Why visual tells stopped working
For a while generated images gave themselves away with extra fingers, melted jewellery or text that looked like letters but was not. Each new generation of models fixes more of these, and a single tell was never proof: real photographs have motion blur, odd reflections and bad crops too. Treat a visual oddity as a reason to check, not as an answer.
Checks that hold up
| Check | How | Limits |
|---|---|---|
| Platform label | Look for the network's AI or edited label | Missing labels mean little; labels can be wrong |
| Provenance data | Ask for the original file and read its credentials | Most shared copies have none |
| Earlier copies | Run a reverse image search | New images have no history yet |
| Other sources | Find other photos or reports of the same event | Takes time for breaking news |
| The account | Check history, location claims and links | New accounts can be genuine |
Meta has described labelling images when it detects industry-standard indicators such as C2PA and IPTC metadata, and asking users to disclose realistic generated media. That makes labels useful when present and weak evidence when absent. Provenance data is covered in checking Content Credentials before you share, and hidden marks in invisible watermarks on AI images.
How to spot AI images: a worked example
Suppose a community moderator logs 20 striking images that appeared in a local group over one week and runs the five checks on each. The counts are invented for this example. Platform labels are present on 3 of the 20. Provenance data can be checked on 4, because only 4 posters supplied an original file; 2 of those carry a valid manifest with a camera capture, 1 records a generative action, and 1 has no manifest. Reverse image search finds earlier copies of 9 images, of which 6 trace to stock libraries or older news events and 3 to a generator's public gallery.
Other sources for the claimed event exist for 7 of the 20. Account checks find 5 posters with new accounts and no history. Putting the checks together, the moderator ends the week with 8 images supported by two or more checks, 6 contradicted by at least one check (an earlier copy from a different event, or a generator gallery), and 6 with no evidence either way. The 6 unsupported images are the hard case: they are left up with a comment asking for a source, rather than removed on the strength of odd-looking hands. Two of them gain a source within a day and move to the supported group; the other four stay unverified, which is an honest end state for a week's work.
Pitfalls
- Sharing with a caveat. A share travels; the caveat rarely travels with it, and the card is what most readers see.
- Trusting a reverse image search that finds nothing. A new image has no history yet, which says nothing about how it was made.
- Treating a missing platform label as clearance. Labels depend on metadata or disclosure, and most generated images carry neither.
- Reading an odd detail as proof. Real photographs have compression artefacts, motion blur and odd reflections too.
- Checking only the image and never the claim. A real photograph attached to a false caption is the older and more common fake.
- Running the checks only on images you distrust. The ones that fit your expectations are the ones you are least likely to question and most likely to share.
What to record
For any image you decide to share or moderate, write one row: which checks you ran, what each found and what you did. The rows below are examples. Over time the log shows which checks earned their keep for your feed; earlier copies and other sources usually settle more cases than labels do. The publisher side of the same checks is set out across the provenance section.
| Image | Checks run | Time (UTC) | What they found | Decision |
|---|---|---|---|---|
| Flooded high street | Label, reverse search, other sources | 2026-09-13 18:02 | No label; earlier copy from 2021 elsewhere | Do not share; reply with the source |
| Concert crowd | Reverse search, account | 2026-09-13 18:15 | No earlier copy; account 8 years old, local | Share with attribution |
| Storm over the harbour | Provenance, other sources | 2026-09-13 18:30 | Manifest records a generative action | Do not share as a photograph |
| New mural | Label, account | 2026-09-13 18:41 | No label; new account, no history | Ask for a source, wait |
Before you share
- If the image makes a claim about a real event or person, find a second source before sharing.
- If you share it anyway, say what you know and do not know in your own words.
- If it turns out to be generated, post the correction where the original share went.
Text in feeds is harder: short posts give detectors almost nothing to work with, as whether an AI detector can judge a social post explains. For images, the evidence-based checks above are about as good as it gets. The provenance section covers the same questions from the publisher's side.
Sources
- Meta, Our Approach to Labeling AI-Generated Content and Manipulated Media (April 2024)
- Google Search Help: search with an image
- IPTC NewsCodes: digital source type vocabulary
Questions
How can I tell if an image is AI-generated?
Combine evidence: a platform label, provenance data in the original file, a reverse image search for earlier copies and other sources for the same event.
Are extra fingers still a sign of AI images?
Less and less. Newer models fix most visual tells, and real photographs have oddities too.
Does reverse image search find AI images?
It finds earlier copies and similar images, which often reveals where an image came from. A brand-new image may have no history yet.
Do social networks label all AI images?
No. Labels depend on detected metadata or user disclosure, so many generated images carry none.
What should I do if I cannot verify an image?
Wait. Ask the poster for the original file or a source, look again in a day when other reports may exist, and if you must share, say plainly that it is unverified.
Hannah Voss, Editor. Checks every guide against a working WordPress install and the networks' current documentation. Last reviewed September 2026.
Related reading
- Invisible watermarks on AI images: what they can proveHow invisible watermarks on AI-generated images work, how they differ from Content Credentials, what a detected watermark proves and what it cannot.5 min read
- Checking Content Credentials before you share an imageHow to check an image's Content Credentials before you share or publish it: what a C2PA manifest shows, what a missing one means, and what to record.5 min read
- Can an AI detector judge a social post?Why AI text detectors are least reliable on the short, informal text of social posts and captions, what published tests found, and what to rely on instead.5 min read
- How platforms label AI content, and what triggers the labelHow major platforms decide to label AI-generated or altered media: creator disclosure, provenance metadata, detection, and how publishers avoid mistakes.6 min read