DEV Community

Paragon for Paragon

Posted on Originally published at paragonaimag.substack.com

How AI Detectors Work

#ai

Developer note: Series 3, Episode 4 deep dive for Paragon AI MAG. How do AI detectors work? The mechanism: statistical signals (perplexity — how surprised a model would be by your exact words — and burstiness — how much that surprise varies), trained classifiers (Turnitin's transformer scores each sentence 0 to 1), and watermarks (Google's SynthID plants a hidden pattern in word choices). We also cover the evidence: Liang et al. 2023 (61.22% average false positives on 91 TOEFL essays; 11.6% after vocabulary enrichment, 56.65% after simplification), Turnitin's intentional 85%/15% trade-off, the 2026 tests, and the honest alternative — C2PA provenance. Canonical URL: https://paragonaimag.substack.com/p/how-ai-detectors-work

Watch the video: https://youtu.be/RyvEyk0pdYs
Read the full article: https://paragonaimag.substack.com/p/how-ai-detectors-work

How AI Detectors Work — and Why They Fail

Paragon AI MAG — Series 3, Episode 4: "How the AI Tools You Use Actually Work"

A 100% AI score on an essay he wrote himself

On 7 November 2024, Orion Newby, a first-year student at Adelphi University in New York, handed in a history essay he wrote himself. On 20 November his instructor filed an academic integrity report: Turnitin's AI detector had scored the essay 100% machine-generated. Newby denied it and showed two other detectors that rated the same essay human-written. On 3 December the university found him responsible anyway; his appeal was denied on 7 December. On 28 January 2026, Justice Randy Sue Marber of the Supreme Court of Nassau County annulled the finding and ordered his record expunged. The court ruled on weak evidence and unfair process — not on whether detectors work.

A computer looked at words a real student wrote and said, with total confidence, "a machine wrote this." How does a detector decide who wrote something — and why does it get it so wrong?

The short answer: a detector never watches you write

A detector never watches you write. It doesn't see your keystrokes, your backspaces, your moments of hesitation, or your thought process. It only reads the finished text. It guesses authorship based on how predictable your words are compared to a massive machine data set. A score is a probability, not proof. It's an estimate of how machine-like your writing seems to a mathematical algorithm. Nothing more.

The core idea is simple. A language model writes by picking the most likely next word. A detector looks for that fingerprint — text that is too predictable. Early detectors measured two things: perplexity, how surprised a model would be by your exact words (low surprise looks machine-made), and burstiness, how much that surprise varies across the text (human writing varies more). But the trap is that plain, simple human writing gets flagged just for being too clear and predictable.

Our verdict: 🟡 WATCHING. AI detectors are a useful clue and a terrible judge. They can point to text that might be worth a second look by a human being, but they can never definitively prove who wrote it — and no score should ever decide a student's future on its own.

The three ways a detector works

Method 1: statistical signals. Perplexity and burstiness — the early GPTZero approach. It works on raw AI text and misfires on plain, simple human writing.

Method 2: a trained classifier. Turnitin's model is transformer-based; it scores each sentence between 0 and 1 and labels sentences above a threshold as AI. It is only as good as its training data, and nobody outside the company can see how it decides.

Method 3: watermark. Google's SynthID plants a hidden pattern in the word choices while the AI writes, so its maker can check later. It only finds text from tools that watermark; heavy rewriting or translation weakens it, and it works less well on factual answers.

The honest alternative is provenance: C2PA Content Credentials are a signed record of how a file was made — like a nutrition label. They are only present if the creating tool adds them.

The evidence: who gets flagged

The most important study in this field comes from Stanford. Liang and colleagues, published in Patterns in July 2023, tested seven popular detectors on 91 human-written TOEFL essays by non-native English speakers and 88 essays by US eighth-graders. On the eighth-graders the detectors were near-perfect. On the TOEFL essays they wrongly flagged an average of 61.22%.

The mechanism is the giveaway. The flagged essays had lower perplexity — simpler vocabulary, safer sentence patterns. When researchers enriched the essays' vocabulary, false positives fell to 11.6%. When they simplified the native eighth-graders' essays, false positives jumped to 56.65%. The detectors were measuring how elaborate the writing is, not who wrote it.

The absurdity has a famous example: in July 2023, Ars Technica documented excerpts of the US Constitution flagged as AI-written by GPTZero and ZeroGPT. If your writing is highly structured, formal, or just lacks wild conversational variation, you look suspiciously like a language model to these scanners.

Vendors will tell you they have fixed this. Turnitin claims it keeps its document-level false-positive rate under 1%. But to achieve that, its Chief Product Officer says it intentionally detects about 85% of AI content and deliberately misses about 15%. It even hides scores from 1% to 19% behind an asterisk, because the data in that range is too noisy to trust. Think about that: the tool admits its own lower-end estimates are essentially statistical static.

And what does a 1% error rate actually look like in a real school? Vanderbilt University submitted roughly 75,000 papers to Turnitin in a single academic year. A 1% error rate would falsely accuse 750 completely innocent students. Vanderbilt disabled Turnitin's AI detection in August 2023. The human cost of that 1% was just too high.

Even OpenAI couldn't detect its own text

You'd think the creators of ChatGPT could spot their own text. OpenAI built its own AI Classifier and launched it on 31 January 2023. In its own tests it correctly flagged only 26% of AI-written text. In July 2023, OpenAI withdrew it, citing "its low rate of accuracy". If the creators of the leading language model in the world can't reliably detect it, the whole premise of the software is in question.

The confusion around this technology produced a viral case at Texas A&M University-Commerce in May 2023: an instructor pasted student essays directly into ChatGPT and asked the bot, "did you write this?" The bot, which is designed to please the user and frequently hallucinates, confidently said yes. Temporary failing grades followed for an entire class until a manual individual review cleared the students. You cannot ask a text generator to act as a lie detector.

What to do if a detector flags you

If you are a student:

  1. Always draft in Google Docs or Word Online. Version history is your bulletproof evidence. It tracks every keystroke and proves your human effort over time.
  2. Preserve all your messy physical notes, outlines, and early drafts.
  3. If flagged, demand the full sentence-level report, not just the final percentage score.
  4. Never edit a flagged document after the fact. Preserve the evidence exactly as it was submitted.

If you teach:

  1. State your exact AI policy in writing before the term begins, so everyone is on the same page.
  2. Request process statements from students outlining exactly how they researched and drafted their work.
  3. Talk to the student first to understand their workflow before ever filing an official academic report.
  4. Remember that even Turnitin itself says its score "should not be used as the sole basis for adverse actions against a student". The score is a prompt for a conversation, not a gavel.

The 2-minute self-test: paste a page of your own writing — something you wrote five or ten years ago, well before ChatGPT existed — into a free online detector and watch the score. Then paste a famous human text like the US Constitution or a classic novel, and watch the software fail right before your eyes. If a detector can call the Constitution AI-written, a percentage is not proof of anything.

What comes next

Real technical solutions are emerging to replace these flawed guessing games. Google has open-sourced SynthID Text and extended watermarking to Gemini. C2PA Content Credentials are spreading to cameras and phones. These tools prove provenance rather than relying on statistical guesswork after the fact. Detectors and "humanizer" tools are locked in an arms race — but the honest direction is proof of process, not guessing after the fact.

Until those solutions are implemented everywhere, you are the final line of defense against a false flag. A percentage is not proof of anything. Process is your only protection. When the algorithm inevitably gets it wrong, your messy human drafts in your version history are the only true way to prove your authorship.

So keep your drafts, know your rights, and never let a probability score erase your hard work.


Paragon AI MAG — one video and one podcast episode every day. Subscribe to the Paragon AI MAG YouTube channel and follow the show on RedCircle so you never miss an episode.

Sources: Liang et al., "GPT detectors are biased against non-native English writers", Patterns, July 2023; Quantslant / OpenClassActions, Newby v. Adelphi (court record, Supreme Court of Nassau County, 28 January 2026); Turnitin (Chief Product Officer statement; FAQ); OpenAI AI Classifier (launched 31 January 2023, withdrawn July 2023); Ars Technica, July 2023; Vanderbilt University; Texas A&M University-Commerce, May 2023; Google SynthID Text; C2PA Content Credentials.

Top comments (0)