If you write online, someone has probably run your work through an AI detector by now. Editors use them to screen submissions. Teachers use them to check essays. Freelancers use them to prove their work is human. The tools advertise accuracy above 99% and sound reassuring.
The reality is messier, and it matters.
In 2026, researchers at the University of Florida published one of the first large-scale independent tests of commercial AI text detectors at the IEEE Symposium on Security and Privacy — one of the most respected security conferences in the world. Their conclusion: current commercial AI detectors are not reliable enough for any high-stakes authorship decision. False positive rates ranged from 0.05% to 68.6% across the five detectors tested, and a single one-step adversarial rewrite could drop the best detector’s detection rate from 94% to 2.5%.
That is not a small caveat. That is the story.
If you are new to this category, it helps to see the full picture. The best AI writing tools produce the kind of text these detectors try to catch, and an AI-assisted editing workflow is where many writers first run into false positives.
What These Tools Claim
GPTZero is the most recognizable name in the space. Founded in January 2023, it claims more than 17 million users and positions itself as the leading AI detector. Its homepage advertises “99% Accuracy” and says it detects content from ChatGPT, GPT-5, Gemini, Claude, and Llama.
Originality.ai takes a different lane. It is built for publishers and SEO teams, sells itself as the “most accurate AI content detector on the market,” and integrates with Chrome, API workflows, and major CMS platforms. It has a free Chrome extension, which is how many people first encounter it.
Both tools explain their method similarly: they look for statistical patterns in text. GPTZero’s documentation describes three main signals:
- Perplexity — how predictable the writing is. AI text tends to be more predictable.
- Burstiness — how much sentence length and style vary. Human writing varies more.
- Style — whether the tone is generic or repetitive.
That sounds reasonable. The question is whether those signals hold up in real, high-stakes text. Independent testing says they do not, at least not reliably.
What Independent Testing Found
The Florida researchers collected every paper published at four top security conferences (IEEE S&P, CCS, NDSS, and USENIX Security) up to the end of 2022 — 6,295 papers total. They generated AI versions of each paper’s abstract and introduction using GPT-4.1, built a balanced dataset of 25,144 human/AI samples, added 6,905 adversarial samples, and ran five commercial detectors over all of it.
The total cost: about $5,000. That is the quiet reason nobody had done this properly before — running commercial detectors over a large academic dataset is expensive.
The results, in brief:
| Detector | False Positive Rate | Accuracy |
|---|---|---|
| Best two (Detector-1, Detector-2) | 0.3% – 0.6% | 96.9% – 97.6% |
| Detector-3 | — | Nearly useless (99.6% false negatives, flagged almost nothing as AI) |
| Detector-4 | 65.8% | Unusable |
| Detector-5 | 68.6% | Unusable |
Two of the five were usable in a clean baseline. Three were effectively broken — one flagged almost nothing, and two flagged more than half of all human text as AI.
Then the researchers pushed further.
Longer Text Makes Detection Worse
You would assume longer documents are easier to detect. The vendors imply it. The data says the opposite.
The researchers took 628 abstracts each detector was 100% confident was AI-generated, appended the corresponding human-written introductions (making the text about six times longer), and resubmitted.
The best detector’s true positive rate dropped from 628 to 467 — a 25.6% relative decline. The second-best dropped 8%. Both shifted toward judging text as “human.”
Why? Academic introductions are full of citations, technical terms, and complex argument structures — features that look statistically similar to human writing. The AI signal in the short abstract gets diluted. Long, complex documents are harder to detect, not easier.
One Step Defeats the Best Detectors
The paper’s core finding is the adversarial part, and it is remarkably simple.
Attack one: ask the LLM to rewrite the abstract at a very low Flesch Reading Ease score — deliberately convoluted, hard-to-read prose, without changing meaning. The best detector’s true positive rate collapsed from 94.2% to 36.2%. The second-best only dropped to 87.8%.
Attack two: run the AI abstracts through a commercial “humanize” service — the kind of tool that rewrites AI text to look more human. This time the roles reversed. The first detector was barely affected (3 misclassifications), but the second detector’s true positive rate collapsed from 95.8% to 2.5% — it recognized only 16 out of 628 AI-generated abstracts.
The two attacks hit the two detectors differently, which suggests they rely on different surface-level statistical features. Neither feature is meaning. Neither detector understands whether text is coherent or sensible.
The paper even tested a tool called SCIgen, which generates fluent-sounding but meaningless nonsense papers. The best detector caught only 12% of them. One detector flagged zero.
What This Means for You
### If you are a writer or freelancer
Do not rely on an AI detector to prove your work is human — and do not let a false positive destroy your reputation. A detector flagging a 0.3% false positive rate still means that, on a large site processing thousands of pieces, real human writing will get flagged. If that happens, you have no meaningful way to “prove” the text is yours, because the detector’s output is not evidence — it is a guess dressed as a percentage.
### If you are an editor or site owner
Treat detectors as a signal, never as a verdict. The IEEE paper’s real-world simulation makes this concrete: assume a venue receives 1,600 submissions, 1% of which are adversarial AI. After an attack, the best detector would correctly flag about 6 of 16 AI papers, miss 10, and mislabel about 5 human papers. Detector-2 would flag roughly 16 human papers and catch essentially none of the AI ones.
Running detection as a gate adds review burden to human work without reliably catching the AI it targets.
### If you are choosing between GPTZero and Originality.ai
Here is the honest version:
- GPTZero is the most accessible and the most widely adopted, especially in education. Its integrations (Google Docs, Classroom, Chrome) are excellent, and its detection philosophy — with its “writing report” and human verification features — is thoughtful about false positives.
- Originality.ai is built for publishers and SEO teams, with stronger API and CMS integration and a pricing model aimed at volume scanning. It markets itself as the most accurate for content teams.
On pure detection accuracy in independent academic testing, both products fall somewhere in the anonymous mix — the study anonymized the five detectors, and the two usable ones (Detector-1, Detector-2) were not publicly identified. Neither vendor’s “99%” marketing survives peer review, and both are vulnerable to the same one-step attacks.
The real difference is use case, not magic accuracy. Pick GPTZero if you want classroom integrations and human-writing verification. Pick Originality.ai if you need API-driven scanning at volume for a content operation.
The Honest Bottom Line
AI detectors are useful as a rough triage tool — a way to say “this looks worth a closer look.” They are not proof, and they are not safe to use as a gate in high-stakes settings.
The Florida researchers end the paper with a pointed recommendation: stop pouring resources into the detection arms race and instead invest in transparency policies — requiring disclosure of substantive AI use, and accepting that detection technology will keep losing as LLMs get better. That is where the field is heading, and it is worth knowing before you stake your credibility, or your clients’, on a percentage from a detector.