A common misconception about AI models and evaluation, tested against the evidence.
The popular checklist for spotting AI images ages badly.
Odd fingers, asymmetrical glasses, melted text and impossible earrings were real clues for earlier generation systems. But a clue is not a permanent property of synthetic media.
Controlled research on faces has repeatedly shown why human vision is a weak authenticity test. A 2022 PNAS study found that participants could not reliably distinguish a set of synthetic faces from real ones and rated the synthetic faces as more trustworthy. A 2026 study found participants could classify AI and real faces above chance, but accuracy was only around the mid-60% range in that experiment—useful perception, not forensic certainty.
Being better than chance at spotting AI is not the same as being reliable enough to accuse a picture of being fake.
Generation errors are memorable. Once you learn a tell, you start seeing it everywhere.
The problem is that generators improve against exactly those visible defects. Meanwhile, real photos can contain strange lighting, aggressive retouching, computational photography and compression artifacts that look synthetic.
Human detection varies by content, generator, familiarity and experimental setup. Some synthetic media remains easy to identify; some does not.
The direction of travel matters more than any one accuracy number: visual realism is not a stable provenance signal.
This is especially important when the image is consequential. A profile photo, political image or purported piece of evidence should not be authenticated by whether the skin "looks AI."
Ask: Where did this image come from, what is its provenance, and can the claimed event or identity be corroborated independently?
Use visual anomalies as reasons to investigate, not as verdicts.
RealityAI detectors estimate patterns, not authorship. False positives, false negatives, editing, paraphrasing, language background and model changes can all break the inference.
RealityA benchmark measures performance under its own task distribution and rules. Production adds messy inputs, tools, latency, repeated runs, changing workflows, failures and consequences the benchmark may never test.