Is this real? How to spot AI-generated images, video and text

Can you still tell by looking?

Less and less. A few years ago, a generative AI picture often gave itself away with six fingers, melted earrings or shop signs in gibberish. Newer tools have fixed most of those slips, and the ones that remain are easy to crop out.

People were already bad at this before the tools improved. In a 2022 study in the journal PNAS, Sophie Nightingale of Lancaster University and Hany Farid of UC Berkeley found that participants spotted AI-generated faces 48% of the time, roughly a coin toss. Training raised that to 59%. The same study found people rated the synthetic faces as slightly more trustworthy than real ones.

So treat visual oddities as a bonus clue, not a test. The FBI still suggests looking for distorted hands, odd teeth or eyes and shadows that don't match. If you see them, good. If you don't, that tells you very little.

What actually works, then?

Habits that check the story around the content, rather than the content itself. They worked for doctored photos long before AI, and they still work now.

Ask who posted it first. A screenshot of a screenshot, shared by an account you have never heard of, is not a source. Try to find the original upload: the first account, the date and whether that person or outlet has a track record.

Run a reverse image search. Google's help page explains that you can upload an image, drag it into the search box or right-click it in Chrome and choose "Search with Google Lens". The results include websites showing the same or similar images. That often reveals an older photo, a different place or an obvious fake already debunked.

See whether reputable outlets report it. A real explosion at a famous landmark, or a real video of a world leader saying something shocking, will be on several established news sites within the hour. If only anonymous accounts have it, wait.

Slow down when it makes you angry or scared. Content built to provoke a strong feeling is built to be shared before it is checked. That is true of human propaganda and of deepfakes. A minute's pause is the cheapest defence there is.

What are content credentials and AI labels?

Some tools now attach a kind of digital receipt to the files they make. The main standard is called C2PA, run by the Coalition for Content Provenance and Authenticity. Its consumer-facing name is Content Credentials, and it is backed by more than 500 companies, including Adobe, Microsoft, Google, Meta, OpenAI, Amazon and the BBC, according to the project's website.

When a file carries Content Credentials, you may see a small "CR" pin. Clicking it shows how the file was made and its editing history. You can also upload a file to the free checker at verify.contentauthenticity.org. The Content Authenticity Initiative says the records are cryptographically signed, so later changes show up.

Two caveats. The initiative itself says credentials "aren't intended to prescriptively indicate whether a piece of content is 'real'", and adding them is optional. So a missing credential proves nothing: most genuine photos online don't have one either.

Google's SynthID takes a different approach: an invisible watermark built into the content as it is made. Google DeepMind says it marks images, video and audio from Google's tools and text from the Gemini app. You can upload an image, video or audio clip to Gemini, Google Search or Chrome and ask whether Google AI created or altered it. Google's SynthID Detector portal is now open to everyone and also checks content from partners including OpenAI and Nvidia, the company says. It only finds SynthID watermarks, so a "no" means "not made by these tools", not "real".

Platform labels rely on these signals plus self-reporting. Meta shows an "AI info" label on Facebook, Instagram and Threads when it detects industry-standard AI indicators or when the poster says so. Since September 2024, posts that were only edited with AI have that label tucked into the post's menu, while fully AI-generated content keeps a visible label.

YouTube requires creators to disclose realistic content that is made or meaningfully altered with AI, such as making a real person appear to say something they didn't. A label may then appear on the video player, or in the description for animated content. YouTube may also add labels itself when it finds C2PA data or its own systems detect AI. Cosmetic edits, captions and AI-written scripts don't need disclosure.

The lesson is the same for all of them: a label is useful information when it is there. Its absence is not a clean bill of health.

What does EU law require?

The EU's AI Act has a transparency rule, Article 50, that applies from 2 August 2026. It has two parts that matter here.

First, anyone who uses AI to make a deepfake (image, audio or video that resembles real people and could falsely appear genuine) must disclose that it was artificially generated or manipulated. For obviously artistic, satirical or fictional work, the duty shrinks to flagging that AI content is present, in a way that doesn't spoil the work. AI-written text published to inform the public on matters of public interest must also be disclosed, unless a human has reviewed it and someone takes editorial responsibility.

Second, companies that provide AI generators must mark their output in a machine-readable way, so that software can detect it. Under the "Digital Omnibus" amendment published in July 2026, systems already on the market before 2 August 2026 have until 2 December 2026 to comply, according to the law firm Garrigues.

In practice, this means more labels in Europe over the coming months. It doesn't mean every fake will be labelled: scammers don't follow disclosure rules. Our EU AI Act explainer covers the rest of the law.

What about phone calls in a loved one's voice?

This is the scam to prepare for now, because it doesn't need you to be online. The US Federal Trade Commission warns that a short audio clip of your family member's voice is enough for a cloning program to imitate them. The FBI says criminals use such clips to fake emergencies and demand money.

The FTC's advice is blunt: "Don't trust the voice." Hang up and call the person back on a number you already know is theirs. If you can't reach them, contact another relative or a friend. Requests to pay by wire transfer, cryptocurrency or gift cards are classic warning signs.

The FBI also recommends creating a secret word or phrase with your family. Pick something that isn't on social media, agree it in person, and use it whenever a call asks for money or urgency. It feels a little silly until the day it isn't.

Can a tool tell me if text was written by AI?

Not reliably. OpenAI released its own AI-text classifier and said it correctly identified only 26% of AI-written text, while wrongly labelling human writing as AI 9% of the time. On 20 July 2023 the company withdrew it, citing "its low rate of accuracy".

Independent research found a worse problem: bias. A 2023 study in the journal Patterns, led by James Zou at Stanford, ran 91 essays written by non-native English speakers through seven detectors. They labelled more than half as AI-generated, and one flagged nearly 98%. The same detectors correctly identified more than 90% of essays by US eighth-graders as human.

The reason is that detectors tend to flag plain, predictable wording, which is also how many careful second-language writers write. A large language model can also be asked to vary its style, which fools them the other way. Treat any detector score as a weak hint at most, and never as proof against a student, colleague or job applicant.

For text, the same source-checking habits apply. Who published it, can you find the claims elsewhere, and do the quotes and links actually exist? AI-written text often contains confident errors, known as hallucinations, and checking a couple of facts is usually quicker than guessing at style.

A quick checklist

1. Pause if it is shocking, urgent or makes you furious.
2. Find the original post, account and date.
3. Reverse image search photos and video stills.
4. Look for reputable reporting of the same event.
5. Check for labels or Content Credentials, remembering that no label proves nothing.
6. Call back on a known number before acting on any voice asking for money, and use your family code word.
7. Don't trust AI-text detectors to judge a person.
8. When unsure, don't share.