AI detection measures whether text looks machine-written. Fact checking measures whether a claim is sourced and true. These are two separate questions, and a detector answers only the first. A high AI score tells you nothing about whether the sentence above it is accurate, checkable, or invented.
Why this matters
A writer submits a draft, a detector returns 87% AI, and the draft gets rejected. Nobody in that chain checked a single claim. The score answered an authorship question and was read as a quality verdict.
The numbers behind that misread are public. Weixin Liang and co-authors, in arXiv 2304.02819 (published in Patterns), found that roughly 61 percent of TOEFL essays written by humans were flagged as AI-generated. Those were real people writing in a second language. The detector failed on the writer, not the content.
So here is the split you need to hold. Authorship signals live in word choice, sentence rhythm, and phrase frequency. Truth lives in whether a claim has a source attached. A detector reads the first and is blind to the second.
Does a high AI score mean the content is wrong?
No. An AI score is an authorship estimate, and authorship has no bearing on accuracy. A human can write a fabricated statistic. A model can produce a correctly sourced one. The score moves with phrasing, not with whether the claim holds up.
The ForgeRank published dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json) makes this concrete. Across 29 drafts, the scorer recorded 2,839 claims. Of those, 51 were judged to need a source, and 38 had none attached. That is a content-level finding — a claim count — and no authorship detector produces it.
The mechanism matters because it explains the failure. Detectors are trained on surface patterns: token frequency, perplexity, burstiness. A claim that needs a citation leaves no detectable surface trace. You can strip every source from a document and the AI score will not move.
The boundary: a detector can flag text that a model generated. It stops there. It cannot tell you whether the generated text is true, and it cannot tell you whether human text is true either.
What does a content-level check actually measure?
It measures whether claims carry support. The unit is the claim, not the document. A checker walks the draft, flags each assertion that needs a source, and counts how many have one attached. That count is the output.
Because the unit is the claim, the result is actionable in a way a detector score is not. When the ForgeRank dataset shows 38 of 51 source-needing claims arriving unattached, the fix is specific: attach 38 sources. You cannot "fix" an 87% AI score the same way, because the score points at style, and style is not the defect.
The consequence for the writer: you get a list of claims to source, ranked by whether each one carries its evidence. That is a to-do list. A detector score is a verdict with no repair path.
The boundary: a claim-level check verifies that a source exists and is attached. It does not verify that the source is correct or that the claim matches it. Sourcing is necessary and not sufficient.
Why do detectors flag human writers so heavily?
Because they score style, and style varies by first language. The arXiv 2304.02819 result — 61 percent of human-written TOEFL essays flagged as AI — shows the failure lands on non-native English writers first. Their phrasing is less predictable to a model trained on native text, and low predictability reads as machine output.
The mechanism is the same one that makes the score unreliable in both directions. Detectors reward text that matches their training distribution and penalize text that does not. A non-native writer sits outside that distribution through no fault of the content.
The consequence: a false positive is not a rare edge case. It is a predictable output of the method, concentrated on a specific group. Treating the score as a gate means rejecting writers for how their English reads.
The boundary: this does not mean every flag is wrong. It means the flag is a signal about phrasing, and phrasing is not the question you are trying to answer when you check a draft.
Two beliefs about AI text quality, and what the data says
The common belief: if the AI score is low, the draft is clean and ready. The evidence-supported belief: phrasing and claim quality move on separate tracks, and a low AI score leaves the sourcing question open.
The ForgeRank dataset resolves this. Eight of the nine drafts scoring 7.0 or above still received a heavy-AI-phrasing verdict. The one draft whose phrasing came back clean scored 4.0. If a low AI score meant a clean draft, that 4.0 would not exist.
The four dimension tiers show the same split. All ten drafts at 4.0 scored 4.0 on all four dimensions. All ten at 4.8 scored 4.0/4.0/7.0/4.0. Each step in the distribution is one dimension moving, not all four rising together. Phrasing and claim quality are separate axes, and the dataset keeps them separate.
A reviewer who scores drafts like this lands a heavy-AI-phrasing draft at 6/10, below the 7.0 pass line, for the same reason: the phrasing habit is visible and the sourcing gap is not.
What most people get wrong about AI detection
The mistake is treating the detector as a fact checker. It is not, and it was never built to be. Google's own spam policies name scaled content abuse — mass-producing pages to manipulate rankings — as the target, regardless of whether a human or a model wrote them. Google does not ask who wrote the page. It asks whether the page has value.
Google's March 2024 core update aimed to reduce low-quality, unoriginal content in search results by 40 percent, per Google Search Central. That target is content quality, not authorship. A detector score cannot measure it.
The misconception persists because both tools return a number, and numbers feel like verdicts. But one number estimates authorship and the other counts unsourced claims. They answer different questions, and only the second one tells you whether the draft is publishable.
Key takeaways
- AI detection scores authorship signals like phrasing and sentence rhythm; it says nothing about whether a claim is sourced or true.
- Weixin Liang and co-authors found roughly 61 percent of human-written TOEFL essays flagged as AI in arXiv 2304.02819, so false positives land on writers, not content.
- The ForgeRank published dataset recorded 2,839 claims across 29 drafts, with 38 of the 51 source-needing claims arriving unattached — a claim-level finding no detector produces.
- Eight of the nine drafts scoring 7.0 or above still carried a heavy-AI-phrasing verdict, which keeps phrasing and claim quality on separate axes.
- Google's spam policies target scaled content abuse regardless of authorship, and the March 2024 core update aimed to cut low-quality, unoriginal content by 40 percent.
Frequently asked questions
Is AI detection the same as fact checking?
No. AI detection estimates whether text looks machine-written by scoring phrasing patterns. Fact checking verifies whether a claim has a source and holds up. A detector reads style and is blind to sourcing, so a clean score leaves every claim in the draft unchecked.
Can a detector tell if a statistic is fabricated?
No. A fabricated number and a real one look identical to a detector, because the score moves on sentence rhythm and word frequency, not on whether a source exists. Catching a fabricated statistic requires a claim-level check, not an authorship score.
Why do AI detectors flag non-native English writers?
Detectors reward text that matches their training distribution. Non-native phrasing is less predictable to that model, and low predictability reads as machine output. Weixin Liang and co-authors found roughly 61 percent of human TOEFL essays flagged as AI in arXiv 2304.02819.
Should I reject a draft because the AI score is high?
Not on the score alone. A high AI score points at phrasing, which is a style question. Check the claims instead: count how many need a source and how many have one attached. The ForgeRank published dataset logged 38 of 51 source-needing claims arriving unattached.
What does Google actually penalize?
Scaled content abuse — mass-producing pages to manipulate rankings — regardless of whether a human or a model wrote them, per Google's spam policies. Google's March 2024 core update aimed to reduce low-quality, unoriginal content in search results by 40 percent, per Google Search Central.
What should I check before publishing a draft?
Walk the draft claim by claim. Flag every assertion that needs a source and confirm each one carries it. That produces a repair list. An AI score produces a verdict with no fix, which is why it cannot be the last check before you publish.
A detector score is a signal about phrasing, and phrasing is not the question you are trying to answer when you decide whether a draft goes out. The claim count is. Open your current draft tonight and mark every sentence that asserts a fact — then check how many of those marks have a source sitting next to them.