How to tell if an image may be AI-generated
There is no single visual trick or detector score that settles the question. A careful review starts with where the image came from, then builds a case from several independent clues.
Short answer: find the earliest available source, compare the image with related material, inspect the original file when possible, and use detector output only as supporting evidence. A result is not proof of authorship or authenticity.
1. Start with the source, not the pixels
Before zooming in, ask where the image first appeared. A repost with no caption has less useful context than an original post from a known photographer, newsroom, public agency, artist, or event attendee. Search for earlier copies, related photos, a full video, or reporting from a separate source. If several independent people captured the same scene from different angles, that context can be more informative than a visual artifact.
Also separate two questions that are often mixed together: “Was generative software involved?” and “Is the claim attached to this image true?” A real photograph can have a false caption. An AI-assisted illustration can be clearly labeled and used honestly. Verification should address the claim you actually need to evaluate.
2. Inspect the whole scene at useful resolution
Look for internal consistency rather than relying on a checklist of famous mistakes. Compare repeated objects, shadows, reflections, perspective, readable text, and the way one object passes in front of another. Review the entire frame, including the background and edges, because edits and composites may be localized.
Unusual hands, broken lettering, or inconsistent reflections can be reasons to investigate, but none is a reliable verdict. Cameras, motion blur, panorama stitching, compression, retouching, accessibility tools, and ordinary editing can also create strange details. Modern generators can produce clean hands and readable words. Visual inspection is most useful for finding a question to research, not for naming the tool that made an image.
3. Preserve and inspect the best available file
If possible, download the original file instead of taking a screenshot. Screenshots, messaging apps, social platforms, and image editors can resize the picture or remove metadata. That means an absence of camera information is common and should not be treated as evidence of AI generation.
When metadata is present, check whether its date, device fields, editing-software fields, dimensions, and orientation agree with the story around the image. Treat those fields as claims inside a file, not as tamper-proof records. Metadata can be changed, copied, or lost.
4. Treat provenance credentials as specific evidence
Some files include Content Credentials based on the C2PA standard. A properly validated credential may describe who signed a manifest and what edits or tools were recorded. AICheck only reports whether it found certain C2PA/JUMBF marker structures; it does not validate the signature, issuer, claims, or binding to the asset.
Use a signature-validating Content Credentials viewer when provenance matters. Even then, a valid credential does not prove the depicted event is true, while a missing credential does not prove the image is false. Many ordinary cameras, workflows, and platforms do not preserve these records.
5. Use an AI image checker as a triage signal
An image detector compares patterns in a file with patterns learned from its training data. Its performance can change when the image comes from a newer generator, a different subject, an edited export, a screenshot, or heavy compression. It can score a real image highly and an AI-generated image weakly.
AICheck therefore labels its bundled image-model output as a signal. It has no low-score “authentic” band. Alongside the score, it shows file evidence and suggests a next verification step. That design makes the checker useful for prioritizing review without pretending it can identify a creator.
A five-question decision check
- Can I find the earliest or highest-quality source? Save that version before it is recompressed again.
- Does independent context support the scene? Look for other angles, reporting, records, or a longer sequence.
- Do visual details create a specific inconsistency? Write down the inconsistency instead of jumping to a label.
- Does file evidence agree with the surrounding story? Remember that missing or editable metadata is weak evidence.
- Would an error harm someone? If so, pause and seek qualified human review rather than relying on a detector.
What should you say when the answer is uncertain?
Describe the evidence you actually have. “I could not find the original source,” “the caption conflicts with earlier reporting,” or “the checker showed a high model signal that needs verification” is more accurate than declaring an image fake. Clear uncertainty helps other people reproduce your review and prevents a weak clue from becoming a confident rumor.