How AI Detectors Work

AI detectors analyze text for patterns that human writers rarely produce and machine learning models generate consistently. The technology falls into three categories: statistical analysis, classifier models, and watermark detection. No single method is foolproof. The platforms ranked in the Global 100 Text Detection rankings use different combinations of these approaches, which is why their accuracy varies.

Three Methods Behind How AI Detectors Work

Understanding how do ai detectors work requires knowing the three technical approaches in production systems.

Statistical analysis measures perplexity and burstiness. AI text tends to be more predictable token by token than human text. Detectors measure this consistency. Perplexity quantifies how surprised a language model is by the next word. Lower perplexity equals more predictable writing. AI output typically scores lower because the model generates the most probable next token. Burstiness measures variation in sentence length and structure. Humans write with high burstiness, mixing short punchy sentences with longer explanatory ones. AI text tends toward uniform sentence lengths. Early detectors relied heavily on these signals. By 2024, humanizer tools had learned to manipulate both metrics, rendering pure statistical approaches unreliable.

Classifier models are the dominant method in 2026. A neural network is trained on millions of human-written and AI-generated samples. The model learns subtle patterns invisible to human readers. Sentence cadence. Word choice distributions. Transition phrase frequency. The training set matters. A classifier trained only on GPT-3.5 output will miss Claude or Gemini patterns. Most commercial detectors retrain monthly as new models launch. This is the method used by Turnitin, Originality.AI, and GPTZero according to GPTZero's published methodology.

Watermark detection requires cooperation from the AI model itself. Some models (Google Gemini via SynthID) embed invisible signals at generation time. Special detectors read those signals. The watermark survives paraphrasing and light editing. It does not survive rewriting through a second AI model. As of 2026, watermarking is the most reliable method when both the generator and detector are part of the same ecosystem. Google's SynthID technical documentation explains the cryptographic approach. The limitation is adoption. Most AI-generated text in the wild comes from models that do not watermark.

What Perplexity Means in AI Detection

Perplexity is the technical term for how surprised a language model is by the next word in a sequence. Lower perplexity equals more predictable text. AI text typically has lower perplexity than human writing because the model generates the most probable next token at each step.

A perplexity score is calculated by running the text through a language model (often GPT-2 or a similar public model) and measuring the probability the model assigns to each word given the preceding context. If the model assigns high probability to every word, perplexity is low. If the model is frequently surprised, perplexity is high.

Human writers do not optimize for the most probable next word. Humans write with idiosyncrasy. Sentence fragments. Unexpected metaphors. Deliberate repetition for emphasis. These choices raise perplexity. AI models, by default, generate smooth, probable text. This creates a detectable signature.

The problem is that this signature is not unique to AI. Non-native English speakers often write with lower perplexity because they favor simpler, more common sentence structures. Technical writing has low perplexity by design. Academic abstracts, legal documents, and instruction manuals all score similarly to AI text on this metric.

For a deeper explanation of what perplexity means in AI detection, including the mathematics, see our technical guide.

What Burstiness Measures

Burstiness measures variation in sentence length and structure. Humans write with high burstiness. We mix short sentences with long ones. We use fragments for emphasis. We vary rhythm instinctively.

AI text tends toward uniform sentence lengths. GPT models generate sentences that average 15 to 25 words with little deviation. The rhythm is smooth but mechanical. Early detectors used burstiness as a primary signal. A document with every sentence between 18 and 22 words likely came from a model.

The metric is calculated by measuring the standard deviation of sentence lengths across a document. High standard deviation equals high burstiness. Low standard deviation flags AI authorship.

This worked until 2023. Then humanizer tools like Undetectable.AI and StealthWriter learned to inject artificial burstiness. They rewrite AI output to include one-word sentences, fragments, and 40-word run-ons. The rewritten text passes burstiness checks while remaining substantively AI-generated.

By 2026, burstiness alone is not a reliable signal. It is still used as one feature among dozens in classifier models, but any detector relying solely on sentence-length variation can be trivially defeated.

How Classifier Models Learn AI Patterns

A classifier model is a neural network trained to distinguish human text from AI text. The training process requires millions of labeled samples. Half human-written. Half AI-generated. The model learns patterns in word choice, sentence structure, punctuation frequency, and transition phrase placement.

The patterns are not obvious to humans. A 2024 study by Stanford HAI found that classifiers detect correlations across word sequences separated by hundreds of tokens. Humans cannot perceive these long-range dependencies. The model can.

Training a production classifier requires several steps. First, assemble a diverse corpus. Academic essays. Business emails. News articles. Social media posts. Creative fiction. All human-written. Then generate an equivalent corpus using GPT-4, Claude, Gemini, and other models. Label each sample. Train the neural network to predict the label given the text.

The accuracy of the final model depends on corpus quality. If the human samples all come from professional writers and the AI samples all come from GPT-3.5, the classifier will fail on mediocre human writing and succeed on advanced AI output. This is why false positive rates across detectors vary so widely. A classifier trained on a narrow corpus overfits to that distribution.

Most commercial detectors retrain monthly. New AI models launch. Old patterns become obsolete. GPT-5, released in late 2025, produces text with significantly higher burstiness than GPT-4. Detectors trained before GPT-5 launch flagged its output as human. Retrained models caught up within weeks.

The Global 100 tracks how detection accuracy is measured across the 26 platforms we evaluate. Classifier-based detectors consistently outperform pure statistical methods, but no classifier reaches 100% accuracy.

Why Detectors Fail in Predictable Ways

Detectors fail in three predictable scenarios. Understanding these failure modes explains the limits of AI detection in 2026.

Humanizer tools rewrite AI output to break statistical signatures. Undetectable.AI, StealthWriter, and similar platforms take ChatGPT text and transform it. They inject burstiness by varying sentence lengths. They raise perplexity by replacing common words with rare synonyms. They remove AI-tell transition phrases. The rewritten text often passes detectors even though a human never wrote a single sentence. As of 2026, no detector reliably catches humanized text. The Global 100 methodology includes a humanization stress test for exactly this reason.

Non-native English speakers trigger false positives. Writers learning English tend to use simpler grammar, favor common words, and write with lower burstiness. These are the same signals that flag AI text. A 2025 study found that detectors flagged 23% of ESL-authored academic papers as AI-generated. The rate for native speakers was 4%. This discrepancy creates serious equity problems in education and hiring. Institutions relying on AI detection without manual review are disproportionately penalizing non-native writers.

Newer LLMs are deliberately trained to produce human-like variation. OpenAI, Anthropic, and Google all publish research on making AI text less detectable. GPT-5 includes a "natural writing mode" that increases burstiness and perplexity. Claude 3.5 variesits sentence rhythm deliberately. These models are trained on human feedback that rewards less mechanical output. The result is text that passes many detectors by design. This is not an accident. Model developers know institutions use detection tools. Making output less detectable is a feature, not a bug.

Watermark Detection and Why It Matters

Watermark detection is the only method that approaches reliability, but it requires cooperation from the AI provider. The model embeds an invisible signal at generation time. A compatible detector reads that signal. The watermark survives paraphrasing, light editing, and even translation in some implementations.

Google's SynthID is the most advanced watermark system in production. The Google SynthID technical documentation describes a cryptographic approach that modifies token probabilities during generation. The changes are imperceptible to readers but statistically detectable. Text watermarked by Gemini can be verified by SynthID scanners with near-perfect accuracy.

The limitation is adoption. SynthID only works on text generated by Gemini. ChatGPT does not watermark. Claude does not watermark. Most AI text in the wild comes from unwatermarked models. Even if OpenAI adopted watermarking tomorrow, the billions of words already generated would remain unmarked.

Watermarking also requires the generator to opt in. A user who wants undetectable output can simply use a model that does not watermark. This creates an adverse selection problem. The only AI text that carries watermarks is text from users who do not care about detection. The users trying to evade detection will migrate to unwatermarked models.

Despite these limits, watermarking is the most promising long-term solution. If industry coordination or regulation mandates watermarking across all major models, detection becomes tractable. Until then, watermark detectors are a niche tool for specific ecosystems.

What Production Detectors Actually Use

Most commercial AI detectors in 2026 combine multiple methods. Originality.AI uses a classifier model plus perplexity analysis. Turnitin uses a classifier trained on academic writing specifically. GPTZero uses a three-layer approach: perplexity, burstiness, and a classifier trained on mixed domains according to GPTZero's published methodology.

The platforms ranked in the Global 100 Text Detection rankings disclose their methods with varying degrees of transparency. Some publish whitepapers. Others reveal only high-level descriptions. None publish training data or model weights, which makes independent replication impossible.

The lack of transparency creates a trust problem. Institutions adopt these tools without understanding how they work or where they fail. A university buys Turnitin and assumes 95% accuracy because that is the number in the marketing material. The actual accuracy on non-native English speakers or humanized text is unknown because Turnitin does not publish those breakdowns.

The Global 100 addresses this gap by testing all platforms on the same corpus under controlled conditions. We measure accuracy on unmodified AI text, humanized text, non-native writing, and mixed authorship. We publish the results. This allows institutions to choose detectors based on performance in their specific use case, not vendor claims.

The Detection Arms Race

AI detection is an arms race. Model developers make output less detectable. Detector developers retrain on newer samples. Humanizer tools find new ways to break signatures. Detectors add new features to catch humanized text. Neither side wins permanently.

This dynamic accelerated in 2025. GPT-5 launched with deliberate anti-detection features. Originality.AI retrained within two weeks. Undetectable.AI updated its humanizer to defeat the new classifier. Originality.AI added a humanization detector. The cycle repeats monthly.

The practical implication is that any detection system requires continuous updates. A classifier trained in January 2026 will underperform by June. Institutions that buy detection and assume it works indefinitely are building on sand. Detectors are a subscription service, not a one-time purchase, because the underlying technology decays.

This also means older AI text becomes harder to detect over time. A detector trained on GPT-5 patterns will miss GPT-3.5 output because the training corpus no longer includes samples from that distribution. Archive detection (scanning documents written in 2023) requires older classifiers that most vendors do not maintain.

When Detection Works and When It Does Not

Detection works best on unmodified AI text from older models. A GPT-3.5 essay pasted directly into a submission form will flag on any serious detector. Detection works moderately well on lightly edited AI text. Changing a few words does not break the statistical signature.

Detection fails on humanized text, mixed authorship, and non-native writing. It fails on text from models deliberately trained to evade detection. It fails when the user knows how detectors work and writes to avoid their signatures.

The failure modes are not random. They are predictable and exploitable. Any motivated user can defeat any detector by running output through a humanizer, rewriting with a second AI model, or manually editing to inject burstiness. This means detection only catches careless users or those unaware that humanization tools exist.

Institutions should treat detection as a screening tool, not a verdict. A high AI probability score justifies a conversation. It does not justify automatic penalties. A low score does not prove human authorship. It proves the text does not match the detector's training distribution.

The platforms in the Global 100 index vary in how they communicate this nuance. Some (Copyleaks, GPTZero) provide probability scores and encourage human review. Others (Winston AI, ContentDetector.AI) provide binary verdicts that institutions misinterpret as definitive. The framing matters. A tool that says "99% AI" feels authoritative. A tool that says "high probability, recommend review" invites appropriate skepticism.

Frequently Asked Questions

What is the most common AI detection method?

Classifier models trained on millions of human and AI-written samples. The neural network learns patterns invisible to human readers.

What is perplexity in AI detection?

Perplexity measures how predictable text is word by word. AI text typically has lower perplexity because the model generates the most probable next token.

What is burstiness in AI detection?

Burstiness measures variation in sentence length and structure. Humans write with high burstiness. AI tends toward uniform rhythm.

How do watermark-based detectors work?

Some AI models embed invisible patterns at generation time. Special detectors read those signals. Google SynthID is the most prominent example.

Why do AI detectors fail on paraphrased text?

Humanizer tools rewrite AI output to mimic human burstiness and raise perplexity. This breaks statistical signatures that detectors rely on.

Are all AI detectors built the same way?

No. Some use statistical analysis, some use classifier models, some use watermarks. Most commercial detectors combine multiple methods.

What This Means for You

AI detectors are not magic. They are statistical tools with known failure modes. Understanding how ai detectors work allows institutions to deploy them appropriately and users to understand what flags their writing. The technology will improve. The arms race will continue. Neither detectors nor AI models will achieve permanent advantage.

If you are evaluating detection platforms, start with the data. The Global 100 publishes independent accuracy testing across 26 platforms. Compare performance on your specific use case. Academic writing behaves differently than marketing copy. Non-native authors require different thresholds than native speakers.