What is AI detection? It is any method for estimating whether a piece of content, usually text but also images, audio or video, was produced by a generative AI system rather than a person. Most AI detection software is itself a machine learning model that outputs a probability, not a verdict, and the leading vendors say plainly that a score should never be the only basis for a decision. A second, newer meaning comes from cybersecurity, where "AI detection and response" describes tools that watch an organization's AI systems for attacks.
This guide explains the different things people mean by AI detection, the four technical approaches behind the tools, how to read an AI detection report, what Google Classroom does and does not check, and where the method breaks down. If you only want to know whether detectors can be trusted, our companion guide on whether AI detectors are accurate goes deep on the 2026 research.
What is AI detection, in one minute
AI detection answers a narrow question: does this content carry the statistical fingerprints of machine generation, or a mark that a generator deliberately left behind? It does not answer whether someone cheated, whether a fact is true, or whether a person used AI responsibly. Those are human judgments that a detector can inform but not replace.
In practice, the phrase covers three different jobs:
| Meaning | What it checks | Typical users | Example tools |
|---|---|---|---|
| AI content detection | Whether text was written by a language model | Teachers, editors, publishers | Turnitin, GPTZero, Pangram, Originality.ai |
| Synthetic media detection | Whether an image, voice or video is AI generated or manipulated | Journalists, trust and safety teams | SynthID Detector, Content Credentials, Hive |
| AI detection and response (AIDR) | Whether an organization's AI apps and agents are being attacked or misused | Security teams | Enterprise security platforms |
Most people searching the phrase mean the first one, so that is the focus here. The second is covered in detail in our deepfake detection guide, and the third gets its own section below.
How AI detection software works
Every AI detector relies on one of four ideas, or a mix of them. Knowing which one a tool uses tells you a lot about when to trust it.
1. Trained classifiers
The most common approach. A vendor collects large sets of human written and AI generated text, then trains a neural network to tell them apart. Given a new document, the model scores it, often sentence by sentence, and rolls those scores up into a document level percentage.
Turnitin is a good example because it documents its method publicly. Its model assigns each sentence a score between 0 and 1 based on the sentence in context, labels sentences above a threshold as AI, and reports the share of evaluated sentences that crossed the line. Turnitin says the system is transformer based and trained on its own corpus of academic writing, and that it is not explicitly programmed to measure single signals such as "perplexity" or "burstiness"; it learns many patterns at once, which is also why individual predictions are hard to explain (Turnitin FAQ). GPTZero, Pangram and Originality.ai are also trained classifiers, each tuned on different data for different audiences.
The strength of classifiers is accuracy on the kind of text they were trained on. The weakness is the flip side: new models, unusual genres and heavy editing can push text outside what the classifier has seen.
2. Zero shot statistical tests
These methods need no training set of AI text. They use an existing language model to measure how "surprising" a passage is. Language models tend to produce word sequences that other models find predictable, while human writing is more idiosyncratic.
Two research methods show how far this idea has come. DetectGPT, from Stanford researchers in 2023, compares a passage with slightly reworded versions of itself: machine text tends to sit at a local peak of the model's probability, so rewording lowers its score more than it does for human text. In the paper, it lifted detection of fake news articles from a 20 billion parameter model from 0.81 to 0.95 AUROC compared with the best earlier zero shot method (DetectGPT paper). Binoculars, published in 2024, compares how two closely related models see the same text, which corrects for prompts that make even machine text look surprising. Its authors report catching over 90 percent of ChatGPT samples across several document types at a false positive rate of 0.01 percent, without training on any ChatGPT data (Binoculars paper).
The caveat sits in the same paper: when the authors tested GPT-4 output in March 2024, Binoculars correctly flagged only about 42 percent of samples. Statistical tests age as models change.
3. Watermarks
Instead of guessing after the fact, a generator can deliberately mark its own output. Google DeepMind's SynthID Text adjusts the probabilities a model uses to pick each next word, leaving a pattern invisible to readers but testable by a detector. Its Nature paper reports a live experiment across nearly 20 million Gemini responses that found no loss in text quality, and the method is open source in Hugging Face Transformers (Nature; Google AI for Developers). Google's documentation says the detector returns one of three states: watermarked, not watermarked or uncertain.
Watermarks are strong evidence when present, but they only exist if the generator chose to add them, and a detector can only read its own scheme. A negative result says nothing about text from other tools. For images, audio and video, our SynthID Detector guide covers Google's public checking portal. Watermarks on content are also different from AI model fingerprinting, which marks a model's behavior to prove who owns the model. For OpenAI's text watermark, which ChatGPT applies in the EU, see ChatGPT watermark: how textGrain works.
4. Provenance metadata
The fourth approach records where content came from rather than analyzing it. The C2PA standard, shown to consumers as Content Credentials, attaches a signed history to a file: which camera or app created it, and whether AI tools were used. It works for media far more than for plain text, and it disappears when metadata is stripped by a screenshot or reupload.
What is an AI detection report?
An AI detection report is the screen a reviewer sees after a scan. Formats differ, but most contain the same parts, and Turnitin's version shows each clearly (Turnitin FAQ):
- An overall percentage. For Turnitin, this is the share of qualifying prose sentences predicted to be AI generated or AI paraphrased, not the share of the whole file. Lists, headers, references, tables and code are left out.
- Highlighted passages. The segments the model flagged, so a reviewer can read them in context.
- Low score handling. Turnitin shows scores from 1 to 19 percent only as an asterisk, with no number and no highlights, to limit false positives.
- Processing limits. For English, a document needs at least 300 words of prose and no more than 30,000 words, or no score appears.
- Audience. Turnitin's FAQ says the indicator is visible only to instructors and administrators, though instructors can download the report as a PDF to share, and Turnitin's newer report guide describes instructors making the report available to students. For the full scoring pipeline, see how does Turnitin detect AI.
Two details matter for interpretation. First, Turnitin says it aims to keep false positives under 1 percent for documents with more than 20 percent AI writing, and accepts missing some AI text to get there; a document scored at 50 percent could contain as much as 65 percent AI writing. Second, on short documents of a few hundred words, the prediction tends to be all or nothing, so mixed writing can be flagged as entirely AI. A report is a map of where to look, not a finding.
How to use AI detection in Google Classroom
A common surprise: Google Classroom's built in originality reports are not AI detectors. Google's help pages describe them as comparing a student's Docs, Slides and Word files against web pages and books, linking detected sources and flagging uncited text, with an optional check against past submissions from the same school (Google Classroom Help). Google does not describe any check for AI generated writing. Schools on the free Education Fundamentals tier get originality reports on five assignments per class; unlimited reports need a paid upgrade.
To add AI detection to a Classroom workflow, schools generally use a third party integration:
- Turnitin's Google Classroom addon. It brings Turnitin Feedback Studio into Classroom assignments, and with a Turnitin Originality license instructors get the AI writing indicator (Turnitin press release).
- Detector browser extensions and LMS integrations. Vendors such as Pangram offer Chrome and Google Docs extensions so a teacher can check a document where it lives.
- Process evidence instead of, or alongside, a score. Google Docs version history often says more about authorship than any detector, because it shows how a draft grew over time.
Whatever you add, set a policy first: who sees the score, what threshold prompts a conversation, and what evidence beyond the score is required before any consequence.
What AI detection can and cannot tell you
AI detection is useful as triage. Modern detectors rarely flag ordinary, unedited human writing, and they can surface passages worth a closer read across a large stack of submissions. A peer reviewed study published in June 2026 found that four major detectors correctly classified every fully human paper in its sample, while differing widely in how much AI text they caught (International Journal for Educational Integrity).
The limits are just as real:
- Misses are common. Edited, mixed or "humanized" AI text often slips through, and results vary sharply between tools.
- Polished human writing can look machine like. Heavily edited, formulaic or nonnative English prose has a history of false positives.
- Detectors age. A tool trained before a new model was released may not recognize that model's style. Styles shift between versions too: in the 2026 Claudish study, Opus 5.5 scored lower than Opus 5 on 19 of 21 measures.
- Grammar help is not the same as generation. Turnitin says grammar and spelling fixes from tools like Grammarly were mostly not flagged in its tests, but text from Grammarly's generative features likely will be.
- The earliest tools were weak. OpenAI withdrew its own AI text classifier in 2023 for low accuracy, a reminder that a confident looking number can still be wrong.
If you want a short list of tools and their free limits, see our best free AI detector guide; for head to heads, see GPTZero vs Turnitin and Pangram vs GPTZero.
What is AI detection and response?
In security, AI detection and response, often shortened to AIDR, means something else entirely. CrowdStrike, which markets an AIDR product, defines it as a runtime security category that detects, investigates and responds to threats targeting and originating from AI systems, covering models, prompts, agents and the data pipelines behind them (CrowdStrike). Here the AI is the thing being protected, not the thing being detected. Typical alerts are a jailbreak attempt, sensitive data leaking into a prompt or an agent calling tools it should not.
If that is the meaning you need, the closest topics on this site are prompt injection attacks, LLM red teaming tools for testing before launch, and AI agent observability for monitoring after it.
Who should use AI detection
- Teachers and schools: as one input inside a written policy, alongside draft history and a conversation with the student.
- Editors and publishers: to screen submissions at volume, then review flagged pieces by hand.
- Trust and safety teams: with classifier APIs such as Hive's (our deepfake detection guide explains what Hive checks), evaluated on a sample of your own content first.
- Writers worried about being flagged: keep drafts and version history; that evidence of process is your best defense.
From Ahzy Capital (our own products, labeled as such): Get the free AI Agent Cheat Sheet: 24 copy and paste AI agent setups for 8 types of small business, with prompts and workflows. We make and sell these, so they aren't independent picks.
FAQ
What is AI detection software?
It is software that estimates whether text or media was produced by generative AI. Most tools are trained classifiers that output a probability; others test for statistical patterns, read watermarks a generator embedded, or check provenance metadata such as Content Credentials.
Is AI detection the same as a plagiarism check?
No. A plagiarism check looks for text that matches existing sources. AI detection looks for patterns typical of machine generation, so AI written text can be completely original and still be flagged. Turnitin keeps the two scores separate, and they do not influence each other.
Does Google Classroom have AI detection built in?
Not according to Google's documentation. Classroom's originality reports compare work against web pages, books and, optionally, past school submissions to flag uncited text. AI detection requires an addon such as Turnitin's Google Classroom integration or a vendor extension.
Can AI detection be wrong?
Yes. Detectors miss much edited or mixed AI text and can misread polished human writing. Vendors including Turnitin state that a score should not be the sole basis for action. Treat any result as a reason to look closer.
What is AI detection and response?
It is a cybersecurity category, not a writing check. AIDR tools monitor an organization's AI apps, prompts and agents for attacks and misuse, such as jailbreak attempts or data leaks, and can block or alert in real time.
