How to Tell if an Image or Video Is AI-Generated: Signs, Tools and Limits

How can you tell if an image is AI-generated?

The honest answer is that you often cannot tell by looking alone. In an online quiz run by Microsoft's AI for Good Lab, participants made about 287,000 judgements and identified real and AI-generated images correctly only 62% of the time ^[1]. A representative study of 3,002 people in the USA, Germany and China, presented at the IEEE Symposium on Security and Privacy in 2024, found that state-of-the-art fakes were nearly indistinguishable from real media and that most participants were simply guessing ^[2].

That does not make checking pointless. It means the reliable approach is layered: look for visual errors, check for machine-readable signals such as watermarks or Content Credentials, and trace the picture back to its first appearance. The US National Institute of Standards and Technology (NIST) describes the same idea at a larger scale: it says there is no silver bullet, and that provenance tracking, labelling and detection each have limits ^[4].

A quick checklist

1. Pause before sharing. Breaking-news footage deserves extra care; in June 2025, a clip of an explosion at Tehran's Evin Prison, which experts told CBS News may have been made with AI, appeared online within minutes of a real Israeli strike on the prison ^[12].
2. Look closely at hands, text, shadows, reflections and objects that should work in a physical way ^[3].
3. Check for credentials or watermarks with a tool that reads them, such as the Content Credentials standard or Google's SynthID checker in Gemini ^[5][7].
4. Search for the source. Google's own help page recommends visual inspection and reverse image search alongside its watermark tool ^[5].
5. Treat every result as evidence, not proof, because each method can miss or misread content ^[4].

What visual signs give away AI images?

Researchers at Northwestern University built a guide from 138 images generated with diffusion models, 9 images curated from social media and 42 real photographs ^[3]. They grouped the tell-tale problems into five categories ^[3]:

• Anatomical: implausible hands, eyes, teeth and bodies, including bodies that merge into each other ^[3].
• Stylistic: images that look a bit too perfect, with glossy or waxy skin, a magazine or film-still look, or parts of the picture with a different level of detail or colour ^[3].
• Functional: objects that could not work, atypical designs and incomprehensible text or logos, because generators lack a structural understanding of how things work ^[3].
• Violations of physics: inconsistent shadows, reflections and problems with depth and perspective ^[3].
• Sociocultural: scenarios that are unlikely, inappropriate, against the norms of a culture or historically inaccurate ^[3].

The authors stress that human-perceptible artifacts are not always present, although they often appear, and that anatomical oddities such as a strange-looking hand do not necessarily mean an image is AI-generated ^[3]. Generators also improve quickly: the Microsoft researchers note that their results probably overestimate how well people can spot current AI images ^[1].

Which fakes are hardest to spot?

In the Microsoft experiment, two kinds of images fooled people more often than not: human faces made with GANs (an older type of image generator) and inpainting, where part of a real photo is replaced with generated content ^[1]. NIST makes the same point: a partly edited image is mostly real, which complicates every detection method ^[4]. A real photo with one altered detail is therefore harder to catch than a fully synthetic scene.

What about video?

Video is harder still. Hany Farid, a professor at the University of California, Berkeley, told CBS News in June 2025 that fakes had moved on from still images to full video with explosions that looks like handheld phone footage ^[12]. In a clip of an explosion at the entrance of Tehran's Evin Prison, which experts said may have been AI-generated, they pointed to an incorrect sign above the door and inconsistencies in the explosion itself ^[12]. Useful things to check include signs and other text, lighting, and whether the same scene appears in footage from other cameras.

Do AI watermarks and Content Credentials prove where an image came from?

Machine-readable signals are stronger than eyesight, but only when they are present and the right tool can read them. There are two main kinds.

Invisible watermarks

A watermark is a signal hidden in the pixels or audio. Google's SynthID embeds imperceptible signals in AI-generated content, and Google says over 20 billion pieces of content had been watermarked with it by November 2025 ^[6]. Since 20 November 2025, people can upload an image to the Gemini app and ask whether it was made with Google AI ^[6]. Google's help page adds video and audio checks, with limits of 100 MB per file, videos under 90 seconds, and roughly 10 checks of each type per 24 hours ^[5].

The limits matter. Gemini can currently recognise only content created by Google AI tools, and Google states that finding no SynthID watermark does not mean the content is not AI-generated ^[5]. Google says watermarks usually survive resizing, recolouring and compression but may not be detected after many alterations ^[5].

Content Credentials

Content Credentials, the standard produced by the Coalition for Content Provenance and Authenticity (C2PA), are a cryptographically secure way to record how, when and where content was made and how it changed ^[7]. If the file or its provenance data is altered, the cryptographic link breaks, which signals tampering ^[7]. Gemini can read Content Credentials version 2.2 and later from products on the C2PA Conforming Products List, on the web and in its Android app ^[5].

Credentials, however, record a claim about origin, not the truth of a scene. Google notes that a result naming a company only shows who provided the information, and that credentials do not definitively prove whether content is AI-generated ^[5]. C2PA says credentials can become separated from a file and supports 'soft bindings', such as watermarks or fingerprints, to find them again ^[7].

Why do labels and metadata so often disappear?

Metadata travels inside the file, so anything that re-saves the file can remove it. NIST notes that embedded metadata is often stripped when files are shared, for example through social media platforms, and that people can strip it either to deceive or for benign reasons such as privacy ^[4].

The Washington Post tested this in October 2025. It made a video with OpenAI's Sora that carried Content Credentials, posted it to eight social apps, and found that none of them kept the credentials or let users view them ^[8]. Only YouTube added a warning, and that notice, reading 'Altered or synthetic content', sat in the video's description ^[8].

Visible watermarks are even easier to defeat. 404 Media reported that three websites it tested removed the visible Sora 2 watermark from videos in a matter of seconds ^[9]. NIST also describes research showing that many invisible image watermarks can be removed, for example by adding noise and then denoising the image ^[4].

What do platforms say?

Meta said in 2024 that it would read AI indicators in the C2PA and IPTC metadata standards to label images on Facebook, Instagram and Threads ^[10]. The same announcement admitted that it is not yet possible to identify all AI-generated content and that people can strip out invisible markers ^[10]. Meta instead requires users to disclose photorealistic AI video and realistic-sounding audio when they post it ^[10].

Are AI image detector tools reliable?

Automated detectors are classifiers that estimate how likely a file is to be synthetic. They can outperform people: the Microsoft team's own detector exceeded 95% accuracy on both real and AI images in its study ^[1]. That figure applies to one research tool on one dataset, not to any free website you might find.

Benefits

• The Microsoft researchers describe detection tools as significantly more reliable than humans, while noting that they still make mistakes ^[1].
• Their accuracy in that study was consistent across image categories, unlike human accuracy, which varied by subject ^[1].

Limits

• NIST says detectors often perform best on content from the generators they were trained on, while new models keep appearing ^[4].
• As soon as a new detection method appears, generators improve and adversaries learn to evade it, according to NIST ^[4].
• False positives, labelling real content as AI, can be extremely damaging, causing reputational harm or adverse treatment ^[4].
• Partially edited images are hard to classify, and whether an edit is 'substantial' can be subjective ^[4].

A detector score is therefore one input. It is most useful when it agrees with other evidence, such as missing original sources or visible errors.

Which method should you use, and when?

Each method answers a different question. Visual inspection asks whether the content looks physically plausible ^[3]. A watermark check asks whether a specific company's AI left its signal ^[5]. Content Credentials ask what the file's signed history says ^[7]. A reverse image search asks where and when the picture first appeared, which Google recommends alongside its own tool ^[5]. The table below compares them.

For a high-stakes decision, such as reporting a news event or accusing someone of faking evidence, use several methods and look for agreement. NIST notes that organisations may use several detection techniques together and says there is no silver bullet ^[4].

What rules require AI images to be labelled?

In the European Union, the transparency obligations in Article 50 of the AI Act apply from 2 August 2026 ^[11]. Under these rules, providers of AI systems must mark audio, image, video and text outputs in a machine-readable way so they can be detected as artificially generated, and deployers must label deepfakes; the European Commission's code of practice sets out how to do this ^[11]. The code defines deepfakes as image, audio or video content resembling real people, objects, places, entities or events that could falsely appear authentic ^[11].

Adherence to the code itself is voluntary, although the Article 50 duties are legal obligations ^[11]. The Commission's timeline lists publication of the final code on 10 June 2026, and about 190 organisations had signed it by the end of July 2026 ^[11]. The EU has also created a set of icons that deployers may use to label AI-generated content ^[11]. Outside the EU, platform policies such as Meta's disclosure rules apply ^[10], but rules differ by country and change often, so check the current version as of 3 November 2026.

What happens next for spotting fakes?

The trend is towards labels built into files at creation, rather than people spotting errors. Google announced plans to extend Gemini's verification to Content Credentials from products outside its own ecosystem ^[6], and its help page now describes those checks ^[5], while the EU's marking rules apply to providers ^[11]. At the same time, NIST describes attacks that remove or forge ('spoof') watermarks ^[4], and the Washington Post's test showed that platforms often discard the signals that do exist ^[8].

What remains uncertain

• How well people will do next year. The Microsoft authors expect human accuracy to fall as generators improve ^[1].
• Whether platforms will preserve credentials. Taking part in Content Credentials is voluntary, and in the Post's test no platform kept the data ^[8].
• Coverage of other companies' watermarks. As of 3 November 2026, Gemini's checker cannot yet read other companies' watermarks ^[5].

For now, checks that need no special tool still matter: Google recommends reverse image search to find an original source ^[5], and Hany Farid advised CBS News readers not to rely on social media for breaking news ^[12].