A year ago you could often spot ChatGPT writing at a glance. Today it can read like a person wrote it. Editors, teachers and recruiters now receive work that may or may not be original, and they are all asking the same question: is this AI? Guessing is no longer safe, and checking is fast. This guide explains what an AI detector does, how it reaches a score, and how to read that score fairly.
Why AI detection matters
AI-generated text is everywhere now
Freelancers use ChatGPT to draft articles. Students use Claude on take-home essays. Job applicants use AI to write cover letters. Most of the time a reader has no reliable way to tell by eye. An AI detector gives you a number to work with, and a sentence-level AI checker shows you where to look.
Search engines penalise scaled, low-effort AI content
Google’s helpful-content systems are designed to demote pages built from low-effort AI text, and a site can lose traffic if too much of it looks machine-written. Running drafts through an AI detector before you publish helps protect your domain. Google’s scaled content abuse policy sets out the specifics.
Schools and clients want proof the work is yours
You may have written every word yourself, but if a professor or client runs your work through their own AI detector and it is flagged, you need a response. Run your own scan first. If it comes back clean, you have a record. If some sentences score high, you can revise them before you submit.
A plagiarism check will not catch AI
AI text is technically original because it was never published anywhere, so a normal plagiarism checker will report it as clean. An AI detector is built for exactly this gap. Run both for a full check, or use our plagiarism checker, which includes AI detection in the same report.
How AI detectors work
AI detectors reverse the logic of AI generation. A language model writes by repeatedly choosing a highly probable next word. So text in which almost every word is the probable choice is a signal of machine authorship. Detectors combine several measurements of that idea.
Perplexity: how predictable the language is
Perplexity measures how surprised a language model is by each word. AI text leans on statistically expected vocabulary, which makes it unusually smooth, or low in perplexity. Human writing is messier and more idiosyncratic. A detector measures this across the text and flags passages that are too predictable to be natural.
Burstiness: rhythm and variation
People vary sentence length naturally: short and punchy, then long and complex. AI output tends toward a more uniform rhythm. Burstiness analysis looks at that variance and catches writing that feels polished but strangely flat.
Structure and semantic patterns
AI writing is locally coherent but often globally template-like: a broad introduction, three to five points and a diplomatic conclusion. Detectors look for that predictable macro-structure, but only count it when it appears together with other signals, which keeps false positives down.
Stylistic consistency
Human writing has shifts in voice, small quirks and the occasional error. Text produced from a single prompt is stylistically flat in ways a trained classifier can pick up.
Why the best detectors combine signals
None of these signals is reliable alone. The most accurate systems use ensemble models, several classifiers weighing the signals against each other, and return a calibrated probability rather than a yes-or-no verdict. Scoring each sentence separately, instead of giving one number for the whole document, shows where AI content is concentrated and where human writing begins.
What an AI detector can spot
AI writing is not one thing. Different models leave different fingerprints, so a detector trained only on ChatGPT will miss much of the rest. The Plagiarism Checker Plus AI detector is trained on output from the major language models and retrained as they change.
- ChatGPT (GPT-4o, GPT-5). The most common source of AI text online. Patterns from each public GPT release are tracked, including the latest models.
- Claude. Claude’s style differs from ChatGPT’s: calmer, more hedged and more nuanced. A good ChatGPT detector is not automatically a good Claude detector, so it needs its own training data.
- Gemini. Gemini text has its own structural tells, including a particular cadence, phrasing patterns and hedging vocabulary.
- Llama, Mistral and other open-source models. Popular with developers building wrappers and content farms, and harder to detect because their output varies more.
- Mixed and lightly edited AI. A common pattern is an AI first draft that a person then edits. Because each sentence is scored on its own, partial AI still shows up after a human has cleaned it up.
- Humanized or paraphrased AI. Humanizer tools rewrite AI text to sound less robotic. They can lower a score but rarely remove it, because the original AI structure and the tool’s reshuffling both leave traces.
Accuracy and limits
No AI detector is 100% accurate. False positives happen, most often on short texts, formal academic writing, or writing that happens to share AI-like patterns. False negatives happen too, especially on AI text that has been heavily edited. That is why useful reports show a confidence for each flagged sentence. A low-confidence flag on a single sentence is rarely worth worrying about; a high-confidence flag across a whole paragraph usually is.
Detection is the first step to an honest conversation
AI detection does not replace human judgment, and it should never be the sole basis for an academic-integrity decision. What it does is identify content that warrants a closer look, and give that judgment something to stand on. Compare the flagged passages with the writer’s other work, ask about their drafts and process, and follow your institution’s policy. Most university academic-integrity offices recommend exactly this approach.
Who uses an AI checker
Teachers and academics
Reading dozens of essays a week leaves no time to question every one. Running them through an AI detector first shows which submissions need a closer read and which do not.
Editors and publishers
Even trusted writers’ drafts need verifying. One scan shows whether a draft reads as hand-written or AI-generated, so you can catch it before it ships rather than after.
SEO and content teams
Teams that publish at scale run every piece through an AI checker before it goes into the CMS, to keep low-effort AI text off the site.
Students and writers
If your professor or client might run your work through their own AI detector, run yours first. If any sentences are flagged, you can tighten them before you submit.
Recruiters and hiring managers
AI-written cover letters are hard to spot at a glance. A quick AI check on shortlisted candidates shows which letters deserve a follow-up question.
AI detection at scale
Universities, schools and large content teams need detection that fits their volume and their workflow. For a single document, the free AI detector is enough. For a whole institution, the pieces that matter are:
- Roles and oversight. Admins, managers, teachers and students work in one workspace, so coordinators can manage access and review reports across courses.
- Bulk and automatic scanning. Upload a full class set or a batch of articles at once and get an individual report for each, filed in folders.
- A private repository. Combine AI detection with plagiarism checking against your own archive of past submissions, kept private to your institution.
- Your own login page. Staff and students sign in on your institution’s own subdomain.
- API access. Paid plans include a REST API, with keys you create yourself in Settings, so you can send text from your own systems and get per-sentence AI results back.
See Plagiarism Checker Plus for institutions for plans and details, or read the API documentation.