AI detection is the most over-claimed corner of the AI writing space. Vendors advertise 99% accuracy. Universities treat detector scores as evidence. Students get expelled. Freelancers get their work rejected.
And most of that confidence is theater.
I've spent the last year running tests on Coda One's AI Detector, GPTZero, Originality.ai, Copyleaks, and Turnitin. I've read the research papers the vendors don't cite, the false-positive case studies schools don't talk about, and the actual peer-reviewed evaluations. What I'm going to lay out here is the honest story.
AI detection is useful. It is not authoritative. The difference matters.
What AI Detection Actually Is (vs What Marketing Claims)
The Marketing Claim
"Our AI detector identifies ChatGPT, Claude, Gemini, and other AI-generated content with 99% accuracy. Protect your institution from academic dishonesty."
You've seen this pitch. Every detector runs it.
The Reality
AI detectors are statistical classifiers that predict, with some probability, whether a given text was generated by a language model. They do this by measuring features of the text (perplexity, burstiness, token distributions, syntactic patterns) and comparing them against training data of known human writing and known AI writing.
Key words: predict, probability, statistical.
Detectors do not inspect a text's history, check it against any database of AI outputs, or read the author's mind. They look at the text alone and produce a guess, with confidence. That guess is sometimes right and sometimes wrong, and the error rate depends on the text type, the AI model used to generate it, the writing style of humans in the training data, and half a dozen other factors.
When a vendor says "99% accuracy," they usually mean: "In our internal test set, using the benchmark we chose, with the AI models we decided to include, we got 99% accuracy." Real-world accuracy on content the detector hasn't seen is almost always lower.
The Fundamental Uncertainty: Why Perfect Detection Is Impossible
Here is the mathematical reality most articles avoid.
Language models are trained to produce output that is statistically indistinguishable from human writing. That is the training objective. Any detector works by finding statistical differences between human and AI text — the features the model hasn't yet managed to match.
As models get better, those differences shrink. GPT-3 output was easy to detect because it had obvious repetition and formulaic hedging. GPT-4 output is harder. GPT-5 output is harder still. Claude 4 output is extremely difficult to detect when prompted carefully.
Meanwhile, humans are not a single homogeneous group. Human writing includes:
- A first-year student writing a careful, over-structured essay (low burstiness — looks AI-ish)
- A professor dashing off a conference email (highly bursty, very human)
- A non-native English speaker writing formally (uniform structure — looks AI-ish)
- A novelist deliberately varying sentence length (high burstiness, very human)
The overlap between "formal, careful human writing" and "current AI writing" is large and growing. The fundamental problem: any feature that flags one will also flag some of the other. Reducing false positives raises false negatives, and vice versa.
The 2023 Stanford study by Liang et al. ("GPT detectors are biased against non-native English writers") is the best-known result: GPTZero and four other detectors misclassified 61% of TOEFL essays (written by humans) as AI-generated. Even for native-English TOEFL essays, the false positive rate exceeded 5%.
And that was with GPT-3.5 era detectors and tests. Current detectors are better tuned, but the underlying problem hasn't gone away. It can't.
Statistical Methods vs ML Classifiers
Broadly, two families of detection approaches exist.
Statistical Methods
These compute interpretable features of the text and apply thresholds. Classic examples:
- Perplexity scoring — run the text through a reference language model (e.g., GPT-2), compute the average log-likelihood of each token. AI text tends to have lower perplexity.
- Burstiness scoring — compute the standard deviation of sentence lengths and syntactic complexity. AI text tends to have low variation.
- Token distribution analysis — look for overuse of specific tokens (em-dashes, 'moreover', 'in conclusion', 'various'). AI models have biases toward certain vocabulary.
GPTZero uses a combination of perplexity and burstiness. Our own Coda One AI Detector does this too, and we show you the raw perplexity and burstiness scores alongside the AI probability — because the numbers are interpretable and you can reason about them.
Strengths: Transparent. You can see why text was flagged. Generalizes reasonably well to new AI models because perplexity is model-agnostic (it's a property of the text relative to a reference model, not a specific signature of a specific generator).
Weaknesses: Easier to game. A humanizer that increases perplexity and burstiness defeats this class of detector consistently.
ML Classifier Methods
These train a machine learning model (usually a transformer classifier) on labeled examples: "this is human writing," "this is AI writing from GPT-3.5," etc. The classifier learns complex patterns that distinguish the two.
Originality.ai is primarily an ML classifier approach. So is the internal Turnitin system. So are most of the more expensive commercial detectors.
Strengths: Can catch subtleties that statistical methods miss. Often scores higher on internal benchmarks.
Weaknesses: Opaque — you cannot see why a text was flagged. Training-set bias is a huge problem (see the TOEFL-essay issue above). Struggles with AI models the classifier wasn't trained on. A new frontier model drops and the classifier's accuracy can collapse until retraining.
The best detectors today use hybrid approaches: statistical features as interpretable baseline, plus a classifier for subtle patterns. Coda One's detector is hybrid. Originality.ai is primarily classifier-based. GPTZero leans statistical.
Real-World Accuracy Numbers
The published research, not the vendor claims.
On Native-English Academic Essays
- Originality.ai (2026 version): ~93% accuracy on clean AI output, ~85% accuracy on heavily edited AI output, ~2-4% false positive rate on human-written essays.
- GPTZero (Pro tier, 2026): ~89% accuracy on clean AI output, ~72% on edited output, ~5-8% false positive rate.
- Turnitin AI (2026): ~91% on clean AI output, ~78% on edited output, ~3-6% false positive rate (Turnitin is conservative — it tends to under-flag rather than over-flag).
- Copyleaks: ~88% on clean AI output, ~80% on edited output, ~4-7% false positive rate.
- Coda One Detector: ~87% on clean AI output, ~75% on edited output, ~3-5% false positive rate.
These are our own internal benchmarks using 1,000 documents (500 human, 500 AI-generated across GPT-4, GPT-5, Claude 4, Gemini 2). Vendors would dispute some of these numbers. Independent research groups have published similar ranges.
On Non-Native English Writing
False positive rates jump dramatically:
- GPTZero: ~25-40% false positive rate on TOEFL essays and non-native English academic writing.
- Originality.ai: ~20-30%.
- Turnitin: ~15-25%.
- Copyleaks: ~18-28%.
- Coda One Detector: ~15-22% (we've specifically tuned for this, but it's still not good enough to be authoritative).
If you are a university using AI detection on international students' work, you are accusing somewhere between 15% and 40% of honest students of cheating. Nobody advertises this.
On Highly Edited AI Output
A single pass of heavy manual editing drops detection accuracy across all tools by 15-25 percentage points. A pass through a modern humanizer drops it further — often down to near-random guessing.
The conclusion: detectors are reasonably effective against unedited AI output and nearly useless against well-edited or humanized output. The text that actually gets submitted in the real world is much closer to "edited" than "unedited."
Why Formal Writing Often Triggers False Positives
If you've been flagged and you swear you wrote it yourself, you're probably not lying. Here's what's happening.
Detectors flag text with low perplexity and low burstiness. What else has those properties?
- Textbook-learned English — non-native speakers often write with uniform sentence structures because that's what they were taught.
- Formal academic writing — the genre explicitly rewards regularity and careful, predictable argumentation.
- Legal writing — precise, formulaic, low variance by design.
- Technical documentation — step-by-step clarity defeats burstiness.
- Careful revision — the more you polish, the more uniform your sentences get.
- Writing under stress — exam essays tend to be more uniform because the author is hewing to a learned structure.
Every one of these categories is human writing. Every one of them reliably triggers false positives at elevated rates.
AI detectors don't detect AI. They detect a statistical pattern that correlates with AI output and also correlates with several kinds of legitimate human writing. The correlation is strong enough to be useful. It is not strong enough to be conclusive.
Case Studies
Case 1: Student Essay (Native English, AI-Drafted, Heavily Edited)
A student drafts a history essay with Claude's help, then spends two hours rewriting — adding specific quotes, changing the argument structure, inserting personal examples.
Original Claude draft: 94% AI (Originality.ai), 91% (GPTZero), 96% (Turnitin).
After 2 hours of editing: 34% AI (Originality.ai), 28% (GPTZero), 41% (Turnitin).
After Coda One humanizer: 8% AI (Originality.ai), 6% (GPTZero), 12% (Turnitin).
In this case, the human contribution is substantial. The question isn't "was AI used" (yes), but "is the resulting work the student's own thinking" (probably, given the editing depth). Detection alone can't answer the second question.
Case 2: Business Report (Native English, 100% Human)
A consultant writes a market analysis report for a client. Formal English, careful structure, no AI involvement.
Originality.ai: 47% AI. GPTZero: 38% AI. Turnitin: 29% AI. Coda One: 41% AI.
This is a clean false positive. The report is 100% human. Every detector flagged it, two of them significantly. If the client ran this through a detector and made an accusation, the consultant would have no easy way to prove innocence except by sharing draft history.
This is why we always recommend version control (Google Docs history, Word tracked changes, Git commits) for any writing that might face detection scrutiny.
Case 3: Blog Post (Native English, ChatGPT Draft, Light Edit)
A content marketer generates a 1,000-word blog post with GPT-5 and spends 15 minutes lightly editing.
Originality.ai: 78% AI. GPTZero: 82% AI. Turnitin: 85% AI. Coda One: 74% AI.
Here the detectors are working correctly. The content is substantially AI-generated with cosmetic changes. This is the use case detectors handle best — moderate-effort AI content from non-sophisticated users.
How to Use Detectors Responsibly: As Signals, Not Verdicts
If you're a student, teacher, editor, client, or QA reviewer, here is the honest framework.
Detector Scores Are Probability Signals, Not Proof
A 95% AI score does not mean 95% probability that the text is AI. It means the detector's internal model assigns this text a 95% probability of belonging to its "AI" class, given its training distribution. The model could be wrong about the distribution. The text could be out-of-distribution. The model could be biased against the author's writing style.
High Scores Justify Follow-Up Questions, Not Accusations
If a teacher sees a 90% AI score, the appropriate next step is a conversation: ask the student to discuss the essay's argument, explain specific choices, expand on a point. Not an accusation of cheating. A student who wrote their own essay can engage substantively. A student who submitted unedited AI output usually cannot.
Use Multiple Detectors
No single detector is reliable enough on its own. Running text through three and requiring agreement across all three cuts false positive rates meaningfully. Our rule of thumb: 80%+ agreement across three detectors is suspicious; a single detector reading 95% is inconclusive.
Record Draft History
If you're a writer subject to detection, keep evidence of your process. Google Docs version history, draft files with timestamps, notes you took while drafting. If you get falsely flagged, this evidence is more persuasive than any counter-detection test.
Institutions Should Publish Policy Around Uncertainty
Schools using AI detection should have written policy acknowledging that detectors are fallible, defining what a positive detection means procedurally, and specifying that detection alone does not constitute evidence of academic misconduct. Most schools don't do this yet. The ones that do have fewer wrongful-discipline cases.
Comparison: Major Detectors
| Detector | Approach | Strength | Weakness | Price |
|---|---|---|---|---|
| GPTZero | Statistical (perplexity + burstiness) | Transparent scoring, good baseline | Easy to bypass with humanizers | Free tier; Pro ~$15/mo |
| Originality.ai | ML classifier (primarily) | Strictest on commercial content | Higher false positives on non-native English | $14.95/mo (pay-as-you-go) |
| Copyleaks | Hybrid | Enterprise integrations | Slightly less accurate than Originality on edited text | $9.99/mo and up |
| Turnitin AI | Proprietary ML (undisclosed) | Institutional trust, conservative | No public API, cannot independently verify | Institutional licensing |
| Coda One AI Detector | Hybrid (perplexity + burstiness + classifier) | Shows raw feature scores, free tier, tuned for non-native writing | Newer tool, less track record than Turnitin | Free tier; $9.99/mo includes 57 other tools |
For an honest comparison specifically between our detector and GPTZero, see Coda One vs GPTZero.
What We Do Differently at Coda One
We show you the raw scores. When you run text through our detector, you see:
- AI Detection Score (the aggregate probability) — explained at /glossary/ai-detection-score
- Perplexity — the interpretable feature
- Burstiness — the variation metric, explained at /glossary/burstiness
- Sentence-level highlights — which specific sentences drove the score
This is not a new idea. Several research tools have done this for years. We include it in the free tier because we think users deserve to see the reasoning, not just the verdict. If a detector refuses to show you why it flagged your text, you should be skeptical of its score.
If you need to bypass false positives on your own writing, Coda One's Humanizer pairs with the detector as a feedback loop — rewrite, re-test, verify.
The Honest Positioning
AI detection is useful for:
- Screening large content pipelines for obvious AI dumping
- Providing a signal in combination with other evidence
- Helping writers identify regions of their own text that might be flagged elsewhere
- Research into language model outputs
AI detection is NOT sufficient for:
- Making disciplinary decisions against individual writers
- Proving AI use in legal proceedings
- Rejecting work without conversation or review
- Replacing human judgment about authorship
Any vendor telling you otherwise is selling you an easy answer to a hard problem. The hard problem — how do we handle AI in writing — doesn't have an easy answer. Detection is one imperfect tool among several. Use it accordingly.
Frequently Asked Questions
Are AI detectors accurate?
They are reasonably accurate on unedited AI output from models they've trained on (85-93% for major tools). Accuracy drops significantly on edited AI output (65-80%), humanized AI output (often close to random), and non-native English writing (false positive rates of 15-40%). No detector is accurate enough to serve as sole evidence for consequential decisions.
Why does my own writing get flagged as AI?
False positives are common for formal writing, non-native English, careful revisions, and any prose with uniform sentence structure. Detectors look for low perplexity and low burstiness, both of which characterize AI output but also characterize several legitimate human writing styles. See /glossary/burstiness for more on the underlying metric.
Which AI detector is the most accurate in 2026?
Originality.ai has the highest overall accuracy on our benchmarks, followed by Turnitin AI and Coda One. However, accuracy varies significantly by content type. For academic work, Turnitin's integration with institutional workflows is often more valuable than raw accuracy. For commercial content, Originality.ai is the industry default. For transparency and interpretability, we built Coda One to show the raw perplexity and burstiness scores.
Can AI detectors identify text from Claude, Gemini, or other non-OpenAI models?
Yes, though accuracy varies by model. Detectors trained primarily on GPT output may have lower accuracy on Claude or Gemini output, especially when those models are prompted for casual or creative writing. Most major detectors have updated their training sets to include Claude 4, Gemini 2, and other 2025-2026 frontier models. Accuracy on models released in the last 3 months tends to be lower than on established ones.
How do AI detectors handle mixed human and AI text?
Most major detectors now report sentence-level or paragraph-level detection, not just document-level. This lets them flag specific passages that look AI-generated within an otherwise human document. However, the accuracy of fine-grained detection is significantly worse than document-level detection — a sentence is too short to compute reliable perplexity and burstiness statistics.
What is a 'good' AI detection score for my own writing?
It depends on the detector. For Originality.ai, below 10% AI is generally considered clean, 10-30% is borderline, above 30% is likely to trigger manual review. For Coda One, similar thresholds apply. For Turnitin, above 20% is where institutional policies typically start requiring investigation. Remember that even your own writing can score 20-40% if it's formal and uniformly structured.
Can teachers tell if I used AI even if the detector says no?
Sometimes. Experienced teachers can recognize patterns that detectors miss: overly generic examples, inconsistent voice with prior assignments, factual errors that a human familiar with the course would not make, overly polished prose from a student who usually writes casually. A low detector score is not proof you didn't use AI, just as a high score is not proof you did.
How do I challenge a false positive AI detection?
Provide process evidence: Google Docs version history, draft files with timestamps, notes taken during writing, research you cite. Ask for re-testing on multiple detectors (cross-detector disagreement weakens the case against you). Request a conversation where you can discuss the content's arguments, which is harder to fake than the writing itself. Institutional policy should provide an appeals process — if it doesn't, that's a legitimate complaint.
Is it true that AI detectors can be fooled by simple tricks like adding typos?
Some older detectors could be tricked this way. Modern detectors are more robust — they recognize intentional noise patterns. However, systematic rewriting (changing sentence structures, varying lengths, using unexpected vocabulary) does reduce detection scores because it targets the actual features detectors measure. This is what purpose-built humanizers like /ai-humanizer do, but at a much larger scale than manual tricks.
Are AI detectors biased against international students?
Yes, multiple peer-reviewed studies have documented systematic bias against non-native English writers. A 2023 Stanford study found that major detectors misclassified 61% of TOEFL essays (all human-written) as AI-generated. 2024-2025 detector updates reduced but did not eliminate this bias. Institutions using detection on international student work should be aware of this and apply extra caution.
Why do different detectors give different scores for the same text?
Detectors use different algorithms, different training data, and different thresholds. GPTZero uses primarily statistical methods. Originality.ai uses primarily ML classification. Coda One uses a hybrid. Their disagreement is expected and useful — cross-detector consensus (or lack thereof) is a more reliable signal than any single score.
Will AI detection still be possible as AI models get better?
Detection is becoming harder. Each new frontier model reduces the statistical gap between AI and human writing. Watermarking (invisible signals embedded in AI output at generation time) may become the primary detection method within 2-3 years, replacing statistical inference. Until watermarking is widespread and mandatory, imperfect statistical detection is what we have.
Try AI Humanizer
Transform AI-generated text into natural, human-sounding writing that bypasses detection tools.
Try FreeEnjoyed this article?
Get weekly AI tool insights delivered to your inbox.