital media. Provenance information can help establish origin, history and whether an asset has been altered, but it cannot by itself tell the reader whether the content is true, accurate or factual.⁹

That is not a weakness in provenance technology. It is the correct limit of the claim. Knowing that a tool participated can trigger the next stage of review. It cannot complete that review.

A competent organisation needs to keep three questions separate, because each produces a different kind of answer and a different management response.

This is a process and provenance question. Which tools were used? At what stage? How much of the wording or analysis did the system contribute? What records exist? A watermark, an audit trail, a tool log or a direct account may help answer it. Punctuation alone cannot.

This is an output question. Is the work accurate, relevant, evidence-backed, original, clear and complete? Does it answer the client’s actual need? Can its claims survive expert challenge? These tests should apply whether every word came from a human, a model or a controlled combination of the two.

This is a governance question. Was the tool permitted? Was confidential information protected? Were contractual and policy requirements met? Was human review proportionate to the risk? Is accountability clear? Was disclosure required?

The same piece of work can score differently across those three questions:

Provenance is therefore relevant, but it is not a master variable that settles the quality and governance questions.

Suppose the client in the advert is right and the consultant used ChatGPT. There may be a serious problem. The firm might have breached a client promise, used an unapproved tool, exposed information, relied on fabricated material or sent work without meaningful review.

But the problem is not the detectable cadence. It is the absent or broken control.

Now suppose the client is wrong. If the firm panics because the proposal contains a stylistic cue associated with ChatGPT, it punishes authentic writing and signals that accusation outranks evidence.

Both outcomes reveal the same organisational weakness. Nobody has agreed what AI use is acceptable, how it should be reviewed, when it should be disclosed, what evidence should be retained and who owns the final judgement.

Governance by embarrassment begins after somebody thinks they have spotted AI. It rewards disguise rather than responsible practice. Staff learn to remove obvious model habits, add deliberate imperfections, use so-called humanisers or suppress their natural style. None of that improves accuracy, confidentiality or accountability.

The right response is neither instant denial nor reflexive confession. Begin with the risk the client is actually raising, then review the work and the process separately.

There is nothing wrong with a publication or firm choosing a house style that avoids em dashes. That is an editorial decision. It becomes a governance problem when the mark is treated as evidence of misconduct.

Good policy should define permitted tools and data, review thresholds, evidence requirements, disclosure triggers, record-keeping and accountable ownership. It should also distinguish low-risk assistance, such as language tidying, from high-risk use in analysis, advice, decisions or client-facing claims.

The NIST AI Risk Management Framework treats trustworthy AI as a combination of qualities, including validity and reliability, safety, security and resilience, accountability and transparency, explainability, privacy and fairness. The framework makes the wider point clearly: trustworthiness is contextual and multi-dimensional.¹⁰

Provenance contributes to accountability and transparency. It cannot stand in for the other dimensions.

The writer who stopped using em dashes solved a reputational problem, but only by giving up part of her natural voice. That is a poor bargain for the individual and a warning for organisations.

It is sensible to edit out formulaic prose because it is dull, repetitive or imprecise. That is quality control. It is very different from removing a legitimate stylistic feature because somebody may treat it as proof of AI use.

The objective is not to make AI-assisted work look human. It is to make all work accountable, defensible and fit for purpose.

A bad proposal does not become good because every word was typed by a person. A good proposal does not become bad because AI helped to produce it. Provenance matters when it changes the risk. It should never be allowed to replace judgement.

The account of a professional writer changing her punctuation after repeated accusations is based on a private conversation reported by the author. It is used as an illustrative vignette, not as general evidence of prevalence.