how do you spot ChatGPT writing?

ChatGPT is the highest-traffic consumer AI product on record. by February 2026 it served 900 million weekly active users and roughly 2.5 billion prompts per day. Claude, Gemini, and open-source contenders trade places on benchmarks version-by-version, but ChatGPT has the broadest cultural reach: when someone says “the AI wrote it,” they almost always mean ChatGPT.

what ChatGPT writing looks like

no single word identifies AI text. the tells are statistical and structural. the four most widely-documented signatures:

em-dash overuse. em-dash frequency in ChatGPT outputs rose from under 10% of responses to over 50% by mid-2025 per Rolling Stone reporting. the model picks the em-dash as a default rhythm device where a human writer might choose a comma, a period, or nothing. OpenAI shipped a fix in 2025 that lets users disable this via custom instruction, so the signal has weakened for power users who configure it but remains strong in default outputs.

vocabulary spikes. a Max Planck longitudinal PubMed analysis (Kobak et al., 2024) tracked specific words that saw greater than 50% post-ChatGPT frequency spikes in academic writing: delve, tapestry, underscore, meticulous, commendable, showcase, intricate, symphony, realm, boast, surpass, unlocking, pivotal, robust. these are normal English words. what flags AI writing is their unusual density, not their presence.

structural fingerprints. heavy use of bulleted lists where prose would do, “It's important to note...” openings, “In conclusion...” closings, three-part parallel constructions, hedge phrases (“while it's true that... it's also worth considering...”), both-sides framing on questions that don't require it.

low burstiness. human writing has variance: easy sentences cluster, hard sentences cluster. ChatGPT writing is rhythmically flat; most sentences sit in a similar difficulty band. this is the core signal AI-text detectors exploit, and it's the same property that gets ESL writers falsely flagged: non-native English is also flat for different reasons.

how amige. detects ChatGPT

amige.'s text detector measures statistical regularities: per-token perplexity (how predictable each word was), variance of perplexity across the document (burstiness), and the characteristic distributional fingerprint of large language model output. one technique that has performed strongly in recent literature is Binoculars (Hans et al., ICML 2024), which computes the ratio of log-perplexities from two reference language models. in the paper's own tests, Binoculars reported greater than 90% true-positive rate on ChatGPT outputs at 0.01% false-positive rate without any ChatGPT-specific training. amige. doesn't publish an accuracy figure of its own.

a trained model first reads the marks in the text and routes each scan to the classifiers strongest for it; that routing hint never enters the verdict. then a panel of independent detectors built by different teams picks up these signals, their reads are fused and calibrated, and the confidence is capped so it never lands on a flat 0% or 100%. the per-detector breakdown lets you see when the classifiers agree and when they're in the uncertain middle band, where amige. abstains rather than guess. amige. recalibrates as new generators ship. very short text scans are too thin a signal for any statistical classifier to give a defensible answer; amige. won't accept scans below the floor the active text models require.

the detection difficulty curve

ChatGPT text detection has gotten harder with each major model release. detectors that worked well on GPT-3.5 outputs reportedly were far less reliable on GPT-4; that pattern has continued. real-world accuracy on edited, human-revised, or paraphraser-processed text drops to the 40-80% range in independent benchmarks. a recent NeurIPS 2025 paper measured an average 88% drop in true-positive rate across all detectors after a single paraphrasing pass.

OpenAI has reportedly built but not shipped a server-side cryptographic watermark for ChatGPT text (Aaronson's scheme biases token sampling toward a keyed pseudorandom pattern). independent benchmarks on shipped watermarks in other models (Llama 3.1, Claude 3.5) push detection accuracy above 90% when the watermark is present. as of May 2026 there's no public confirmation that watermarking is enabled on ChatGPT outputs; treat watermark-based detection as research rather than production reality.

the false-positive risk to know about

the key caveat for any ChatGPT text verdict is the non-native English speaker bias. a 2023 Stanford / Cell Patterns study (Liang et al.) found AI text detectors unanimously misidentified roughly 20% of essays written by non-native English speakers as AI-generated. detectors have improved on the original test set since then, but the structural reason persists: non-native English tends to have lower burstiness for ESL reasons unrelated to AI, and burstiness is the signal detectors rely on most.

if you're a teacher, an admissions officer, or anyone in a position to make a consequential call on someone's writing, please don't use any AI detector's output (ours included) as a sole accusation. it is structurally biased against a population of writers who do nothing wrong. read more on this in the machine.