Detecting AI writing in 2026 is harder than ever — and more overconfident detectors are being used to make consequential decisions than ever before. Students are being accused of cheating on the basis of detection scores. Job candidates are being filtered out. The problem is that current detectors can be wrong often enough to matter. This is what genuinely works, what doesn't, and how to use detection responsibly.
What changed in 2026
- Models got better at sounding human. GPT-4o and Claude 3.7-class models produce text that easily clears older detection thresholds, especially with light editing.
- AI watermarking arrived — partially. OpenAI implemented soft watermarking (C2PA metadata and statistical signal embedding) for ChatGPT outputs; detection works when the signal is preserved, but copy-paste often strips metadata.
- Detectors improved, then hit a ceiling. Tools like GPTZero, Copyleaks, and Originality.ai made real gains in 2024–2025, but their false positive rates on human-written text remain ~5–15%, which is too high for disciplinary use.
- Paraphrasing tools defeated detection. Any halfway-motivated student can paste AI output into QuillBot or a similar tool and clear most detectors.
How the main detectors work
AI detectors look for two main signals:
| Signal |
What it measures |
Weakness |
| Perplexity |
How "surprising" each word choice is (AI tends to be lower perplexity) |
Human experts also write low-perplexity prose |
| Burstiness |
Variation in sentence complexity (humans vary more) |
Varies by writing style and genre |
| Watermark signals |
Statistical patterns seeded by the model |
Only works if the signal survives editing |
Most commercial detectors combine these into a probability score. The score is not a percentage certainty — it's a classifier output with a threshold, and both sides of the threshold have error rates.
Detector accuracy: the honest picture
| Tool |
Claimed accuracy |
Reported false positive on human text |
Notes |
| GPTZero |
~85–98% |
~5–12% |
Best-known; widely used in education |
| Originality.ai |
~90–99% |
~4–8% |
Stronger on long documents |
| Copyleaks |
~90%+ |
~8% |
Multilingual; integrates with LMS |
| Turnitin AI |
~98% (claimed) |
~1% (claimed, contested) |
Used by many universities; opaque methodology |
| Winston AI |
~88–94% |
~6–10% |
Good for short texts |
A 5% false positive rate means 1 in 20 human-written documents gets flagged. In a class of 100 students, five innocent students get accused. This is not a rounding error.
Human signals that are more reliable
Rather than relying on detection percentages, look for these qualitative markers:
Likely AI patterns:
- Uniform hedging throughout ("It is important to note that...", "This can be seen as...")
- No specific examples — only categories and generalities
- No real opinion, only both-sides framing
- Perfect structure (intro, 3 points, conclusion) without deviation
- Absence of lived experience, error, or surprise
- Repetition of key phrases at regular intervals
Unlikely AI patterns:
- Specific personal anecdotes tied to verifiable context
- Unusual word choices or stylistic quirks consistent with the writer's other work
- Domain-specific jargon used precisely, including niche subtleties
- Errors that reflect a human's specific misunderstanding
- Temporal references the model couldn't know (published after the model's training cutoff)
Watermarking: the most reliable detection method
The C2PA (Coalition for Content Provenance and Authenticity) standard now embeds provenance metadata in AI-generated images and, increasingly, text. OpenAI's ChatGPT uses a version of statistical watermarking — the model skews token probability distributions in a way that's detectable with the right key.
Limitations:
- Works only for output from tools that implement it
- Copy-paste to a plain text editor often strips metadata watermarks
- Statistical watermarks survive better but can be partially degraded by heavy editing
- No universal standard across providers yet
How to pick your approach
- If you're an educator: Use detectors as one signal among many, not a verdict. Ask the student to explain their reasoning verbally or rewrite a section in class.
- If you're a publisher or editor: Look for human-signal patterns. Ask for specifics the author should know if they wrote it.
- If you're evaluating a job application: Writing tests in a controlled environment are more reliable than detection tools.
- If you need high-confidence detection: Use tools that support provenance standards (C2PA) and only accept documents submitted directly — copy-paste defeats metadata.
- If you need to audit content at scale: Originality.ai and Copyleaks have API access with bulk detection; treat the output as a triage filter, not a verdict.
Common mistakes
Using a detection score as proof. A 94% AI probability score means there's a 6% chance it's human — at scale, that's a lot of wrongful accusations.
Forgetting that editing defeats detectors. AI text that's been lightly rewritten by a human often scores as human. The detector isn't measuring effort; it's measuring signal.
Ignoring false negatives. High-quality AI writing from capable models often passes detection. Detectors are not airtight in either direction.
Treating all AI-assisted writing the same. There's a difference between a document that is 100% AI-generated and one where AI assisted with grammar or structure. The spectrum matters.
What to skip
- Building any policy around detector scores alone. The legal and ethical liability is real; no institution should discipline someone solely on a detector output.
- Free "AI detector" browser extensions — most are poorly calibrated and even more prone to false positives than the established tools.
- Prompt injection "tests" like asking AI to add phrases. These tricks mostly stopped working in 2025.
FAQ
Can an AI detector tell if I used ChatGPT?
It can estimate a probability, not confirm a fact. The score is a signal, not proof. Current top tools are ~85–95% accurate under ideal conditions and worse in practice.
Does rewriting AI text fool detectors?
Often yes. Paraphrasing tools and manual rewrites significantly lower AI scores, sometimes to zero. This is a known and unfixed weakness.
What is C2PA and does it matter?
C2PA is a standard for embedding content provenance metadata — who created what, when, with which tool. It matters for images now and will increasingly matter for text as adoption grows.
Should schools ban AI writing entirely?
That ship has sailed. Better policies focus on how AI is used, require demonstration of understanding, and evaluate process alongside output.
Where to go next