How to check for AI text: a reliable method and its limits

A tool has just flagged a text as "probably AI", and you run SEO, marketing or editorial for a B2B company. Before you publish, revise or reject it, you need to know how to check whether text was generated by AI without over-interpreting a score.

This page sets out the five-step protocol, the decision grid, what to do when a human text is wrongly flagged, and how to document provenance. The overall framework is covered in our guide to AI content detection.

How to check whether text was written by AI: a reliable method and how to interpret results

What checking a text lets you decide

Checking a text is about making a decision: publish, revise, request sources or escalate to an expert. In B2B, that decision affects compliance and brand credibility, so it must rest on evidence rather than on style.

What a check can confirm (and what it should not conclude)

A detection score is a signal, not proof. In practice, a check can give you three types of information:

• A probability: the likelihood that the analysed sample was generated.
• Atypical segments: the passages to review first.
• Form risks: repetition, stock phrasing, inconsistencies.

It proves no intent, because detectors do not have access to the writing history. Human text can "look like AI" (standardised style, translation) and generated text can "look human" after rewriting. How detectors work and what hybrid text is are explained in our article on how to detect text generated by artificial intelligence.

Clarify the goal of the check: compliance, quality, GEO and SEO

Before opening a detector, set the objective. The same text may be acceptable for marketing, risky for regulated communications, and counterproductive for SEO if it is not useful.

B2B use cases

Acquisition pages, white papers, help centres and multi-country content combine volume with rigour. Regulated content (finance, healthcare) calls for stricter checks than a blog post, and what you check changes with the content type:

• Marketing: accuracy, differentiation, the offer, and supporting evidence.
• Documentation: correctness, prerequisites, and reproducibility.
• Regulated content: compliance, mandatory statements, and sources.
• Multi-author content: tone alignment, cross-page consistency, and traceability.

SEO vs GEO: why being "detectable" is not the real issue, and what you should actually secure

With generative answers, a page must also be able to be summarised or cited; making it "undetectable" is not a goal. Google judges usefulness and compliance with its spam policies, not the production tool, as stated by Google Search Liaison (Danny Sullivan, 12 January 2023).

What you need to secure, regardless of how the text was produced:

• Usefulness: a real answer to the reader's question.
• Accuracy: verifiable, dated facts.
• Structure: headings, lists and direct answers.
• Evidence: named sources for every key claim.
• Quotability: passages a generative engine can reuse without misinterpretation.

What makes a page get picked up is detailed in our guide to Generative Engine Optimization (GEO).

Reliability, limitations and risks: reading results without over-interpreting

What reduces detection reliability

Several situations distort the signal, in either direction:

• Short texts: too little statistical material, so the score is unstable.
• Standardised texts: procedures, terms and conditions and technical sheets naturally look like "regular" output.
• Jargon and acronyms: training corpora often under-represent your industry.
• Quotes, tables and lists: they interrupt stylistic continuity without indicating origin.
• Machine translation: it standardises the style of a text that was written by a human.
• Stitched-together sources: tone breaks and inconsistencies mislead the tool.

Reliability therefore depends more on your context than on the tool: the remedy is to segment, analyse by block, and compare against human texts in the same format.

Reading a tool's outputs: overall score, segments, highlighting and explanations

The overall score is a warning signal, not a verdict. The real value lies in the highlighted segments, which focus human review. Each output calls for a different reading:

False positive rate: what it means

A false positive flags a human text as "probably AI"; a false negative lets generated text pass. The false positive rate, the share of human texts wrongly classified, can only be estimated with a corpus of texts of known origin in your own formats and languages. The stricter your editorial guidelines, the higher this risk.

Neither error is the more serious in absolute terms: a false positive costs review time and human content reworked for nothing, while a false negative costs compliance and trust. The stakes of the page tell you which weighs more.

Implement evidence + context validation rather than automated verdicts

Validate with two inputs: evidence and context. Each dimension raises a question to settle:

An operational step-by-step process to check text

Formalising the process makes checking repeatable and reduces arbitrary decisions. The protocol follows five steps.

Step 1: define the scope (goal, audience, risk level, legal constraints)

A scoping sheet prevents you applying the same check to an internal email and to a page that represents the brand. It covers four fields:

• Content objective: conversion, information or compliance.
• Audience: prospects, customers, regulator or internal teams.
• Page stakes: standard, sensitive or critical.
• Legal constraints: required statements and validations.

It is the stakes of the page, not the score, that set the depth of the check:

Step 2: prepare the text (source version, quotes, references, reused sections)

An analysis that can be reproduced in an internal audit rests on three actions:

• Lock a source version: otherwise, you end up comparing variants and lose traceability.
• Exclude or tag reused passages: quotes, legal excerpts, standard definitions and reused blocks skew scores.
• Keep the original of a translated text: along with the name of the translation tool used.

Step 3: cross-check signals (detection, plagiarism, sources, factual coherence)

Cross-check at least two detectors, then add checks that do not depend on any model:

• Detection: two tools on the same version, overall score and risk areas.
• Similarity: an anti-plagiarism check that fits your constraints.
• Sources: links, dates, internal documents and quotations verified.
• Coherence: figures, definitions and no internal contradictions.

Add a stability test: analyse the text on day zero, then again after a minor edit, and compare the flagged segments. If the same areas come up, the signal is usable. It is this test, not a minimum length, that tells you whether the sample is long enough.

Step 4: decide (accept, revise, request writing evidence, rework)

The decision grid crosses the observed signal with the stakes of the page. It contains no numerical threshold, because the same score does not call for the same response on a blog post and on a regulated page:

A human text that is wrongly flagged follows the same grid, with a route of appeal described below.

Step 5: document (traceability, versioning, validation criteria, internal audit)

A short but systematic record protects the team if decisions are challenged and reveals which formats trigger the most false positives. For each check, record:

• The content version (URL or internal ID) and the date of analysis.
• The tools used and the summary result (score, main areas).
• The decision taken according to the grid and the name of the person who approved it.
• The evidence added (sources, figures, quotations) and where it sits on the page.

Documenting a text's provenance: versions, sources and responsibilities

The more personalised the AI, the harder it is to tell where a text came from: verification shifts towards coherence and evidence. Every published piece should be able to explain what it claims, which sources it relies on, and which intent it serves. A provenance file keeps:

• Versions: source version, intermediate versions and the date of each change.
• Contributors: writers, assistance tools used and reviewers.
• Sources: links, internal documents and data cited, with their date.
• Validations: subject-matter, legal or editorial sign-off.
• Control reports: detection, similarity and the recorded decision.

This file makes audits easier and keeps teams aligned. A detection report sits in it as an indication, never as proof of authorship: version history, briefs and human validations describe the actual writing process.

What to do if a human text is flagged as AI-generated

Do not conclude from the score. A human text that is wrongly flagged is an ordinary false positive, handled through a procedure everyone knows, not through suspicion.

Re-examine the text with the protocol

Re-run the protocol on the flagged version and compare it with other texts by the same author and in the same format. A signal that is unstable, or that also appears on reference human texts, supports no decision. The author can request this re-examination and attach their own evidence.

The evidence to provide

Proof of origin comes from the writing process, not from one more detector:

• Version history and edit logs.
• The initial brief, notes and sources consulted.
• Exchanges and approvals with reviewers or experts.
• The factual check of the text (sources, dates, internal coherence).

A documented decision

Decide using the decision grid: solid, well-supported content is accepted as is, with no rewriting aimed at lowering the score. Record the incident (score, evidence, decision, approver) and, if it recurs, adjust your calibration.

Reducing false positives: editorial best practice and control protocols

False positives damage trust and slow teams down. You reduce them through method more than by switching tools.

Make diagnosis more robust

Do not judge an isolated excerpt. If the score shifts dramatically between the "before edit" and "after edit" versions, that often signals style bias rather than certain origin. Three practices limit gut-feel decisions:

• Segment by sections (H2, H3) rather than analysing the whole document as one block.
• Exclude quotes, standard excerpts and bibliographies.
• Compare against a human text of the same template, which serves as an internal baseline.

Avoid "high-risk" styles: overly rigid templates, artificial synonym swapping and mechanical paraphrasing

A rigid template repeated across series of pages resembles automated production. Automatic paraphrasing sometimes lowers certain signals but makes the text less useful, and Google lists it among spam examples when it disregards quality. Lowering a score changes nothing about the decision, as our article on AI text rewriting shows.

Factual checking: sources and coherence

Factual checking does not depend on any model and also strengthens GEO, because generative engines prefer well-sourced passages. Check:

• The presence of primary sources (documents, studies, official pages).
• The accuracy of dates, units, scope and definitions.
• The internal coherence of the document (same figure, same method, same assumption).

Improving AI-assisted text without "masking" it: quality, evidence and quotability (SEO + GEO)

Improving AI-assisted text means increasing human-added value, which serves both usefulness for SEO and reuse for GEO.

Increase usefulness and reduce generic phrasing

A generic model rarely produces trade-offs and realistic examples without a brief and data. Add them, and replace boilerplate wording:

• Specify the context (industry, risk level, audience).
• Add actionable steps (checklists, matrices).
• Include a counter-example (what not to do, and why).
• Introduce evidence (sources, definitions, scope).
• Replace "revolutionary solution" with the precise context in which the solution applies.
• Replace "guaranteed results" with the condition that produces the result, measured by a KPI.

Structure for extractability: definitions, direct answers, tables, FAQs and sources

Extractable content stays useful even without a click. Three structural habits:

• Open some sections with a one- or two-sentence direct answer.
• Add decision tables (risks, actions, owners).
• End with an FAQ that covers real objections.

Checking only makes sense if it improves performance or reduces a risk. Tracking in Search Console and Analytics, like choosing a tool, is covered in our article on measuring the reliability of an AI detector.

Depending on your situation

Depending on what the check revealed, these articles take over:

• You suspect sources were reused without attribution: the situation, risk and decision grid is in our article on plagiarism in the age of AI.
• You need to interpret a similarity report: our guide to anti-plagiarism software explains how to read it before deciding.
• You are checking dozens of pages at once: the method for scaling up to a whole site is described in our article on the website checker.

If you would rather frame generation upstream than check after the fact, Incremys offers a personalised AI built on your brand and your own data, with tracked validations and versions.

FAQ: how to check whether text is AI-generated

How do you check whether text is AI-generated?

Check in two stages: a signal, then evidence. Analyse the text with two detectors and test whether the result is stable between two runs. Then confirm through sources, factual coherence, similarity and traceability, and decide according to the stakes of the page.

How reliable is checking whether text is AI-generated?

The reliability of a check varies. It drops for short, standardised, translated or quotation-heavy texts, and domain jargon skews it. Treat the score as a triage indicator, confirmed by a stability test and factual checks.

Why do false positives happen when checking whether text is AI-generated?

False positives come from human writing styles that statistically resemble generated text: strict editorial guidelines, repetitive templates, translation, technical documentation. To reduce them, segment the text, exclude quotations, apply the stability test and compare against human texts in the same format.

What score threshold is genuinely usable?

No universal threshold is usable. A score only becomes meaningful once calibrated on your own content and languages, using texts of known origin. Instead, set the decision by the stakes of the page (standard, sensitive, critical) and by content type.

What should you do if a human text is flagged as AI-generated?

A wrongly flagged human text is not judged on its score. Re-run the protocol, gather the writing evidence (version history, brief, approvals), then decide using the decision grid. Document the incident without rewriting the text to lower the score.

Which content types are hardest to assess (legal, technical, translation, summaries)?

Legal and technical content, translations and summaries are the hardest to assess, because their regularity resembles automated writing. They require evidence + context validation and often expert human review.

How do you balance AI control, GEO and SEO without producing "templated" content?

Optimise for usefulness rather than stylistic compliance. Use structure for readability, but enrich each section with domain angles, real constraints and evidence. Templated content without differentiation harms both GEO and SEO quotability.

What evidence and sources should you keep to justify the quality of published content?

Keep the source version and its date, the contributors, the sources used, expert or legal validations, and detection and similarity reports. This provenance file makes audits and updates easier, and defends the credibility of challenged content.

Can an AI scan serve as proof?

An AI scan is an indication, not proof: it demonstrates neither the author nor the writing process. It belongs in a documented traceability file (version, date, tools, human validations), where version history and briefs carry more weight.