Everyone wants a single number that answers 'did a person write this.' It's an understandable ask — and it's a harder question than the clean percentage on a detector's results page suggests. This is a plain explanation of how three widely used AI-detection tools, GPTZero, Originality.ai, and Copyleaks, actually work under the hood, based on what each company publishes about its own methodology, plus the independent research on where these tools reliably break down. It ties into Citeya's own AI Detection feature, but the goal here is an honest industry explainer, not a pitch — the limitations below apply to every detector on the market, ours included.
The Two Techniques Underneath Almost Every Detector
Most AI-text detectors lean on one or both of two underlying techniques. The first is perplexity and burstiness analysis: perplexity measures how 'surprised' a language model is by the next word in a sequence, and burstiness measures how much sentence length and structure vary across a passage. Human writing tends to be less predictable word-to-word and more variable sentence-to-sentence — we ramble, we go short then long, we make idiosyncratic word choices. Early AI-generated text tended toward the opposite: smoother, more statistically predictable, more uniform in rhythm. Low perplexity and low burstiness became an early proxy signal for 'machine-generated.'
The second technique is a trained classifier: a supervised machine-learning model fed large labeled datasets of known-human and known-AI text, learning to separate the two the way a spam filter learns to separate spam from real mail. This approach doesn't rely on a hand-picked metric like perplexity — it learns whatever statistical patterns best distinguish the two classes in its training data, which can be more accurate but is also a black box, and only as good as how well its training data resembles the text it's later asked to judge.
GPTZero: Perplexity and Burstiness, With Classifier Layers Added Over Time
GPTZero built its name on perplexity and burstiness — that was the founding technical premise of the tool, originally aimed at flagging AI-written student essays. Per GPTZero's own technology page and FAQ, the product has since layered classifier-based models on top of those original statistical metrics, reflecting a broader industry shift: as language models got better at producing varied, less 'smooth' text, perplexity and burstiness alone became weaker signals, and a trained classifier became necessary to catch what those metrics miss.
Originality.ai: A Classifier Tuned Specifically for Paraphrase Resistance
Originality.ai markets itself squarely at publishers and SEO teams, and its own explanation of how its AI content detection works describes an internally built classifier — a modified BERT/RoBERTa-style supervised model — trained on labeled samples generated across GPT, Claude, Gemini, and open-source models, with manual human review of the training data. It outputs a 0–100% probability score with per-sentence highlighting, and bundles a plagiarism check alongside the AI score. Its own guidance on choosing an AI detector is notably candid that paraphrased or 'humanized' AI text is harder to catch than raw model output, which is precisely the failure mode a classifier trained mostly on unedited generations would be expected to have.
Copyleaks: Enterprise Positioning, Multi-Model Attribution Claims
Copyleaks positions its AI content detector toward enterprise, education, and publishing customers who need audit trails and integration into existing plagiarism-checking workflows. Architecturally it follows the same broad pattern as the others — a trained classifier evaluating statistical patterns in text — with added claims around identifying which model family a passage may have come from and lower false-positive rates than earlier-generation detectors. As with any vendor's own accuracy claims, these are worth treating as a starting point for evaluation, not a substitute for it.
The Practical Differences, Side by Side
Stripped of marketing language, the three tools differ more in packaging and target customer than in core technique. GPTZero leans on its original perplexity-and-burstiness framing as an explainable, education-oriented product, positioning itself toward teachers and academic integrity use cases where a sentence-level explanation of why something was flagged matters as much as the score itself. Originality.ai leans hardest into raw classification accuracy and paraphrase resistance, positioned squarely at content teams and SEO agencies who need to screen large volumes of purchased or outsourced writing before it publishes. Copyleaks leans into enterprise integration and audit trails, positioned at organizations that need the AI check bundled with existing plagiarism-detection infrastructure and compliance reporting. None of the three has published independent, peer-reviewed accuracy benchmarks that would let you rank them cleanly against each other — the accuracy numbers you'll see cited are almost always the vendor's own, on the vendor's own test set.
Where All Three Actually Struggle
This is the part vendors talk about less than they should. The most serious, well-documented failure mode across AI detectors generally is a bias against non-native English writers. Research out of Stanford HAI found that widely used GPT detectors misclassified a meaningful share of TOEFL essays written by non-native English speakers as AI-generated, apparently because non-native writing patterns are often more formulaic and lower-perplexity in ways that overlap with what these tools flag as machine-like. That's not a minor edge case — it's a structural blind spot with real consequences for students and professionals writing in a second language.
The second honest limitation: detectors get less reliable as the underlying models improve. As generation models are tuned to sound more varied and less mechanically smooth, the statistical gap between human and AI text that perplexity, burstiness, and even trained classifiers rely on keeps narrowing. It's telling that OpenAI — with direct access to its own models' output distributions — shut down its own AI text classifier in 2023 citing a low rate of accuracy, rather than continuing to maintain a tool it couldn't stand behind. Detection here isn't a solved problem that vendors are merely refining; it's a moving target where the thing being detected keeps adapting.
Light paraphrasing or 'humanizing' passes also defeat most detectors with only minor edits, since they specifically disrupt the smoothness and predictability signals the models are trained to catch, without meaningfully changing the underlying content.
There's a second-order consequence worth naming directly: because non-native English writers and heavily edited human writing both tend to score as 'more AI-like' on these tools, a policy that auto-rejects or penalizes based on a raw detector score doesn't just risk false positives — it risks a specific pattern of false positives that falls disproportionately on the people least equipped to contest it. Any educator, publisher, or employer relying on these scores should treat that as a known, documented property of the technology, not a rare edge case.
So What Are These Tools Actually Good For?
Used as one signal among several — a prompt for a closer read, a spot-check before publishing, a data point in a larger editorial conversation — these tools have real value. Used as a sole, deterministic verdict that a specific person did or didn't write something, especially in contexts with real consequences like academic discipline or employment, they're operating well past what the underlying technology can responsibly support. The false-positive rate on non-native English writing alone should rule out any policy that treats a percentage score as proof.
Citeya's own AI Detection feature exists for the same reason every tool in this space exists: to give writers and editors a signal on their own drafts before they publish. We describe it the same way we'd describe any other detector here — a probabilistic estimate worth factoring into a decision, not a verdict that replaces one.
No AI detector on the market today can give you certainty. The honest ones tell you that; the honest use of them treats the score as a prompt to look closer, not a verdict.