How to detect artificial intelligence in text: methods and limits

You run marketing, SEO or a B2B editorial line, and you need to understand what a detector measures before you trust its score. This page explains how to detect text generated by artificial intelligence: the methods (statistics, classifiers, stylometry, watermarking), their limits, and the most common case in production, the hybrid human + AI text.

You will know which signal to believe, which to set aside, and which checks to trigger to protect your visibility in Google and in generative engines.

What It Means to "Detect" AI Content: Scope, Risks, and Confidence Levels

Detecting AI-generated content is not the same as proving its origin: it is about establishing a confidence level based on clues such as overly regular style, lack of evidence, inconsistencies and production history. The right scope is reducing risk, not chasing infallibility.

A detection score is a signal, not proof: what it establishes and what gets decided next are set out in our guide to AI content detection.

Detection, Attribution, Traceability: Do Not Mix Up the Terms

These three concepts sound similar, but they support different decisions:

• Detection: a probabilistic estimate of "AI or human", or "highly automated or lightly automated".
• Attribution: an attempt to identify a specific model, family or technique, often fragile and easy to bypass.
• Traceability: your internal ability to show who did what, when, with which sources and which validations.

Only traceability produces verifiable evidence: it enables audits and compliance, even when the text itself is indistinguishable.

False Positives and False Negatives: Why a "Score" Is Not Enough

A detector score is a model output, not a truth. It depends on length, technical depth, language, level of editing and the detector's training data. B2B content, with its standardised style, is particularly exposed. The table pairs each error with a mitigation.

The two errors do not carry the same cost: a false positive costs review time and human content reworked for nothing, a false negative costs compliance and trust. The stakes of the page decide which one weighs more.

GEO and SEO Impact: Google Visibility, Quotability, and Trust in Generative Engines

Across Google and generative engines, trust becomes a filter: a text with no evidence and no sources is harder to cite. So useful detection is about spotting what is missing to be credible, not a "robotic" tone.

Scientific Detection Methods to Spot AI-Generated Text

Serious methods do not try to read an author's mind: they measure distributions, learn decision boundaries or look for voluntary marking. The table compares them.

Statistical Approaches: Perplexity, Burstiness, and Distribution Signatures

Some generated texts used to show lower perplexity and more regular burstiness than human writing, but models have improved and editing flattens these differences. The validity conditions are narrow:

• they work best on long, unedited texts;
• they degrade sharply on short texts, translations or rewrites;
• they must be calibrated by domain, because technical B2B content does not read like a consumer blog.

One line of research measures the geometry of a text rather than its vocabulary. The study "Intrinsic Dimension Estimation for Robust Detection of AI-Generated Texts" (arXiv:2306.04723) reports an average "intrinsic dimension" of around 9 for fluent human texts in several alphabetic languages, around 7 for Chinese, and roughly 1.5 times lower for generated texts. These are methodological reference points, not thresholds to apply to your pages.

Classification Approaches: Supervised Training and Multi-Model Generalisation

Classifiers learn to separate human and AI examples. Their core limitation is generalisation: their performance drops when faced with new models or assisted texts. For their score to be readable, three properties matter:

• Calibration: an 80% output should reflect an 80% probability; otherwise the score misleads.
• Explainability: the tool shows the signals, the sections of text concerned and why uncertainty exists.
• Uncertainty handling: the tool can say "I don't know" in ambiguous cases.

Without these properties, you get a score… and a poor decision.

Watermarking Approaches: Principle, Promise, and Operational Limits

Watermarking marks an AI output with a statistical pattern, which makes detection independent of style. Still, the model must enable it, the mark must survive paraphrasing, and the ecosystem must adopt it. In SEO, where texts are rewritten and localised, its persistence remains a hard constraint.

Hybrid Approaches: Combining Linguistic Signals, Metadata, and Usage Context

Hybrid approaches hold up best in production because they ask "what risks exist before publishing?" rather than "human or AI?". They combine three families of signals:

• Text signals: regularity, redundancies, cautious phrasing, generic statements.
• Metadata: version history, production time, validation chain.
• GEO and SEO context: intent, sources, evidence, E-E-A-T, quotable elements.

Stylometry has its place here: it spots a style break inside a document, which a global score does not show.

Why AI Detection Remains Difficult: Technical and Human Challenges

Real-world content is almost never raw AI: assisted, edited, translated, it turns detection into a governance problem as much as a technical one.

Paraphrasing, Translation, Rewriting: How Editing Erases Signals

Paraphrasing often makes detection collapse, and a human text translated automatically can pass for AI. A detector alone therefore mostly measures the level of editing, not the origin.

Model Drift and the Bypass Race: The Problem Moves

Once a signal becomes known, bypass strategies emerge (artificial variability, noise, post-editing): producing credible text costs less than maintaining a universal detector.

So can you make AI undetectable? Some techniques lower a score, but that changes nothing about the decision, which rests on usefulness, sources and editorial responsibility; what you risk in SEO when you paraphrase with AI is covered separately.

The Hybrid Human + AI Text: Why the Score Gets Blurred

Hybrid text is the normal case in production: a generated draft then reviewed, a human text reworded, a touched-up translation, AI aligned with your tone of voice. Yet detectors have learnt to separate two pure classes, and a mix belongs to neither.

What Happens to the Score on a Mixed Text

Several mechanisms blur the signal at the same time:

• Averaging: a global score smooths out very different passages and does not say which ones raise questions.
• Oscillation: a generated draft that is then rewritten makes the score vary from one passage to the next.
• Erasure: human review neutralises statistical markers without changing where the ideas came from.
• Outside training: mixed texts are the ones on which classifiers generalise worst.

Short, technical or brand-aligned texts add ambiguity: beyond a certain level, detecting turns into auditing. The table summarises the common cases.

Reading a Score on a Hybrid Text

On a hybrid text, the question "AI or human?" often has no answer. The useful questions concern how the work was shared and the substance:

• Which passages does the detector highlight, and what do they contain (generic statements, repetitions, unsourced claims)?
• Who produced each part, with which brief and which sources?
• Are the passages where the style breaks accurate and useful for the page's intent?
• Was the published version approved by an identified person?

A documented hybrid text can be defended. The protocol applied to a text, from framing to the stability test between two runs, is described in our guide to check whether a text was generated by AI.

Building a Robust Detection Workflow (B2B-Focused)

A robust workflow does not monitor writers: it automates triage and keeps humans for arbitration. Before you try to detect anything, clarify what you want to avoid:

• Compliance: regulated sectors, mandatory notices, claims that require substantiation.
• Quality: accuracy, clarity, depth, non-redundancy, freshness.
• Reputation: visible mistakes, hallucinations, vague promises.

No numerical threshold works for every page: whether to accept, revise, escalate or document is decided according to the stakes of the page.

Sampling, Double Review, and Traceability: How to Improve Reliability Without Slowing Production

At high volume, you cannot review everything. Three levers make checks more reliable without blocking the pipeline:

• Risk-based sampling: high-stakes commercial pages, YMYL pages, conversion content, pages already performing.
• Double review: on content that triggers a signal (detector output, complaints, SEO drop, inconsistencies).
• Traceability: versioning, status (draft, approved) and a change log.

Pair Detection With GEO/SEO QA: Usefulness, Evidence, Sources, Coherence, and Intent Alignment

Google judges usefulness and compliance with its spam policies, not the production tool, as stated by Google Search Liaison (Danny Sullivan, 12 January 2023). So turn detection into a quality-control checklist:

• clear intent: what, for whom, and at what level of expertise;
• evidence and primary or secondary sources: links, standards, studies, auditable internal data;
• concrete examples, stated limitations, validity conditions;
• a quotable structure (definitions, steps, tables), useful for generative engines.

AI Detector: How to Evaluate Detection Tools Without Getting It Wrong

A detector becomes dangerous when used as a judge: only its performance on your own content counts. Selection criteria and the hidden cost of false positives are brought together in our method for evaluating an AI detector.

Internal Testing: Build a Realistic Evaluation Set (Human, AI, Mixed, Edited)

Without an internal test set, you do not know how your detector performs on your content; the criteria and metrics for testing an AI in production apply to the detector itself. Build a corpus with four families:

• fully human texts, written by different authors;
• fully generated texts, as raw outputs;
• mixed texts, generated then reviewed by a human;
• texts edited by paraphrasing, translation, rewriting or factual enrichment.

Measure errors by content type; the mixed and edited families reproduce the hybrid text you actually publish.

Business Constraints: Confidentiality, Data Retention, and Auditability

In B2B, check the confidentiality of sensitive content, data retention (logs, reuse for training), auditability and integration with your workflows: a tool that gets bypassed protects nothing.

Practical Cases: AI Signals in SEO Content (and in GEO Answers)

Useful signals are about structure and evidential strength more than style.

SEO Content: Over-Regular Structure, Redundancy, Vague Claims, and Lack of Evidence

These signals trigger a check, without proving the origin of the text:

• plans that are too symmetrical, mechanical transitions, repeated definitions;
• vague promises ("significantly improves", "optimises easily") with no criteria;
• few data points, few examples, no limitations or conditions;
• unsourced claims on sensitive topics.

To prioritise, start with pages that stagnate outside the first page of results: the blockage there is more often about usefulness than about "AI or human".

GEO Content: What Makes Text "Quotable" (and What Triggers Distrust)

Generative engines favour passages that are easy to extract and justify:

• What helps: crisp definitions, numbered steps, tables, identifiable sources, careful phrasing around uncertainty.
• What raises concern: absolute statements with no evidence, internal contradictions.
• What is often missing: anchoring, meaning data, links and context.

Being cited as a source in an AI Overview can increase CTR by +1.08% (Semrush, 2025): quotability has a measurable value, which our SEO statistics put into context. Its levers are detailed in our guide to generative engine optimization.

Action Plan: Turn Suspicious Content Into Publish-Ready Content Without Losing Performance

Moving from suspicious content to publish-ready content follows five steps:

• Reduce vagueness: replace generalities with criteria, conditions and limits.
• Add evidence: external sources, auditable internal data, contextualised examples.
• Strengthen structure: lists, tables, definitions, steps.
• Verify facts: figures, dates, standards, quotes, overall coherence.
• Measure: track the effect on impressions, clicks and engagement in Google Search Console and Google Analytics.

Depending on Your Situation

Depending on the question you are asking, these articles take over:

• You want to read a text's signals beyond the score: identify the ones that matter in SEO with our guide to AI text analysis.
• A corrector or rephrasing tool has been run on a human text: check what it changes with our method to evaluate an AI corrector.
• You want to judge the quality of a generated text, not its origin: rely on the criteria for the quality of AI-generated text.
• You fear that assisted content reuses sources without attribution: the decision grid is in our article on plagiarism in the age of AI.

To produce controlled hybrid texts from the outset, true to your tone of voice and grounded in your own data, Incremys offers a personalised AI that frames generation upstream rather than checking it after the fact.

FAQ: How to Detect Artificial Intelligence, Methods, Tools, and Limits

How do you detect artificial intelligence in a text?

You combine a detector, used for triage, with a quality-control framework, used for decisions. The review looks for missing sources, vague claims, inconsistencies and overly regular structures. For sensitive content, you add factual verification and traceability of versions and approvals.

What are the most reliable detection methods in 2026?

Hybrid methods, which add metadata and usage context to text signals, are the most robust in production. Statistics and classifiers become fragile as soon as a text is edited or translated, and watermarking often fails to survive rewriting. Reliability comes above all from calibration tested on your own content.

Why is detecting AI difficult?

Detection is difficult because most real-world content is hybrid, and editing removes the signals. Short and technical texts multiply false positives. An AI aligned with a brand voice can become indistinguishable on reading, which shifts the question away from origin and towards traceability.

What is the role of an AI detector within an organisation?

An AI detector is a triage and alert tool, not a judge. It helps you prioritise human review and trigger source checks on high-stakes content. Properly integrated into your workflow, it accelerates quality control; used alone, it produces decision errors.

Is a "humanised" text still detectable?

A humanised text is sometimes detectable, sometimes not, because editing dramatically reduces detector reliability. It can fool a tool without becoming better on substance. It is better to base the decision on verifiable criteria, such as sources, accuracy, examples and coherence.

How can you reduce false positives when analysing expert (B2B) content?

You reduce false positives by segmenting by content type and never concluding on a short or highly standardised text without additional checks. A double review compares decisions against evidence criteria. Finally, you calibrate tools on an internal corpus of human, AI, mixed and edited texts.

Does AI detection actually help SEO and visibility in generative engines (GEO)?

Yes, if it helps you improve publishable quality: evidence, usefulness, quotable structure and coherence with intent. No, if it only classifies origin without improving substance. Rankings depend on the value delivered to users, and quotability on the trust that sources provide.

What evidence and sources should you require before approving content?

Before approval, every figure, sensitive claim or comparison must point to a verifiable, recognised source. You also ask for definitions, limits of application and contextualised examples, then keep a record of the references. Without evidence, content is fragile, whether it is human-written or generated.