How the AI Detector Works

A Different Approach to AI Detection

Most AI detectors work by calling a commercial AI API and asking "is this AI-generated?" That approach is expensive, slow, and circular — you're asking an AI to detect another AI. It also leaks your text to third-party servers, introduces latency, and costs money per query.

Our detector works differently. Every analysis runs entirely in your browser using pure linguistic pattern analysis derived from published academic research. No text leaves your device. No API is called. The algorithms run in JavaScript in under a second, no matter the text length. The science comes from peer-reviewed research on how AI writing statistically differs from human writing at the vocabulary, syntax, and structural level.

The Two Detection Modes

AI writing looks very different depending on whether it's being used for formal content or creative content. A ChatGPT-written essay has completely different patterns from a Claude-written short story. Our detector runs two parallel analysis engines simultaneously on every text:

Academic / Formal AI Mode

Designed to catch AI writing in essays, emails, reports, academic papers, business documents, cover letters, and professional communications. This mode looks for the elevated vocabulary, structural predictability, and transition overuse that characterize AI in formal contexts. It is the most accurate mode for most content people need to check — the majority of AI-assisted writing occurs in formal settings.

Narrative / Creative AI Mode

Designed to catch AI writing in stories, fiction, social media posts, creative writing, blog posts, and narrative content. Creative AI has completely different tells — it doesn't use "furthermore" and "in conclusion", but it does repeat sentence subjects obsessively, loop the same phrases, and produce suspiciously uniform sentence lengths in a different range. This mode catches what Academic mode misses in creative contexts.

Both modes produce a weighted score from 0–100. The final score combines both, weighted toward whichever mode shows stronger signal. If a text is half essay and half story, both modes contribute proportionally.

The 6 Academic AI Signals

Each signal is independently scored and contributes a weighted percentage to the Academic AI score.

1. AI Vocabulary Signal (weight: 30%)

This is the single most reliable signal for formal AI writing, and it's rooted in the most significant peer-reviewed study on AI detection to date. Kobak et al. published "Delving into ChatGPT usage in academic writing through excess vocabulary" in Science Advances (2025), analyzing 14 million scientific abstracts. They found that after ChatGPT's release in November 2022, specific words appeared in academic writing at rates 5–28× higher than before — without any corresponding real-world reason for the increase.

The flagged words include: delve, meticulous, nuanced, pivotal, tapestry, testament, commendable, intricate, multifaceted, comprehensive, showcase, underscores, paramount, groundbreaking, and 60 others. We check the density of all 74 documented excess-vocabulary words per 100 words of text. High density = strong AI signal.

2. Significance Phrases Signal (weight: 15%)

AI text constantly reaches for phrases that sound meaningful but say very little: "it's worth noting that", "marks a pivotal moment", "in today's rapidly evolving landscape", "it is important to consider", "plays a crucial role in". These phrases are linguistic padding — they fill space while signaling importance without conveying specific information. Humans writing naturally almost never produce these phrasings at the rate AI does. We check for a library of 40+ such significance-signaling phrases.

3. Burstiness Signal (weight: 20%)

Burstiness is the most mathematically rigorous signal. Human writing has what linguists call "burstiness" — dramatic variation in sentence length. A human might write a 3-word sentence, then a 45-word sentence, then two 12-word sentences, then a fragment. AI produces sentences in a remarkably narrow range, typically 15–28 words each, with very little variation.

We quantify this by measuring the Coefficient of Variation (CV) of sentence lengths: standard deviation divided by mean. Human writing typically produces a CV of 0.5–1.2. AI writing typically produces a CV of 0.2–0.45. A low CV is a strong AI indicator. This methodology was pioneered by Edward Tian (Princeton, 2023) in the GPTZero detector and has since been validated by multiple independent studies.

4. Not-But Contrast Pattern Signal (weight: 10%)

AI writing has an unusual affinity for contrastive sentence structures. It constantly says "not X, but Y", uses "while X, Y" constructions, and peppers text with adversative conjunctions: "however,", "nevertheless,", "on the other hand,", "conversely,". This reflects AI's training to present "balanced" perspectives on every topic — it has been trained to show multiple sides, which manifests as constant contrast signaling. Human writers use these constructions but not at AI's characteristic frequency.

5. Closing Ritual Signal (weight: 15%)

Perhaps the most stereotyped AI behavior: almost every AI-generated piece ends with a formal concluding paragraph beginning with "In conclusion,", "In summary,", "To summarize,", "To conclude,", or "Overall, it is clear that." Real human writers almost never end an email, a quick article, or a blog post with a formal conclusion statement — that's a convention from formal academic essays that AI applies indiscriminately to all writing types. The presence of a closing ritual phrase is a very high-precision signal (low false positive rate).

6. Transition Overuse Signal (weight: 10%)

AI uses paragraph-linking transitions at a rate of 3–5 per 100 words. Human writers, even professional ones, use them far less frequently and with more variety. AI nearly always uses the same set: "Furthermore,", "Moreover,", "Additionally,", "Consequently,", "Thus,", "Therefore,". Seeing multiple of these in a single paragraph — especially at the start of consecutive paragraphs — is a strong AI indicator. We measure the per-100-word transition density and compare it to human baselines.

The 6 Narrative AI Signals

Creative AI writing has its own distinct fingerprints, unrelated to the formal vocabulary signals above.

1. Phrase Loop Signal (weight: 20%)

AI narrative writing repeats 3–4 word phrases across paragraphs at rates no human writer naturally produces. A human author writes a draft in continuous forward motion; an AI generates each paragraph somewhat independently, causing the same phrases to recur. We detect this by computing n-gram repetition ratios — counting how often 3-gram and 4-gram phrases appear more than once relative to total word count. High repetition ratios above a human-calibrated threshold indicate AI generation.

2. Subject Monotony Signal (weight: 15%)

AI narrative stories have a distinctive sentence-opening pattern: "She walked." "She felt." "She looked." "She couldn't help but." Human writers naturally vary their sentence structures — they use participial phrases, time clauses, dialogue, action beats, and inverted syntax to create rhythm. AI defaults to subject-verb openings with the same subject repeatedly. We measure what percentage of sentences begin with the same pronoun or article, and flag texts where one subject word-class dominates sentence openings beyond human norms.

3. Semantic Circularity Signal (weight: 15%)

AI romance, drama, and emotional fiction has a small vocabulary of emotional clichés it cycles through repeatedly: "heart raced", "breath caught", "couldn't help but", "felt a sense of", "a wave of emotion", "tears threatened to fall". These phrases appear so frequently in AI training data (from romance novels, fan fiction, and popular fiction) that the model learns to reach for them whenever generating emotional content. We count their occurrences per 100 words against a calibrated human baseline.

4. Duplicate Sentence Signal (weight: 20%)

AI generators sometimes produce sentences that are nearly identical in structure, vocabulary, or both. We calculate the Jaccard similarity between all sentence pairs in a text — the ratio of shared words to total unique words between two sentences. When any pair of sentences has a Jaccard similarity above 0.7, we flag it as a near-duplicate. This catches AI's tendency to restate ideas across different parts of a text rather than genuinely developing them.

5. Temporal Artifact Signal (weight: 15%)

AI fiction generates impossible simultaneous actions and overused moment-description clichés that are statistical artifacts of training data. Phrases like "his eyes met hers across the room", "she felt her heart skip a beat", "he couldn't help but notice", "she met his gaze" appear in AI fiction at rates that reveal training data contamination. These are AI fiction fingerprints — phrases so overrepresented in the romance and drama fiction that makes up AI training data that the model uses them compulsively.

6. Structural Monotony Signal (weight: 15%)

While Academic AI produces sentences in the 15–28 word range, Narrative AI produces sentences in a slightly shorter 12–24 word cluster. Human fiction writers use extreme sentence length variance — fragments for impact, long flowing sentences for description, medium sentences for action. AI narrative clusters in the middle range with much less deviation. This signal specifically targets the distribution shape of sentence lengths rather than just the average, detecting whether sentences are uniformly mid-length rather than naturally distributed.

How the Final Score Is Calculated

The scoring system was designed to reduce false positives while maintaining sensitivity to genuine AI signals:

1. Both modes run simultaneously on every submitted text, regardless of content type.
2. Each mode produces a weighted score from 0 to 1.0, combining its six signal weights.
3. A cluster boost of +15% applies if 4 or more signals in either mode fire above their individual thresholds. When multiple independent signals all fire together, the probability of AI generation increases non-linearly.
4. The final score combines both mode scores, weighted toward whichever mode shows the stronger signal. A 60/40 blend is used when both modes show moderate signal.
5. A word count penalty applies for texts under 50 words. Short texts don't provide enough data for reliable signal analysis, so confidence is scaled down proportionally.
6. The score is capped at 97%. We never claim 100% certainty. Any AI detector claiming absolute certainty is overstating what linguistic analysis can achieve. The 3% floor on "human" probability reflects the inherent uncertainty in all statistical text analysis.

Accuracy and Limitations

Honest accuracy assessment matters more than marketing claims. Here is what our detector can and cannot reliably do:

• ~70–80% accuracy on clearly AI-generated text (200+ words, unedited, formal register). This is the performance range for browser-based linguistic detectors without access to model APIs.
• False positives are possible for ESL (English as a Second Language) writers who tend toward formal vocabulary and uniform sentence structures. Very formal academic writers may also score higher than expected. Always consider the author's writing context.
• False negatives are possible for humanized AI — text run through AI rewriting tools like QuillBot, Undetectable.ai, or manual editing. Very short texts (under 100 words) also produce less reliable results due to insufficient signal data.
• Not suitable as sole evidence of AI use in academic integrity proceedings. AI detection scores should be one data point among many, not a verdict on their own.
• Always consider context alongside the score: writing history, the task requirements, the student or writer's known style, and whether the score is borderline (40–65%) or clearly high (75%+).

Research Citations

• Kobak, D., González-Márquez, R., Horvát, E-Á., & Lause, J. (2025). "Delving into ChatGPT usage in academic writing through excess vocabulary." Science Advances, 11(3). — The foundational peer-reviewed study identifying excess vocabulary as a ChatGPT signal across 14 million scientific abstracts.
• Tian, E. (2023). GPTZero: Towards Detection of ChatGPT Writing in Scientific Articles. Princeton University. — The original burstiness methodology for AI text detection, based on perplexity and sentence-length variation.
• Stanford HAI (2023). Writing pattern analysis of large language model outputs. Stanford Human-Centered AI Institute. — Structural pattern analysis of LLM writing across domains.
• Wikipedia AI Cleanup Project (2024–2025). Community-documented AI writing detection patterns. — Practitioner-validated list of AI writing tells used by Wikipedia editors to identify and flag AI-generated article content.

Explore Further

• Free AI Detector Tool — Try the detector yourself
• AI vs Human Writing Examples — See side-by-side comparisons of real AI and human writing
• ChatGPT Detector — Specialized detection for ChatGPT writing patterns
• What Is AI Writing? — Background on how AI writing is generated