Short answer: asking a chatbot “is this real?” is not a reliable check on an image or video. In one audit of Grok’s replies to viral World Cup posts it was right on image and video claims only 49.4% of the time, and it often missed AI-made or edited media. That sample was narrow, but it matches what verification experts advise: use a chatbot to describe what is visible, then check the source, the original file and any provenance data.
Last checked 12 November 2026. General information, not legal advice. TimeBinder does not detect AI or prevent fraud, and a dated record is not proof of authorship, ownership or originality.
Table of Contents
Why people ask a chatbot
On X, replying “@grok is this real?” to a post has become a common reflex, and other assistants are used the same way in private chats. It is fast, it sounds confident, and it feels like a neutral third party. The problem is that confidence is not evidence.
What one audit found
Fundación Maldita.es and the University of Jaén published an audit on 27 October 2026 of 766 queries to Grok about 46 viral claims during the 2026 World Cup. Its main findings:
- Grok answered fewer than four in ten queries. Of the 317 responses about veracity that were analysed, 59.9% were correct, 26.8% incorrect and 13.2% partly correct.
- Format mattered most. On images and videos it was right in 121 of 245 responses (49.4%). On written claims, including screenshots, it was right in 69 of 72 (95.8%).
- The most common error was failing to notice that media was edited or AI-generated: 81 of the 127 incorrect or partly correct responses (63.8%). It called an AI-made video of a dog in a team shirt real.
- It often contradicted itself on the same content without new information, and almost never named its sources.
The authors describe a black-box audit of one event, one chatbot and one period. The figures should not be read as a general accuracy score for every model or task. Other assistants may do better or worse, and models change.
Factnameh reported a different failure from the Iran war: users asked Grok to identify videos of the 28 February 2026 strike on a school in Minab, and it replied that they came from a 2014 school attack in Peshawar, Pakistan, when the footage was from Minab. Those were real images; the error ran the other way, which is why neither “it says real” nor “it says fake” settles anything.
It is the same limit as AI detectors
At the US judiciary’s evidence-rules committee meeting on 15 October 2026, according to a Reuters report carried by Investing.com, an OpenAI policy lawyer said she did not expect a scientific way to detect AI imagery in future, and Meta’s representative said recent AI media is hard for people to tell apart (see also our piece on what to do if the other side says your evidence is a deepfake). A chatbot judging pixels is a detector with a chat interface. Our post on working drafts and timestamps versus AI detectors covers why scores are not proof.
What a chatbot can reasonably do
Poynter’s Alex Mahadevan, quoted by Factnameh, advises narrow prompts about what is directly visible, such as “describe only what you can see”, “read any visible text” and “what is too blurry or cropped to identify?”, and warns against “is this real?”, which pushes a model to guess. In other words, use it as a describer and note-taker, never as the judge.
Checks that do more than a guess
- Find the earliest source. Who posted it first, and where? Reverse image search tools can help; none is conclusive.
- Ask for the original file. A re-upload or screenshot has lost metadata. The original from the camera, phone or software is the thing worth checking.
- Look for provenance data. Some files carry Content Credentials (C2PA). If present and valid they show a signed edit history; if absent they show nothing. You can check whether a file carries a C2PA manifest in your browser; it shows what the manifest declares and does not verify the signature. See our C2PA field guide.
- Compare fingerprints. If someone says a file was changed, a SHA-256 fingerprint of each version shows whether two files are identical. Use the hash checker; the fingerprint is made in your browser and the file stays on your device.
- Corroborate. Independent reports, other angles, location and time details, and the people involved are stronger than any single tool.
- Say what you do not know. “I could not verify this” is a better answer than a guess.
If it is your own file that is being doubted
The harder case is when real work is called fake. A chatbot or detector verdict is easy to quote and hard to rebut. The better defence is a history: the original file, earlier drafts, and a dated record. A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. It does not detect AI, show who made a file, or prove it is genuine. See how to protect yourself from AI forgeries and how to prove a file existed before a date.
For legal disputes, see our legal hub and the lawyers’ guide.
Tools that can help
There is no affiliate link on this page: nothing here depends on a paid third-party tool. Our affiliate disclosure explains when we do use them.
See the glossary for text-to-image model, training data and lateral reading.
The same limit applies to watermarks and platform labels. The AI watermark and label decoder explains what a SynthID result, a C2PA manifest or an AI label can and cannot show. For a product clip someone forwarded, use the viral product-video red-flags checker.
If you would rather test your eye on food advertisements, the free Real or AI? food ads quiz shows a real photograph next to an AI-generated dish of the same kind. It is a practice test. It cannot tell you whether any other file is AI-made.
Frequently Asked Questions
-
Can I ask a chatbot whether an image or video is AI-generated?
You can, but you should not rely on the answer. In a Maldita.es and University of Jaén audit of Grok on World Cup claims, image and video verdicts were right 49.4% of the time. That is one chatbot and one event, but it fits expert advice to treat chatbot verdicts as guesses.
-
Are chatbots better at checking written claims?
In the same audit, 95.8% of verdicts on written claims, including screenshots of text, were right, against 49.4% for images and video. Even so, written claims can be checked against sources directly.
-
What should I ask a chatbot instead of “is this real?”
Narrow questions about what is visible: describe what you see, read any text, and say what is unclear. Then verify the source and original file elsewhere.
-
Does a timestamp show a file is real?
No. A TimeBind shows that an exact file existed no later than a recorded time and still matches. It does not show who made it or that its content is true.
-
How do I show my own work is not AI when a tool says it is?
Keep the original file, earlier drafts and any dated records. A tool score is not proof either way. See our calm reply to an AI accusation.
Sources checked
- Maldita.es, “Grok, is this real?”: X’s AI gets an average of four in ten answers wrong about the 2026 World Cup, 27 October 2026 (English edition of the report; the site also publishes it in Spanish).
- NeoTeo, Grok fact-check accuracy varied by content type, cross-check of the 49.4% and 95.8% figures and the limits of the sample.
- Factnameh, Grok amplifies misinformation as users turn to it for Iran war content verification (quotes Alex Mahadevan, Poynter; also the source of the Minab example).
Keep a dated record of the original, before anyone asks whether it is real
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
If you want a short, interactive walk-through of labels, reverse image search, voice-clone scam checks, unreliable video tells and dated originals, see the free Spot AI fakes in 5 short lessons mini-course. It does not identify whether a file is synthetic.
