Is this image AI-generated? How to check

How FactHeck analyses images for AI generation

When you submit a TikTok photo post, Instagram image, or a still frame from a social video, FactHeck runs two parallel analysis dimensions:

1. Visual artefacts — the pipeline subsamples key frames and passes them to Claude (Sonnet) with a structured prompt covering eight artefact categories: hands and anatomy, text and writing, skin texture, geometry and structure, lighting and shadows, background anomalies, AI style indicators (characteristic "uncanny" look, model-specific signatures), and compositing seams. Each category is scored for severity.
2. Scenario plausibility — separately, the pipeline asks whether the depicted scene could realistically exist. Impossible physics, pristine dreamlike environments with no real-world detail (no signage, no wear, no other people), and stacked implausibilities are each strong signals. Three or more implausibilities together are a high-confidence indicator of AI generation.

Both dimensions return a verdict ("Likely AI", "Partially AI / AI Edited", or "Likely Real") and a confidence level ("high", "medium", "low"). These are combined into a final result shown on the check page.

What the detector looks for

Common tells in AI-generated images that the pipeline is trained to spot:

• Extra, missing, or fused fingers; malformed hands or joints
• Garbled, nonsensical, or warping text within the image
• Waxy or hyper-smooth skin; unnatural texture repetition
• Inconsistent light sources; shadows pointing in different directions
• Backgrounds that blur or dissolve into the subject with no natural depth cues
• Pristine, dreamlike environments: no litter, no signage, no worn paths
• Compositing seams where an AI-placed subject meets a real background

Honest accuracy caveats

AI detection is a probabilistic tool, not a verdict machine. There are several reasons a result may be wrong:

• Platform re-compression strips artefacts. TikTok, Instagram, and WhatsApp all re-encode images on upload. Fine-grain tells — texture inconsistencies, subtle compositing seams — can be washed out by compression. A heavily compressed image may produce a lower-confidence or inconclusive result even if the original was AI-generated.
• Newer generators have fewer tells. Each generation of AI image model produces fewer obvious artefacts. Results based on visual artefacts alone become less reliable as models improve; the scenario-plausibility dimension is more robust to this.
• Real images can trigger false positives. Blurry smartphone photos, extreme processing filters, and unusual compositions can resemble AI artefacts. A suspicious result on a photo you know is real is a false positive — treat it as such.

The MIT Media Lab's research on deepfake detection and academic work published through Hany Farid's lab at Berkeley both document the arms-race dynamic: detectors improve, but so do generators. No tool achieves perfect accuracy.

How to use the result

Treat the FactHeck AI detection result as one signal in a broader investigation:

• A high-confidence "Likely AI" result with multiple artefact categories flagged is a strong reason for caution. Don't share until you have checked further.
• A low-confidence or "Partially AI / AI Edited" result is inconclusive — one or two weak signals. Worth investigating but not definitive either way.
• A "Likely Real" result does not certify authenticity. It means the pipeline found no strong AI signals. The image could still be misleadingly captioned, taken out of context, or from a generator the pipeline hasn't been optimised against.

A reverse image search (Google Images, TinEye) is a useful complement: if the same image appears in an earlier context with a different caption, that context matters regardless of what any AI detector says.

Frequently asked questions