AI-Generated Text: Authorship, Detection, and the New Evidential Frontier

Large language models impose their own statistical signature on text, disrupting traditional authorship attribution, while AI detection tools carry documented false-positive rates and no validated court-ready method yet exists for determining whether a specific human used an LLM to write or rewrite content.

Last updated: 10 Jun 2026

For decades, forensic authorship attribution worked from a stable assumption: each writer has a distinctive style that persists across samples, measurable by statistical analysis of vocabulary, syntax, and function-word patterns. That assumption is now under pressure. Large language models can produce fluent, coherent text on demand, write in any requested register, and when used to edit or rewrite a human author's draft, replace the features a stylometrist would rely on.

This creates problems in several directions at once. Academic integrity systems flag student essays. Fraud investigators wonder whether a signed document was actually written by the person who signed it. Defamation cases hinge on authorship. Copyright claims require knowing how much of a creative work is human. And in all of these cases, the tools that law, courts, and institutions want to reach for : AI detection software : are producing false-positive rates that make them dangerous to use as primary evidence.

This topic covers how LLMs generate text and why that process disrupts stylometry, what the detection tools actually measure and what their documented limitations are, the legal questions that turn on human-AI authorship attribution, and what forensic linguistics can honestly say it can and cannot determine today.

Key terms

Large language model (LLM): A neural network trained on large text corpora to predict likely next tokens. At inference, it generates text by sampling from probability distributions over vocabulary, producing output that is statistically smooth and consistent with its training distribution.
Stylometric signature: The set of measurable linguistic features : function-word frequencies, sentence length, punctuation patterns, vocabulary richness : that characterises a specific writer's output and allows their texts to be distinguished from others.
Human-LLM collaboration: Text production in which a human contributes prompts, editing decisions, and intentional content while an LLM generates or transforms the prose. The resulting text carries both human creative intent and the model's statistical signature.
AI detection classifier: A machine-learning system trained to discriminate between human-written and LLM-generated text by measuring features such as perplexity, burstiness, and n-gram probabilities. Current classifiers are not reliable enough for evidential use.
Perplexity: A measure of how surprising a sequence of words is to a language model. LLMs tend to generate low-perplexity text (predictable word choices); human writing often shows higher and more variable perplexity. Detection tools exploit this, but the distinction breaks down with formal academic prose.
Burstiness: The variance in sentence-level complexity across a passage. Human writing tends to mix complex and simple sentences unevenly; LLM output tends to be more uniform. Some detection tools use burstiness as a discriminating feature, though its reliability is contested.

How LLMs generate text and what it looks like

The model does not write the way a person writes : and the difference is measurable, but not always in the right direction.

A large language model generates text token by token, at each step sampling from a probability distribution over the vocabulary conditioned on everything that came before. The model has no communicative intention, no memory of authoring previous documents, and no idiolect of its own. What it has is a statistical pattern learned from billions of words of training text, which it reproduces in a characteristically smooth and predictable way.

The result is text with several measurable properties. It tends to be low in perplexity: the word choices are statistically predictable given the context. It tends to be low in burstiness: the variation in sentence complexity across paragraphs is smaller than in typical human writing. It tends to avoid the kinds of slightly-off collocations and idiosyncratic constructions that human writers produce when their phrasing is slightly non-standard. And it is nearly free of the consistency over time that makes a person's writing recognisable as theirs : each generation is independent.

Perplexity and burstiness: human writing vs LLM output.

The complication is that these properties are not unique to LLMs. A human writer producing formal academic or professional prose makes similar choices: simple sentence structures, cautious vocabulary, avoidance of colloquialisms, consistent tone. A non-native English speaker constructing carefully correct text may produce low-perplexity, low-burstiness prose indistinguishable by a classifier from LLM output. This is where the false-positive problem originates.

Why stylometry fails on LLM-assisted text

When the model edits your prose, it replaces your fingerprint with its own.

Traditional stylometry assumes that a text reflects a single author's stable linguistic habits. Feed 10,000 words by a known author into Burrows's Delta or a similar method, and the software learns to recognise that author's function-word profile, sentence-length distribution, and vocabulary range. Then compare an unknown text, and the distance metric tells you whether it falls within that author's cluster.

This breaks in two distinct ways when an LLM is involved. The first is full generation: if a human provides a detailed prompt and the model generates all the prose, the output will cluster with the model's characteristic distribution, not the human's. The human's creative decisions : the ideas, the structure, the argument : are present, but the linguistic surface, which is what stylometry measures, belongs to the model.

The second failure mode is editing or paraphrasing. A human writes a rough draft, passes it to an LLM with the instruction "polish this", and accepts the output. The ideas and rough structure are human, but the function-word distribution, the sentence boundaries, and the vocabulary choices have been modified by the model's statistics. The resulting text will not cluster reliably with the human's known writing, and it will not cluster cleanly with pure LLM output either. It sits in an ambiguous middle ground that current tools cannot resolve.

Scenario	Human idiolect preserved?	LLM signature present?	Stylometry useful?
Human writes without LLM	Yes	No	Yes, as baseline
LLM generates from detailed prompt	No	Yes	No : attributes to model
Human draft, LLM polishes	Partial	Partial	Unreliable
Human writes, LLM rewrites entirely	Minimal	Yes	No : model dominates
Human writes with light LLM suggestions	Mostly yes	Minor	Partially useful

AI detection tools: how they work and where they fail

The tools courts and universities want to use are not ready for evidential weight.

AI text detection tools : GPTZero, Originality.ai, Turnitin's AI detection feature, and the classifiers built into academic integrity platforms : work by measuring properties of the text that correlate with LLM output in their training data. The most commonly cited are perplexity (computed against a reference language model) and burstiness (variance in per-sentence perplexity). Some tools also measure coherence, structural regularity, and the frequency of specific n-gram patterns associated with LLMs.

The underlying problem is that the classifiers are trained on a snapshot of LLM output at a specific moment. As models improve and as humans learn to prompt them to produce less stereotypically "AI" text, the classifiers' training distribution drifts from the current target. Furthermore, watermarking schemes : where text is generated with a statistical pattern embedded that a detector can verify : exist in research form but are not yet deployed consistently by major model providers in a way that allows third-party verification.

OpenAI released and then withdrew its own AI text classifier in 2023, citing accuracy concerns. The withdrawal itself is informative: a provider with access to its own model's internal probability distributions could not build a reliable public classifier even with that advantage. Third-party tools working only from the surface text face a harder problem still.

AI detection pipeline and its failure modes.

Legal questions: copyright, defamation, and fraud

The law has not caught up with the technology, and forensic linguistics sits in the gap.

Three legal contexts are currently generating the most live questions about AI-assisted authorship: copyright, defamation, and fraud. Each asks a slightly different version of the same underlying question: to what extent is this text attributable to a specific human being?

Copyright: the US Copyright Office's guidance confirms that purely AI-generated material lacks the human authorship required for copyright protection. For human-LLM collaboration, the degree of human creative selection and arrangement is the test. Establishing that degree evidentiary requires some ability to disentangle the human and model contributions : a task current forensic linguistics cannot perform reliably.
Defamation: if a defamatory statement was published with an AI tool producing the offending text based on a human's prompt, questions arise about whether the human publisher holds responsibility for the model's outputs. Several US cases involving AI-generated fake news stories and fake legal citations (the Mata v. Avianca case, 2023) have established that submitting LLM-generated content without verification is sanctionable, though the defamation liability question is less settled.
Fraud: where a document is purported to have been authored by a specific person : a signed report, a witness statement, a professional assessment : and AI generation is suspected, the authenticity question requires showing that the document's linguistic surface is inconsistent with the named author's established style. This negative finding is within forensic linguistics' current capability; the positive attribution of AI generation is not.

What forensic linguistics can and cannot determine today

Honesty about the limits of the field is its most important current contribution.

The current state of the field can be summarised plainly. A forensic linguist examining a questioned text can offer the following:

Negative authorship findings: the text lacks the function-word profile, sentence-length distribution, and vocabulary choices that characterise the named author's established writing. This is a well-grounded negative claim that stylometry supports.
Consistency-with-LLM observations: the text has statistical properties : low perplexity, low burstiness, absence of idiolectal markers : consistent with LLM generation. This is an observation, not a proof. The same properties can arise from careful human formal prose.
Model-specific attribution: no court-accepted method exists for attributing a specific text to a specific LLM version. Watermarking could provide this, but it is not yet deployed in a court-ready form.
Human-LLM collaboration proportion: no reliable method exists for estimating what percentage of a text was generated by an LLM versus written by a human. This question is currently unanswerable by any published method.

Professional bodies including the International Association of Forensic Linguists and the UK's Forensic Science Regulator have not yet produced specific guidance on AI-text evidence, reflecting how recently the issue has become practically urgent. The absence of a validated, peer-reviewed, court-ready methodology is not a gap that expert testimony can currently bridge by assertion.

The field's trajectory and emerging approaches

The tools are catching up, slowly, with the problem they need to solve.

Several lines of work may change the picture. Cryptographic watermarking, where tokens are generated with an embedded statistical pattern that allows the model provider to verify later whether a text came from their system, is technically viable and has been prototyped by researchers at Google and elsewhere. If model providers adopt standardised watermarking, it could eventually provide a reliable mechanism for court-admissible AI attribution. The adoption problem is commercial: providers may be reluctant to enable easy identification of their outputs.

Authorship verification : rather than attribution : is a more tractable short-term problem. Rather than asking which of many possible authors wrote a text, verification asks: is the distribution of features in this text consistent with the known writing of person X, yes or no? Applied to the negative case (it is not consistent), this can reliably flag the possibility of AI assistance without claiming certainty about the mechanism. This framing is more defensible in court than a positive AI-generation finding.

The deeper question the field has not yet resolved is whether human-LLM collaborative writing, as it becomes a normal part of professional document production, remains a forensically tractable question at all. If everyone uses LLMs to polish their prose, the population of comparison texts will itself be LLM-influenced, and the baseline shifts. At that point, the question "did this person use an LLM?" may become as unanswerable as "did this person use a spell-checker?"

Worked example

A fraud investigation: is the signed report genuine?

A professional report is disputed. The named expert says they did not write it.

A dispute arises in civil litigation over a forensic accounting report tendered in evidence. The named expert testifies that the report is not their work and that they did not write or authorise it. The opposing party claims the report was produced by the named expert with AI assistance and submitted as entirely their own work. The question for the forensic linguist is whether language evidence can settle who wrote the report.

Building the known corpus. The analyst collects 15 reports, two conference papers, and 40 emails written by the named expert over the preceding five years, totalling approximately 80,000 words. This gives a statistically reliable baseline.
Stylometric comparison. Burrows's Delta and a support vector machine classifier are applied to compare the questioned report against the known corpus. The report falls outside the expert's cluster on multiple measures: function-word frequency profile, average sentence length, and punctuation rate all diverge from the baseline at a statistically significant level.
LLM-consistency analysis. The analyst measures perplexity and burstiness of the questioned report using a reference language model. Both measures fall in the range associated with LLM output in published research. The analyst notes this is consistent with but does not prove AI generation.
Idiolect markers. The named expert's known writing contains characteristic features: a preference for numbered lists in narrative sections, frequent use of the phrase "it is submitted that" as a hedge, and consistent use of em-dashes in parenthetical asides. None of these features appear in the questioned report.
Expert opinion. The analyst reports that the questioned report is not consistent with the named expert's established writing style on multiple independent measures, and that its statistical properties are consistent with text generated or substantially rewritten by a large language model. The report is framed as moderate-to-strong support for the view that the named expert did not author the document. No claim is made that the specific LLM used can be identified.

The negative finding : this does not look like the expert's writing : is the most defensible part of the opinion. The positive consistency-with-LLM finding is expressed as supporting but not conclusive. Together they establish that the attribution of the report to the named expert is linguistically unsupported, which is the relevant question for the court.

Check your understanding

Question 1 of 4· 0 answered

Why does LLM-assisted editing of a human draft make stylometric attribution unreliable?

Key Takeaways

LLMs generate text by sampling from learned statistical distributions, producing output with characteristic low perplexity and low burstiness that is measurably different from human writing : but that difference is not reliable enough to serve as proof in court.
When a human uses an LLM to write or rewrite text, the model's statistical signature replaces the human's idiolect, making traditional stylometric authorship attribution unreliable in the collaborative or edited case.
AI detection tools carry documented false-positive rates of 4-20% or more and disproportionately flag non-native English writers; they should trigger further investigation, not serve as primary disciplinary evidence.
Forensic linguistics can currently support negative authorship findings (this text is inconsistent with author X's known writing) and consistency-with-LLM observations, but cannot attribute a text to a specific model or quantify human versus AI contribution.
Cryptographic watermarking offers a potential future path to court-ready AI attribution, but has not been deployed consistently by model providers in a verifiable form as of the field's current state.

Can forensic linguists tell whether a human used an AI to write a piece of text?

Not reliably, and not to a standard currently accepted by courts. AI detection tools measure statistical properties of text but produce significant false-positive rates. A forensic linguist can note features consistent with LLM output but cannot at present attribute a text to human-LLM collaboration with the precision required for expert court evidence.

Why do large language models disrupt traditional stylometric authorship attribution?

Stylometry works by measuring an individual writer's stable linguistic habits. When a human writes through an LLM, even using it only for editing or rewording, the model replaces the human's idiolectal features with its own statistical patterns. The output reflects the model's training data rather than the human author's style, making the human's fingerprint invisible.

What false-positive rates have been documented for AI text detection tools?

Studies have found that tools like GPTZero and Turnitin's AI detector incorrectly flag human-written text as AI-generated at rates ranging from roughly 4% to over 20% depending on the corpus tested. Non-native English writers have been shown to be flagged at disproportionately high rates because formal, simple prose resembles LLM output statistically.

Can AI-assisted text be attributed to a human author for copyright purposes?

This is an unresolved legal question across most jurisdictions. US copyright doctrine requires human authorship; the Copyright Office has indicated that purely AI-generated material is not protectable, but human-AI collaboration may be depending on the degree of human creative input. The evidential question of how much a specific human contributed is precisely what forensic linguistics cannot yet reliably answer.

What can forensic linguistics currently establish about AI-generated text in legal proceedings?

An expert can testify that a text lacks the usual idiolectal features of a named author's established writing style : a negative finding. They can note statistical properties consistent with LLM output. They cannot affirmatively identify which LLM was used, that a specific person used an LLM, or that a document is definitively AI-generated rather than written by a person with an unusually formal, consistent style.

Test yourself on Forensic Linguistics with free, timed mocks.

Practice Forensic Linguistics questions

Found this useful? Pass it along.

Spotted an error in this page? Report a correction or read our editorial standards.

AI-Generated Text: Authorship, Detection, and the New Evidential Frontier

Last updated: 10 Jun 2026

Scenario

Human idiolect preserved?

LLM signature present?

Stylometry useful?

Human writes without LLM

Yes

Yes, as baseline

LLM generates from detailed prompt

Yes

No : attributes to model

Human draft, LLM polishes

Partial

Unreliable

Human writes, LLM rewrites entirely

Minimal

Yes

No : model dominates

Human writes with light LLM suggestions

Mostly yes

Minor

Partially useful

Key Takeaways

LLMs generate text by sampling from learned statistical distributions, producing output with characteristic low perplexity and low burstiness that is measurably different from human writing : but that difference is not reliable enough to serve as proof in court.

When a human uses an LLM to write or rewrite text, the model's statistical signature replaces the human's idiolect, making traditional stylometric authorship attribution unreliable in the collaborative or edited case.

AI detection tools carry documented false-positive rates of 4-20% or more and disproportionately flag non-native English writers; they should trigger further investigation, not serve as primary disciplinary evidence.

Forensic linguistics can currently support negative authorship findings (this text is inconsistent with author X's known writing) and consistency-with-LLM observations, but cannot attribute a text to a specific model or quantify human versus AI contribution.

Cryptographic watermarking offers a potential future path to court-ready AI attribution, but has not been deployed consistently by model providers in a verifiable form as of the field's current state.

Can forensic linguists tell whether a human used an AI to write a piece of text?

Why do large language models disrupt traditional stylometric authorship attribution?

What false-positive rates have been documented for AI text detection tools?

Can AI-assisted text be attributed to a human author for copyright purposes?

What can forensic linguistics currently establish about AI-generated text in legal proceedings?

Your journey to becoming a forensic professional starts here.

Key Takeaways

Key Takeaways