TL;DR
- An AI label is a model output, not a record of the writing process. Define what kind of AI help your brief permits.
- Older tests found false positives on human essays. A later NIST test also found strong detectors. Neither result settles the origin of a particular B2B draft.
- Check the exact file, detector version and decision threshold. Review drafts, source notes and the author’s account before drawing a conclusion.
- A low AI score does not make a claim true. Accept content against its sources, buyer usefulness and agreed process, with open issues recorded.
What Question Is the Score Supposed to Answer?
Stanford HAI’s May 2023 research article raised concerns about false accusations from AI text detectors. That concern matters when a B2B team reviews agency work. A score may become a reason to reject a writer, with little review of how the draft was made.
For generative engine optimization (GEO), define the task first. Are you checking whether a supplier followed an agreed writing process? Whether the facts are correct? Or whether the page answers a buyer’s question? One detector label cannot answer all three.
Even the process rule needs plain wording. A ban on generated first drafts differs from permission to use AI for spelling or translation. Without that distinction, people may dispute the same score while applying different rules.
A Human Essay Could Receive an AI Label
Liang and colleagues’ April 2023 paper tested seven detectors on 91 human-written Test of English as a Foreign Language (TOEFL) essays and 88 US eighth-grade essays. Mean false-positive rates were 61.22% for the TOEFL set and 5.19% for the US set. These rates average errors across the tested detectors. A false positive labels human text as AI.
The sets differ in task and age as well as language background. This is a historical sample comparison, not an isolated causal test of native language or a current B2B error rate.
The same study found that all seven detectors flagged 18 of 91 TOEFL essays (19.78%). Agreement among tools was still not proof of AI authorship. This result concerns that essay set and those detector versions.
Liang et al., 2023, Figure 1 and text. Human samples: 91 TOEFL and 88 US essays. Different tasks and populations; not current tool ratings or B2B rates. Study, samples and Figure 1.
Strong Detection Results Also Need a Scope
The US National Institute of Standards and Technology (NIST) offers a useful counterpoint. Its June 2025 report on the 2024 pilot, Tables 5–6, gives detector submission 12 an area-under-curve (AUC) score of 0.996 against GPT-3.5 and 1.000 against GPT-4. AUC measures how well scores separate human and generated summaries across thresholds. It is not the share of words written by AI.
The first round had 104 summaries on ten topics: 24 from GPT-3.5, 40 from GPT-4 and 40 from humans. This benchmark did not trace the revision history of a B2B page. Strong separation in a defined test is useful evidence of detection ability. It does not make every score a verified authorship finding.
| Direct research | Actual result | What the test establishes |
|---|---|---|
| Liang et al., 2023 | Mean false positives: TOEFL 61.22%; US essays 5.19%; all seven flag 19.78% of TOEFL set | Human essay errors under historical detectors; not a present-day B2B estimate |
| NIST, 2025 report | Submission 12: AUC 0.996 / 1.000 for GPT-3.5 / GPT-4 summaries | Strong benchmark separation; not an audit of a mixed writing workflow |
Figure 2. Evidence comparison. Different research teams, text tasks and metrics. A false-positive rate at operating thresholds and AUC across thresholds do not form a shared tool ranking.
The buyer’s useful question is narrower than “Do detectors work?” Ask whether the stated evidence supports this decision on this kind of text. Keep the test population, tool version and pass rule visible. A good benchmark and an incomplete writing record can both be real.
A Completed Review of a Disputed Agency Draft
This example is fictional. An agency submits a pump-maintenance article. The client permits AI spelling help but prohibits generated first drafts. A detector marks the final file as AI-written. The editor has interview notes and a tracked draft, but one rewritten paragraph has no recorded source. No live detector test was run for this example.
| Review question | Completed fictional record | Decision or next check |
|---|---|---|
| What was agreed? | Human first draft; AI spelling help allowed | Review the process against this rule, not an undefined AI-free standard |
| What exactly was scored? | Final submitted file; saved detector output and tool version | Keep the matching file and threshold; the label does not identify a writer |
| What supports the writing account? | Interview notes and tracked revisions explain most sections; one paragraph lacks a source | Ask for that paragraph’s origin and supporting material; leave its status open |
| What can be accepted now? | Verified maintenance steps match approved notes; the disputed paragraph is held | Resolve process compliance and factual support separately before final acceptance |
The detector has prompted a review. It has not settled the dispute. The editor can point to a specific gap and request the material needed to close it. The right fix is not to keep changing words until the score drops.
A draft history is supporting evidence, not an infallible record. Read it with the interview notes and the author’s explanation. If a step remains unclear, say so. Do not call a missing record proof of misconduct or treat a friendly explanation as proof that the rule was followed.
Check the content on its own terms too. Does each maintenance instruction have an approved source? Are limits and service conditions clear? A human can write a false claim. AI-assisted text can contain a correct one. Neither observation excuses a breach of the agreed process.
Our AI content and human expertise guide covers which expert inputs a page needs. Here the separate task is to decide what evidence is needed when a detector label is used to approve or reject the draft.
AI Detectors and B2B Authorship: Common Questions
Does an AI score show the share of words written by a tool?
Do not assume that it does. Read the provider’s definition of the score and its unit. A document-level classification, sentence flags and an estimated share are different outputs. None is a revision log.
Is a low detector score a quality check?
No. Review factual support, fit, clarity and usefulness directly. A score about likely text origin does not check product specifications or customer permissions. Keep those acceptance checks in the brief.
Should we reject work whenever several detectors agree?
Use agreement as a review signal, not a verdict. Preserve the outputs and ask what evidence supports the writing account. Tools can respond to similar text features. Count the reasons and records, not just labels.
Can an agreed process include AI editing?
Yes, if the team permits it and defines the boundary. State which tasks are allowed, what must be disclosed and who checks the facts. This is a workflow choice, not a claim that a particular AI share improves search results.
Bring the Draft, the Brief and the Disputed Score
If a detector score is driving your GEO content review, bring the file, process brief, source notes and saved output to SemShift’s GEO service discussion. We can discuss which editorial checks answer the buyer’s question and which claims still need evidence before publication.