Guide

How to Tell If an Image Is AI-Generated in 2026: The Checks That Actually Work

A practical 2026 guide to checking whether an image is AI-generated: the three free provenance checks that give a real answer, the geometry tests that still beat the eye, why 'look at the hands' stopped working, what the 2026 security research says about Content Credentials, and the checks that prove nothing at all.

September 29, 2026 · Updated September 30, 2026 · The AI Rankings

Quick answer: Do not start by looking at the image. Start by asking whether the file still carries its origin, because that is the only check that can produce evidence rather than a guess: upload it to the Gemini app and ask whether it was created or edited by Google AI, which reads both Google’s SynthID watermark and C2PA Content Credentials (Google); run it through openai.com/verify, which checks for the C2PA metadata and SynthID watermark OpenAI embeds in images from ChatGPT, Codex and its API (OpenAI); and inspect the file’s Content Credentials with any C2PA reader. If those come back empty — which they often will, because a screenshot, a re-save or a platform upload can strip metadata — fall back to geometry rather than vibes: check whether the lines joining objects to their shadows and to their reflections all converge on a single point, the test digital-forensics researcher Hany Farid published for exactly this purpose (Content Authenticity Initiative). The honest caveat that governs everything below: in the largest published experiment on the question, people scored 62% across roughly 287,000 judgements, barely above a coin toss (Microsoft AI for Good Lab, arXiv 2507.18640), and on faces specifically 2026 participants managed 58.4% (Journal of Vision, July 2026). Your eye is not the instrument. Provenance is.

This guide is the procedure: what to check, in what order, what each check can and cannot prove, and when to stop. It does not rank detection tools — for tested tools with independent accuracy figures and pricing, see best AI image detectors. For AI-written text rather than images, see best AI detectors. For how the invisible marks themselves work, see what is AI watermarking.

The short version: check in this order

The order matters more than the individual checks, because each step is cheaper and more conclusive than the one after it. Work down the list and stop as soon as you have an answer.

Steps 1 to 4 look for evidence. Steps 6 to 8 look for symptoms. Evidence beats symptoms every time, which is why a guide that opens with “count the fingers” has the order backwards.

Why “look at the hands” stopped working

The visual tells that defined 2023 are mostly gone, and the research now says so in numbers rather than impressions.

People are barely above chance overall. Researchers at Microsoft’s AI for Good Lab analysed approximately 287,000 image evaluations from more than 12,500 participants playing an online real-or-not quiz and found an overall success rate of 62% (arXiv 2507.18640). The paper’s own summary calls this “a modest ability, slightly above chance”. Participants did best on human portraits and, in the authors’ words, “struggled significantly with natural and urban landscapes” — the opposite of the folk assumption that faces are the hard case.

On faces, 2026 performance is worse than the folk assumption too. A study published in the Journal of Vision in 2026 by Alexis McGuire, Paul Taylor and Sophie Nightingale of Lancaster University with Maty Bohacek of Stanford and Hany Farid of UC Berkeley put 169 participants in front of 96 faces and recorded 58.4% accuracy at separating real from synthetic (Lancaster University, 7 July 2026).

And synthetic faces now read as more trustworthy than real ones. In the same study, on a seven-point scale, real faces were rated 4.03, faces from generative adversarial networks 4.36, and faces from diffusion models 4.70 (Lancaster University). The most synthetic faces were the most trusted. McGuire describes the result as “a paradox” that points to realism and trustworthiness being “driven by two different psychological mechanisms”. For anyone checking an image, the practical reading is blunt: a face looking honest is evidence of nothing.

The face-detection gap is measured, not anecdotal. A study of 2,203 undergraduates judging 200 face images, published in Cognitive Research: Principles and Implications on 7 January 2026, put average human discrimination at an AUC of 0.53 — chance is 0.50 — with participants correctly identifying 67% of real faces but only 31% of deepfakes, while a convolutional neural network on the same images reached 97% accuracy (Pehlivanoglu et al.). The same paper found the reverse for video: on 70 clips, humans reached an AUC of 0.67 while the FaceForensics algorithm managed 49% accuracy. Machines win on stills; people win on motion. This page is about stills, so the machine-first ordering above is the one the evidence supports.

The generators that make the artefacts have moved on. The current frontier — OpenAI’s GPT Image 2.5, Microsoft’s MAI-Image-2.6 and Google’s Nano Banana family among them — is several model generations past the ones the “six telltale signs” listicles were written against. We track the generators themselves, with capabilities and pricing, in best AI image generators.

A practical consequence follows: any check based on how an image looks is a check on a specific generation of model, and it expires. Provenance checks do not expire in the same way, because they do not depend on the model being bad at something.

Step 1 to 3: the three free provenance checks

These are the checks that can return evidence. All three are free and take under a minute each.

Ask the Gemini app

Upload the image to Gemini and ask whether it was created or edited by Google AI. Google states the feature checks two things: SynthID, its invisible watermark, and Content Credentials, which it describes as “a digital passport, documenting the digital content’s origin and history” (Google).

The limits are published, and they are the reason this is step one rather than the only step. Files must be under 100MB. Quotas are roughly 10 image checks per rolling 24 hours. And Google is explicit about what a negative result means: “If a SynthID watermark isn’t detected, it means the image or video wasn’t created or edited by Google AI, but it could have been created by other AI systems.” Google also advises against uploading screenshots or collages, recommending a tight crop around the image, and says Content Credentials checks are available on the web and Android, with iOS to follow.

Two things about this check are worth knowing. It is currently the most accessible route to a SynthID result, because Google’s dedicated SynthID Detector portal is still restricted — DeepMind says it is “currently collaborating with journalists and media professionals to test the portal”, with access by waitlist (Google DeepMind). And the scale behind it is large: Google reported more than 10 billion pieces of content watermarked with SynthID when it announced the portal (Google, 20 May 2025).

Run it through openai.com/verify

OpenAI embeds both C2PA metadata and a SynthID watermark in supported images, and states that “supported images generated with ChatGPT, Codex, and the OpenAI API include both signals” (OpenAI). Upload the file at openai.com/verify to check. Organisations can use OpenAI’s verification API for the same check at volume.

OpenAI publishes two caveats that matter. Coverage “can vary by product, model, export path, file type, and when the content was created”. And durability is limited: C2PA Content Credentials “can sometimes be removed by platforms, editing tools, or file conversions”, while watermarks can be degraded by “compression, cropping, noise, edits, format conversion, or other transformations”.

Inspect the Content Credentials

For what a credential contains and how the standard works, see what is C2PA.

C2PA Content Credentials are a signed record attached to the file describing what captured or edited it. Where present, they are the strongest single signal available, because they are cryptographically signed rather than inferred. The C2PA steering committee is Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic (C2PA), and C2PA published Content Credentials Deployment Guidance 1.0 in July 2026 (C2PA).

Capture-side support reached consumer hardware in 2025 and 2026. The Content Authenticity Initiative reports that the Google Pixel 10 supports C2PA credentials and that Sony shipped the PXW-Z300 video camera with Content Credentials support, in a post that also puts the initiative at more than 6,000 members (Andy Parsons, CAI, 18 January 2026).

The catch is coverage in the other direction. A signed credential saying an image came from a camera is meaningful. No credential at all is meaningless, because the overwhelming majority of images on the internet have never carried one, and the ones that did frequently lose them on upload.

Step 4: read the metadata field that names the generator

The IPTC — the standards body for news metadata — recommends that AI-generated images carry the “Digital Source Type” property with the value trainedAlgorithmicMedia, expressed as the URI http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia (IPTC, 9 May 2023). Related values in the same vocabulary include compositeSynthetic for composites mixing synthetic and camera-captured elements, digitalArt for human-made digital art, and virtualRecording. IPTC says the value can be written into the file’s XMP packet or carried inside a C2PA manifest, and that C2PA has integrated digitalSourceType into its own specification.

In practice: open the file in any metadata viewer that shows XMP, and look for a digital source type. If it says trainedAlgorithmicMedia, the file is declaring itself. That is not proof — a declaration can be written by anyone — but a positive hit is a strong lead and takes two minutes.

Step 5: reverse image search the image, not the claim

Reverse image search answers a different and often more decisive question: has this image existed before now, and in what context?

Run the file through more than one index, because their coverage differs. Google Lens reaches Google’s image index and its “About this image” panel; Google has committed to surfacing C2PA metadata in that panel across Google Images, Lens and Circle to Search so that users can “see if it was created or edited with AI tools” (Google, 17 September 2024). Microsoft Bing Visual Search, TinEye and Yandex Images each index material the others miss.

Three outcomes are worth acting on. If the image appears years before the event it supposedly shows, the claim is false whether or not the image is synthetic. If the image appears nowhere at all and is presented as a news photograph, that absence is itself a signal, because real news photographs travel. And if the image appears on a stock or generator-gallery site, you frequently get the generator named for free.

This step also answers the question that often sits behind “is this AI”: is the claim true? A fabricated caption on a genuine photograph is a deception no detector will ever catch.

Step 6: the shadow and reflection test, step by step

This is the strongest check available to the naked eye, and it is the one worth learning properly, because it rests on physics rather than on a generator’s current weaknesses.

Hany Farid set out the technique for the Content Authenticity Initiative on 14 September 2023 (CAI). The underlying fact is simple: “A point on an object, its corresponding shadow, and the light source responsible for the shadow all lie on a single line.”

For shadows:

1. Pick a distinct point on an object — a corner, a spike, the tip of a nose.
2. Find the same point on that object’s shadow.
3. Draw a straight line through the two points and extend it.
4. Repeat for at least three other object-and-shadow pairs elsewhere in the image.
5. In an authentic photograph lit by a single source, all those lines meet at one point — the light source. In a generated image, they frequently do not.

For reflections, in water, glass or a mirror:

1. Pick a distinct point on an object and the matching point on its reflection.
2. Join them with a line and extend it.
3. Repeat across several matched pairs.
4. All the lines should intersect at a single point. Farid’s summary: “This geometry of reflections suggests a simple forensic technique for verifying the integrity of reflections.”

For straight lines and perspective, the same logic applies to the scene itself: parallel real-world lines such as floor tiles, floorboards, railings or window mullions must converge on a common vanishing point. Generated images routinely produce tiles that drift.

Two things make this check credible rather than folklore. First, Farid’s own assessment when he published it: “Today’s AI-generated images seem to struggle to produce perspectively correct shadows and reflections.” Second, it has been tested at scale. A study titled “Shadows Don’t Lie and Lines Can’t Bend! Generative Models don’t know Projective Geometry…for now” built classifiers that look only at derived geometry — shadows, perspective fields and line convergence — and reported AUCs between 0.72 and 0.97, including 0.80 to 0.82 on an “unconfident” indoor test set where pixel-based classifiers perform at chance, generalising across Stable Diffusion XL, DeepFloyd IF, PixArt-α, DALL-E 3 and Adobe Firefly (Sarkar et al., arXiv 2311.17138). Its conclusion is the sentence to remember: “Generated images contain geometric structures not seen in real images.”

The limits are equally published. Farid names three: matched points need distinct shapes to be identifiable, constraint lines that are nearly parallel introduce large errors, and lens distortion corrupts the construction. And he flags the expiry date himself — “it may just be a matter of time before generative AI will learn to create images with full-blown perspective consistency”.

Step 7: the visual artefacts that are still worth a look

These are symptoms rather than evidence. Treat a hit as a reason to look harder, never as a verdict, and treat a miss as meaning nothing at all.

Notice what is missing from that table: hands, and the number of fingers. Hands were a genuine tell against 2022 and 2023 models. Any guide still leading on finger-counting was written against models that have since been superseded several times.

Platform labels: what they do and do not tell you

Platform labels are a free signal sitting on top of the image, and they are worth reading before you do any forensic work.

TikTok announced on 19 November 2025 that it would start adding invisible watermarks to AI-generated content made with TikTok’s own tools — described as “a robust technological ‘watermark’ that only we can read” — specifically so that its labels survive a download and re-upload, and that it would “continue reading C2PA Content Credentials and adding them into AI-generated content made on TikTok”. The same post says TikTok has labelled more than 1.3 billion videos to date and has added context to its AI-generated-content labels explaining whether a label came from TikTok’s own detection, from the creator, or from TikTok’s AI tools (TikTok Newsroom).

Read a platform label as follows. A label present is a reasonable signal that something in the pipeline detected a mark or the creator disclosed. A label absent proves nothing, for the same reason a missing credential proves nothing: the file may never have carried a mark, or may have lost it. And labels attach to how a file was processed, not to whether its content is honest — a photograph lightly retouched with a generative tool can be labelled while a wholly fabricated claim on an unedited image is not.

What provenance actually proves — and the 2026 security findings

This is where the honest version of this guide parts company with the optimistic one, and it is the single most important section for anyone whose check has real consequences.

Content Credentials are the best signal available. They are also not yet courtroom-grade, and the strongest statement of that comes from a security analysis published on 23 April 2026 by researchers at UMBC, Hacker Factor and the NSA, titled “Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short” (Golaszewski et al., arXiv 2604.24890). Examining C2PA versions 2.2 to 2.4, the paper documents six classes of weakness:

• Timestamps can be altered undetected, because nothing in the signed data references them.
• Revocation is frequently not checked, so many validators accept signatures from compromised keys.
• Validators disagree. The paper’s own demonstration has a Nikon Z6 III used to sign AI images, with Adobe Inspect reporting revoked certificates as valid while Verifieddit reported them invalid.
• Some file regions sit outside the signature. Exclusion ranges allowed modified GPS coordinates in a Pixel 10 Pro image to go undetected by the Proofmode validator.
• Credentials expire. The paper shows an Arizona Secretary of State image that validated in January 2025 and failed validation a year later, which the authors note conflicts with legal retention requirements.
• The conformance programme is weak, lacking security requirements and source-code review.

The authors’ conclusion, stated plainly, is that C2PA “should not yet be relied upon for high-stakes uses”.

That finding does not make the provenance checks above pointless — it makes them a strong signal rather than proof. The practical translation:

• A valid credential from a known signer is good evidence about origin, and should be checked in more than one validator if anything turns on it.
• A watermark hit is good evidence that a specific provider’s model touched the file.
• Neither tells you who took the photograph, when, or whether the caption is true.
• Absence of any signal is not evidence of anything, in either direction.

For how the standard works, see what is C2PA; for the invisible marks, see what is AI watermarking. This page stops at what they mean for a single check.

Asking a chatbot is not a check

The most common shortcut in 2026 is to paste an image into a general-purpose assistant and ask “is this AI?”. Unless the assistant is running an actual detector, the answer is a guess dressed as an analysis.

NewsGuard tested this directly on 22 January 2026 with Sora-generated videos shown with and without watermarks. On the items without a visible watermark it reported failure rates of 95% for Grok, 92.5% for ChatGPT and 77.5% for Gemini (NewsGuard). The test was on video rather than stills, so it is indicative rather than directly transferable. The pattern it exposes transfers cleanly, though: the models did well when a visible watermark was present and failed when it was not, which is to say they were reading a label, not analysing the content.

There is an important exception, and it is why step one above is what it is. Gemini’s verification flow is not the chatbot guessing — it is Gemini running SynthID and Content Credentials checks and reporting the result (Google). In the same NewsGuard report, Gemini identified all five watermark-removed Nano Banana Pro images it was shown as AI-generated. Ask Gemini to verify, and you get a check. Ask any assistant for an opinion on how an image looks, and you get its opinion on how an image looks.

When to reach for a detector tool

Detector tools have a place, and it is late in the process and narrow in scope. They produce a probability, not a finding.

Use one when the provenance checks came back empty, the geometry is inconclusive, and you need a second opinion before making a decision. Run two or three rather than one, and treat disagreement between them as the answer it is — a reason to investigate further, not a tie to be broken.

The single figure to hold in mind before you quote a detector score to anyone: in NewsGuard’s audit — the most rigorous public test of consumer image detectors, run on 15 authentic news photographs — five leading tools collectively labelled genuine photographs as AI-generated 13.33% of the time, with the worst tool wrong on 40% of real images (NewsGuard, 8 May 2026). Tools that avoided false positives in that audit were the weakest at catching genuinely manipulated images, and the reverse. There is no tool that is currently good at both.

We rank the tools, with the full audit results, pricing and a best-for verdict on each, in best AI image detectors. For faces and video specifically, see best deepfake detectors. For the mechanics of how classifier-based detection works and why it fails, see how AI detectors work.

Checks that prove nothing

Each of these circulates widely and none of them supports a conclusion.

• Counting fingers. A 2023 tell against 2023 models, which current generators have superseded.
• “It just looks AI.” The smooth, high-contrast, over-lit look is a style, reproducible by any photographer with a preset, and current models will produce grain, blur and imperfect exposure on request.
• File size or dimensions. Generated images are exported at ordinary sizes and resolutions, and any real photograph can be resized.
• Absence of EXIF camera data. Most social platforms strip EXIF on upload. A photograph with no camera data is the normal case, not a suspicious one.
• Absence of Content Credentials. Same logic, stated by OpenAI itself: metadata “can sometimes be removed by platforms, editing tools, or file conversions”.
• A single detector’s score. See the 13.33% false-positive rate above.
• A chatbot’s opinion. See the 77.5% to 95% failure rates above.
• The image being of a real, named person or place. Current models generate recognisable people and locations without difficulty.

Two procedures: 60 seconds and 20 minutes

If you have 60 seconds — you are deciding whether to share something:

1. Upload to the Gemini app and ask whether it was created or edited by Google AI.
2. Reverse image search it and read the oldest result’s date.
3. If both come back clean and the image is being used to make a factual claim, do not share it as fact. Treat it as unverified.

If you have 20 minutes — you are about to publish, escalate, or accuse someone:

1. Run all three provenance checks: Gemini, openai.com/verify, and a C2PA inspector.
2. Read the raw XMP for an IPTC digital source type.
3. Reverse image search across at least two indexes and establish the earliest appearance.
4. Run the shadow and reflection vanishing-point construction on at least four matched pairs.
5. Check straight lines against a common vanishing point.
6. Run two or three detector tools and record the disagreement rather than the average.
7. Identify who first published the image and what they say about where it came from.
8. Write down which of the eight steps produced evidence and which produced impressions. If only impressions, you do not have a finding.

Step 8 is the one people skip, and it is the one that prevents the failure mode this whole field keeps producing: a confident accusation built entirely on symptoms.

If you are about to accuse someone

The stakes are asymmetric, and the research is unambiguous about which way the errors fall.

A detector calling a real photograph fake happened in 13.33% of cases in NewsGuard’s audit of five tools on 15 authentic photographs (NewsGuard, 8 May 2026). Human judgement on faces sat at an AUC of 0.53 with only 31% of deepfakes correctly identified (Pehlivanoglu et al., 7 January 2026). And the provenance layer that is supposed to be the trustworthy one carries documented weaknesses that its own analysts say make it unsuitable “for high-stakes uses” (Golaszewski et al., 23 April 2026).

The defensible position: treat every signal on this page as a reason to ask the person where the image came from, and treat their answer — the original file, the camera, the download history, the context — as the evidence. A checkable provenance story is worth more than any score. This is the same conclusion the text side of this problem reached: a detection signal is a prompt for a conversation about process, never standalone proof of misconduct.

Frequently asked questions

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

Check the file’s origin before you look at the picture. Upload it to the Gemini app and ask whether it was created or edited by Google AI, which checks both Google’s SynthID watermark and C2PA Content Credentials; run it through openai.com/verify, which checks for OpenAI’s C2PA metadata and SynthID watermark; and inspect its Content Credentials with a C2PA reader. If all three come back empty, fall back to geometry: extend lines from objects to their shadows and to their reflections, and check whether they converge on a single point, which they should in a real photograph and frequently do not in a generated one. Visual artefacts such as garbled small text or broken repeated patterns are worth a look but prove nothing on their own.

Is there a free tool to check if an image is AI?

Yes, three of them, and the useful ones are provider verification tools rather than detectors. The Gemini app checks for SynthID and Content Credentials free, with a quota of roughly ten image checks per rolling 24 hours and a 100MB file limit. OpenAI’s verify page at openai.com/verify checks for the C2PA metadata and SynthID watermark OpenAI embeds in images from ChatGPT, Codex and its API. Any C2PA Content Credentials reader will display a signed credential where one survives. Free classifier-based detectors also exist, but they return probability scores with documented false-positive rates rather than a verifiable answer.

Can AI-generated images be detected reliably?

Only when the image still carries a watermark or signed credential from a provider that embeds one. In that case detection is genuinely reliable, because there is a real signal to find rather than a style to guess at. When no mark survives, reliability collapses: five leading consumer detectors in NewsGuard’s May 2026 audit labelled genuine photographs as AI-generated 13.33% of the time, and a 20,000-image benchmark found mean detector accuracy dropping from 86.7% to 51.8% on text-rich images, with no method above 80%. Reliable detection in 2026 means reading provenance, not classifying pixels.

Do AI images still have weird hands and fingers?

Usually not. Hand and finger errors were a dependable tell against 2022 and 2023 models, and those models have been superseded several generations over. Any guide that leads on counting fingers is describing an older generation of model. The visual checks that have held up better are geometric rather than anatomical: shadows and reflections that do not converge on a single point, parallel lines that fail to meet at a common vanishing point, and repeated structures such as railings or windows that lose their count across a frame.

What are Content Credentials and how do I check them?

Content Credentials are a signed record attached to an image file describing what captured or edited it, built on the open C2PA standard. Check them by opening the file in any C2PA reader, which will display the manifest if one survives. The C2PA steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic, and capture-side support has reached consumer hardware — the Content Authenticity Initiative reports that the Google Pixel 10 signs images with C2PA credentials. A credential present is strong evidence about origin; a credential absent is evidence of nothing, because platforms, editors and file conversions routinely strip them.

Does a missing watermark mean an image is real?

No, and this is the single most common error in reading these checks. Google states the position directly for its own tool: if a SynthID watermark is not detected, it means the image was not created or edited by Google AI, but it could have been created by other AI systems. Most AI image generators embed nothing, open-weight models run locally embed nothing, and any mark that was present can be degraded by compression, cropping, editing or format conversion. A negative provenance result narrows the field of possible generators. It never establishes that a human took the photograph.

Can I just ask ChatGPT or Gemini if an image is AI-generated?

Only if you ask for a verification rather than an opinion. NewsGuard’s January 2026 test of non-watermarked Sora videos recorded failure rates of 95% for Grok, 92.5% for ChatGPT and 77.5% for Gemini, which shows the models were reading visible watermarks rather than analysing content. Gemini’s dedicated verification flow is different, because it runs actual SynthID and Content Credentials checks and reports the result — in the same NewsGuard report it correctly identified all five watermark-removed Nano Banana Pro images it was shown. Ask Gemini to verify the file. Do not ask any assistant whether a picture looks real.

How accurate are people at spotting AI images?

Barely better than chance, and the figures are consistent across studies. Microsoft’s AI for Good Lab analysed roughly 287,000 judgements from over 12,500 participants and found an overall success rate of 62%, with people doing best on human portraits and worst on natural and urban landscapes. A 2026 study in the Journal of Vision recorded 58.4% accuracy on 96 faces. And a January 2026 study of 2,203 participants judging 200 face images put average discrimination at an AUC of 0.53, with only 31% of deepfakes correctly identified, against 97% for a convolutional neural network on the same images.

What is the shadow test for AI images?

It is a geometric check based on the fact that a point on an object, its corresponding shadow and the light source all lie on a single straight line. Pick a distinct point on an object, find the matching point on its shadow, draw a line through both and extend it, then repeat for several other pairs. In a photograph lit by one source, all those lines meet at a single point. In generated images they frequently do not. Hany Farid published the technique for the Content Authenticity Initiative in September 2023, and classifiers built on the same geometry — shadows, perspective fields and line convergence — have reported AUCs of 0.72 to 0.97, including 0.80 to 0.82 on test sets where pixel-based classifiers perform at chance.

Does metadata show if an image was made by AI?

Sometimes, and it is worth two minutes of checking. The IPTC recommends that AI-generated images carry a “Digital Source Type” property with the value trainedAlgorithmicMedia, written into the file’s XMP packet or carried inside a C2PA manifest, and C2PA has integrated that property into its own specification. Related values include compositeSynthetic for mixed synthetic and camera-captured content. Open the file in any metadata viewer that shows XMP and look for it. A positive hit is a strong lead, but a declaration can be written by anyone, and most files carry no such field at all.

Will an AI label on Instagram or TikTok tell me an image is AI?

It will tell you that something in the platform’s pipeline detected a mark or that the creator disclosed, which is not the same question. TikTok said in November 2025 that it adds invisible watermarks to AI content made with its own tools, continues to read C2PA Content Credentials, and has labelled more than 1.3 billion videos, with labels now indicating whether they came from TikTok’s detection, the creator, or TikTok’s AI tools. Read a label present as a reasonable signal and a label absent as no signal at all. Labels also attach to how a file was processed rather than to whether its caption is true.

Are Content Credentials proof that an image is authentic?

No, and the most careful assessment of that comes from security researchers rather than critics of the standard. A paper published on 23 April 2026 by researchers at UMBC, Hacker Factor and the NSA examined C2PA versions 2.2 to 2.4 and documented timestamps that can be altered undetected, revoked certificates that many validators accept, validators that disagree on the same file, file regions left outside the signature, and credentials that stop validating within months. The authors conclude that C2PA “should not yet be relied upon for high-stakes uses”. Treat a valid credential as strong evidence about a file’s origin, verified in more than one validator if anything turns on it, and never as a verdict on whether the image is honest.

What should I do if I cannot tell whether an image is AI-generated?

Stop trying to judge the image and start checking the claim. Reverse image search it across at least two indexes and find its earliest appearance; if the image predates the event it supposedly shows, the claim is false whether or not the image is synthetic. Then ask whoever published it where the file came from, and treat a checkable provenance story — the original file, the camera, the download history — as the evidence. If you are left with impressions rather than a provenance result, a reverse-search result or a geometric finding, you do not have a conclusion, and the correct action is to label the image unverified rather than to call it fake.

← All guides