How to Tell If Text or Images Are AI-Generated (and Why AI Detectors Are Not Proof)

Short answer: for text, there is no reliable way to prove a passage was written by AI. Detector tools give a guess, they flag human writing as AI, and they can be fooled with light editing, so a score should start a conversation, never end one. For images, video and audio, you have better options: some AI tools now embed invisible watermarks or attach "Content Credentials" that free checkers can read. A missing signal still proves nothing. This guide explains what each method can and cannot tell you, and gives a fair step-by-step process.

Before you accuse anyone: do not treat a detector score as evidence on its own. Even the companies that make detectors say so. If you are a teacher or manager, follow your school's or employer's policy, and do not paste student or employee work into unapproved tools (see our student privacy and FERPA basics).

Why AI text is so hard to detect

An AI chatbot writes by predicting likely next words (our guide to how AI chatbots work explains this). Most text detectors look for the same thing from the other side: writing that is very predictable and evenly structured. The problem is that plenty of people write that way too, especially in formal, formulaic or second-language writing. And AI text becomes harder to spot as soon as a person edits it or asks the chatbot to vary its style.

The evidence on this is clear:

• OpenAI withdrew its own detector. OpenAI launched an AI text classifier in January 2023 and reported that, on its test set, it correctly flagged only 26% of AI-written text as "likely AI-written" while wrongly labeling human text as AI-written 9% of the time. It was withdrawn on July 20, 2023 "due to its low rate of accuracy" (OpenAI's note on the classifier).
• Detectors can be biased against non-native English writers. A Stanford-led study published in the journal Patterns (2023) ran seven widely used detectors on 91 essays written by non-native English speakers for the TOEFL exam. On average, the detectors labeled more than half of these human-written essays as AI-generated, while classifying essays by U.S. eighth graders almost perfectly (Liang et al., "GPT detectors are biased against non-native English writers").
• Accuracy claims are not always true. In 2025 the U.S. Federal Trade Commission finalized an order against Workado, a company that advertised its AI Content Detector as "98 percent" accurate. The FTC's complaint said independent testing found about 53 percent accuracy on general-purpose content (FTC press release, April 2025; final order, August 2025).
• Detector makers warn against relying on the score. Turnitin's own guide says its AI writing detection "may not always be accurate" and "should not be used as the sole basis for adverse actions against a student" (Turnitin: Using the AI Writing Report).

What about watermarks in AI text?

Some AI companies are starting to add hidden statistical watermarks to the text their models write. For example, OpenAI says ChatGPT text generated for users in the EU now carries an invisible watermark to comply with the EU AI Act, but access to its text detector is limited to approved research organizations, and checking student work is outside that program (OpenAI Help: provenance signals). OpenAI also notes that short passages, code and heavily rewritten or translated text may not carry a detectable watermark. In practice, ordinary readers cannot check text watermarks today, and text from other AI tools may have none.

Clues in text: useful, but not proof

These signs are worth noticing. None of them proves AI use, and a careful writer can produce any of them:

• Made-up facts, quotes or sources. Citations that do not exist, or that do not say what the text claims, are the strongest practical clue. Our guide to checking AI answers for mistakes shows how to verify sources quickly.
• Leftover chatbot phrases. Lines such as "As an AI language model" or "Here is a revised version of your essay" pasted in by accident.
• A sudden change from the person's usual work in vocabulary, structure or knowledge, compared with earlier writing you know is theirs.
• Generic content that never touches specifics the writer should know, such as what was said in class, a real customer's situation or a local detail.
• Uniform, polished but vague paragraphs with the same rhythm throughout. This one is weak: many good writers and editors produce exactly this.

How to check an image, video or audio file

Here the tools are more useful, because some AI generators label their output when they create it. There are two main kinds of signal:

• Content Credentials (C2PA): signed information attached to a file that records where it came from and how it was edited. It is an open standard from the Coalition for Content Provenance and Authenticity, used by AI tools and also by some cameras, editing apps and news organizations. Its weakness is that it is metadata: screenshots, file conversion and many social media uploads can strip it.
• Invisible watermarks such as SynthID: a signal built into the pixels or sound itself, developed by Google DeepMind. It is designed to survive common edits like resizing or compression, though heavy changes can still remove it.

Free ways to check, according to the companies' own help pages (October 2026):

1. Google Gemini: signed in to the Gemini app or gemini.google.com, upload a single image, video or audio file and ask "Was this created or edited by Google AI?" Gemini checks for SynthID and, where supported, Content Credentials. Google notes it can only recognize SynthID from Google's own AI tools, so "no watermark found" does not mean the file is real (Gemini Apps Help: verify AI-generated content).
2. OpenAI: OpenAI says images made with ChatGPT and its API carry both Content Credentials and a SynthID watermark, and its checker at openai.com/verify looks for those signals. It is designed to detect only OpenAI's signals, not content from other AI tools (OpenAI Help: provenance signals).
3. Use the original file when you can. A screenshot or a re-saved download is much less likely to keep its credentials.

Then do the old-fashioned checks, which still catch a lot:

• Reverse image search (for example with Google Images or Google Lens) to find where the picture first appeared and whether a news outlet or the original photographer has it.
• Look closely at details AI often gets wrong: garbled text on signs or labels, hands and teeth, jewelry, reflections, shadows that point different ways, and repeated background patterns. Newer generators make fewer of these mistakes, so a clean image is not proof of a real one.
• Check the context: who posted it first, when, and whether trustworthy sources report the same event.

If you make images yourself, our beginner's guide to AI image generators covers copyright, licensing and platform rules for images you make yourself.

A fair process when you suspect AI-written work

1. Check the rules first. Was AI use actually banned for this task? Many schools and workplaces allow some uses, such as brainstorming or grammar help.
2. Look for evidence you can verify: fake citations, factual errors, content copied from elsewhere, or leftover chatbot text.
3. Compare with known work by the same person, written in class, in a meeting or on an earlier task.
4. Ask about the process. Request drafts, notes or version history (Google Docs and Microsoft Word keep it), and have a calm conversation: "Walk me through how you wrote this section." People who did the work can usually explain it.
5. Treat any detector score as one weak signal, not the decision. Never act on a score alone.
6. Design it out next time. For teachers: in-class drafting, oral check-ins and assignments tied to class discussion reduce the problem more than detectors do. Our AI for Teachers hub has policy-aware ideas.

If someone says your work is AI-generated

False flags happen, so stay calm and gather your evidence: version history, drafts, notes, browser history for your research, and earlier writing samples. Offer to explain your work or rewrite a section in person. You can point out that detector makers themselves say the score should not be the only basis for a decision. If you used AI in a permitted way, say exactly how. For students, our guide to AI tools for students shows ways to use AI for studying and research that keep the thinking and writing yours.

A prompt to help you review a suspicious source

An AI chatbot cannot tell you whether text is AI-written, so do not ask it that. It can help you list claims to verify:

Here is a passage I need to fact-check. Do not tell me whether it was written by AI. Instead: 1) list every factual claim, number, quote and source it contains, 2) for each, say what kind of source would confirm it (official site, original study, news report), and 3) flag any citation that looks incomplete or unusual. Passage: [PASTE TEXT WITHOUT PERSONAL INFORMATION]

Then check those claims yourself in the original sources. New to prompting? See how to write better AI prompts.

Key takeaways

• No text detector is reliable enough to be proof; they produce false positives, including against non-native English writers.
• Verifiable evidence (fake sources, factual errors, drafts and version history, a conversation) matters more than a score.
• For images, video and audio, check Content Credentials and watermarks with the free tools from Google and OpenAI, plus a reverse image search. "No signal found" does not mean "real".
• Follow your school's or employer's AI policy and keep personal information out of third-party tools.

Frequently asked questions

Can AI detectors tell for sure if text was written by ChatGPT?

No. Detectors estimate how likely text is to be AI-written. They miss AI text, especially after editing, and they flag human writing as AI. Even Turnitin says its result should not be the sole basis for action against a student.

Can I just ask ChatGPT whether it wrote something?

No. Chatbots do not keep a record of everything they or other users have generated, and they may answer confidently either way. Their answer is not evidence.

How can I check if an image was made with AI?

Upload the original file to a provenance checker such as Google's Gemini app (for Google AI content) or OpenAI's verify tool (for OpenAI content), run a reverse image search, and look for visual errors and the original source. A file with no AI signal can still be AI-generated, because metadata is easily stripped and not every tool adds a watermark.

Why do detectors flag non-native English writers?

Many detectors treat simple, predictable wording as a sign of AI. Writers using a second language often choose more common words and simpler structures, so their genuine work can look "predictable" to the detector. A 2023 study in the journal Patterns found this bias across seven popular detectors.

What should a teacher do instead of relying on a detector?

Set a clear AI policy for each assignment, look for verifiable problems such as fake citations, compare with in-class work, ask for drafts or version history, and talk with the student. Follow your school's academic integrity process.