News Analysis

Where AI Chatbots Draw the Line on Imitating Authors

A comparative test of ChatGPT, Claude, Gemini, Copilot and Perplexity found sharply different responses to living-author imitation, deceased authors, broad literary styles and style analysis.

As we explored in our earlier piece on why ChatGPT won’t write in the style of living authors, ChatGPT will often refuse requests to produce new text in a living author’s distinctive voice. The model typically redirects users toward broader craft traits instead: atmosphere, pacing, characterisation, tone or genre conventions.

That boundary reflects a mix of concerns: copyright risk, creator consent, reputational harm, market substitution and the broader question of whether a living writer’s voice should be treated as something AI can freely reproduce.

But is that boundary specific to ChatGPT? Or is it part of a wider industry move to dissuade users from asking models to write in the style of a living author?

To find out, No Latency ran a small comparative audit of five major AI chatbots: ChatGPT, Claude, Gemini, Microsoft Copilot and Perplexity. The goal was not to generate imitative prose, nor to identify workarounds. We wanted to test how different systems draw the line between author imitation, broad literary influence and legitimate style analysis.

The results suggest that AI companies are not applying a single shared rule. They are making different product decisions about where imitation ends and influence begins.

What We Tested

We tested each chatbot across four categories of prompt:

1. Direct imitation of living authors
2. Direct imitation of deceased authors
3. Broad literary-style prompts with no named author
4. High-level style analysis of a living author, with an explicit instruction not to imitate

The test used simple, direct prompts. We did not use jailbreaks, adversarial phrasing or follow-up pressure. We also do not reproduce any generated imitation outputs here. The purpose was to record behaviour categories, not to publish pastiche.

The Results

Our results are summarised in the table below. Note that our table does not tell us which model is “right”. It tells us something more interesting: the models do not behave alike.

Responses were classified using the following codes:

Living Authors Split the Field

The sharpest difference appeared in the living-author tests.

ChatGPT and Perplexity refused direct imitation and redirected the request toward broader craft traits. In effect, they drew a boundary around close or exact imitation of a named living author, while still offering to help with original writing inspired by general features such as atmosphere, characterisation, momentum or tone.

Claude and Copilot occupied a greyer middle ground. They did not refuse the prompts outright, but they added forms of qualification. Copilot tended to translate the named-author cue into broad stylistic traits and then stress originality. Claude produced outputs while showing visible notes about originality, feasibility or stylistic synthesis.

Gemini was the most permissive in this small test. It complied with both living-author prompts without a visible caveat or redirect.

That gives us three broad behaviours:

Behaviour	Systems Observed
Refusal plus safe alternative	ChatGPT, Perplexity
Qualified compliance	Claude, Copilot
Direct compliance	Gemini

This matters because living-author imitation is the most ethically and commercially sensitive category. It raises questions not just about copyright, but about consent, creative identity, reputation and substitution risk.

For a deeper look at the legal question, see our guide to whether AI can copy an author’s style without breaking copyright law.

Can AI Copy an Author’s Style Without Breaking Copyright Law?

ChatGPT Distinguished Living from Deceased Authors

One of the clearest findings was that ChatGPT did not refuse all named-author style requests.

It refused both living-author prompts. But it complied with prompts involving both a long-deceased public-domain author and a recently deceased author.

That suggests ChatGPT’s boundary, in this test, was not “no named literary styles”. It was more specific: no close imitation of distinctive living-author style.

That is an important distinction. Broad literary style, deceased-author pastiche and living-author imitation are not being treated as the same thing.

Perplexity Was Stricter Around “Exact Style”

Perplexity behaved differently.

It refused direct or “exact style” requests not only for living authors but also for the deceased-author prompts. It then offered safer alternatives based on broad stylistic traits.

The model even added a related caveat to one broad-style prompt, despite no author being named.

This suggests that Perplexity’s boundary may be organised less around whether the author is alive and more around the idea of exact stylistic reproduction itself.

That is a stricter rule than ChatGPT’s in some respects, though Perplexity still allowed analysis and broad-style generation.

Broad Style Was Allowed

The most consistent result came from the broad literary-style prompts.

Every system complied with a gothic literary-style prompt that named no author. Every system also complied with a spare noir prompt, though Perplexity added a caveat about not writing in any one author’s exact style.

This is crucial. The stricter systems were not refusing “style” as such. They were drawing a boundary around named-author imitation.

That boundary is easy to blur in practice. A user may think they are asking for a mood, genre or prose texture. A model may interpret the same request as an attempt to imitate a living creative identity.

This is where product design becomes cultural governance. The chatbot has to decide whether a name refers to a protected voice, a general influence, a genre signal, a teaching example or a request for impersonation.

AI companies are not drawing a single line around copyright or style. They are drawing product boundaries around creative identity.

Style Analysis Was Permitted

All five systems allowed high-level analysis of a living author’s prose style when explicitly asked not to imitate it.

That distinction matters. A model can refuse to perform a living author’s style while still helping users understand craft features that are broadly applicable: clarity, pacing, atmosphere, point of view, tension, dialogue, narrative momentum.

For writers and educators, this is the useful line. Studying style is not the same as simulating style. The more restrictive systems appear to recognise that distinction.

Conclusions: The New Boundary is Creative Identity

The test suggests that AI companies are not drawing a single line around copyright or style. They are drawing product boundaries around creative identity.

Those boundaries vary by system.

Some models treat a living author’s name as a red line. Others treat it as a cue to generate something in the general territory, or comply but add caveats about originality. Some avoid “exact style” more broadly. All are more comfortable with unnamed literary modes and style analysis.

This is not just a technical distinction. It affects how AI systems mediate creative work.

The model is no longer just a writing tool. It is also a rule-enforcing layer between the user and the cultural material the user invokes. It decides whether a request is imitation, analysis, homage, genre writing or something too close to a living author’s voice.

That is a significant shift. AI companies are quietly becoming arbiters of what kinds of creative influence are acceptable at the point of generation.

What This Means for Users and Businesses

For individual writers, the safest and most productive pattern is clear: ask for broad craft traits, not author impersonation. A prompt that asks for “slow-building unease, psychologically grounded characters and a sense that something is wrong” is ethically and practically cleaner than asking for a living author’s style.

For businesses, the lesson is slightly different. AI guardrails are not uniform. A workflow that is blocked or redirected in one system may be allowed in another. That creates governance risk, especially for marketing, publishing, entertainment, education and brand work.

If companies are using generative AI for creative production, they cannot assume that platform boundaries will be consistent. They need their own internal rules about imitation, attribution, originality and acceptable influence.

That is because model behaviour is not just a function of capability. It is also a function of product design. AI systems are increasingly being trained not only to answer, but to decide when a request should be refused, redirected or reframed.

OpenAI’s Model Spec is one example of how these trade-offs are formalised, balancing usefulness, safety and alignment. Microsoft has already framed Copilot partly through copyright risk and customer protection, which underlines why these product boundaries matter for enterprise users.

This is part of the broader challenge we cover in Demystifying AI: understanding not just what AI systems can do, but how their design choices shape practical business risk.

Demystifying AI

The Bottom Line

The audit found no single industry standard for literary imitation.

ChatGPT and Perplexity drew relatively clear boundaries around living-author or exact-style imitation. Copilot and Claude operated in a greyer zone of qualified compliance. Gemini was more permissive in the prompts tested.

But across all systems, broad literary style and high-level craft analysis were allowed.

That suggests the key boundary is not “style” versus “no style”. It is named creative identity: whose voice is being invoked, how directly, and whether the model treats that invocation as inspiration, analysis or imitation.

Download the full research note: AI and Literary Imitation: A Comparative Audit of Chatbot Style Boundaries.

---

Method Note

This was a small comparative audit, not a large-scale benchmark. Results may vary by date, model version, account tier, region and product settings. The study did not test jailbreaks or policy-bypass attempts, and it does not reproduce generated imitation outputs.

Testing was conducted in July 2026 using the consumer-facing versions of ChatGPT, Claude, Gemini, Microsoft Copilot and Perplexity available during the testing period. The same seven prompt types were submitted to each system, producing 35 first responses in total.

The prompts were simple, direct and non-adversarial. The audit recorded each system’s initial response only: it did not use jailbreaks, evasive wording, repeated follow-up pressure or attempts to turn a refusal into compliance. Responses were classified according to whether the system complied, qualified or reframed the request, refused and redirected it, or permitted high-level style analysis.

Because consumer AI products can change rapidly — and may vary by model version, interface, account tier, region, configuration or product rollout — the results should be read as a record of observed behaviour during the July 2026 testing period, rather than as permanent claims about any system or company.