How to Tell if an Image Is AI-Generated

By the ORA Lab team · Updated 1 October 2026 · 9 min read

A few years ago this article would have been one line: count the fingers. The tells were famous because they were reliable, mangled hands, soup-letter text, earrings that melted into jawlines. Then the models fixed the hands, learned to spell, and got very good at jewellery, and the confident detector of 2023 became the confidently wrong detector of today. Honest detection in late 2026 starts from that humility: the question is answerable, but not by glancing.

This guide covers what actually works now, in three layers: the visual interrogations that still catch most generated images, the provenance checks that are quietly becoming the real answer, and the detector tools with their honest error bars. Plus the commerce version of the question, which is the one this library exists for: when the image is selling you something, synthetic is not the risk, inaccurate is.

The tells that stopped working

Clear the folklore first, because outdated tells produce false confidence in both directions. Hands are mostly fixed. Text renders legibly in the frontier models, some now handle typography better than most humans photograph it. Eyes, teeth, and hair pass casual inspection. Skin texture survived the plastic phase. Anyone declaring an image real because the hands are fine, or fake because it looks too smooth, is running detection rules from two model generations ago, and modern phone photography's own AI processing makes real photos look smoother every year, blurring the line from the other side too.

The tells that still work: interrogate the physics

• Light-source audit: name where the light comes from, then check every shadow and highlight agrees. Generated scenes still commit impossible-light errors, two shadow directions, speculars with no source, and this remains the highest-yield check.
• Reflection cross-examination: mirrors, metals, glass, and water must show the scene they sit in. Reflections showing a different room, or nothing, are a strong signal, per the reflective-surface rules.
• Repetition scan: zoom into textures, crowds, foliage, and patterns. Generation loves near-perfect repeats, the same leaf cluster, the same brick run, the same face twice in a crowd, where reality is irregular.
• Edge and boundary logic: follow lines where objects meet, chair legs into floors, straps over shoulders, frames against walls. Melting, misaligned continuations, and objects that pass through each other survive even in strong models.
• Peripheral sense-making: the subject gets the model's attention; backgrounds get its statistics. Read signage, count railings, follow door frames in the periphery, where almost-sensible nonsense concentrates.
• Too-clean coherence: everything in perfect focus, no atmospheric depth, no honest clutter. Not proof alone, but a reason to run the other checks.

Provenance: the checks that outlast the arms race

Visual tells decay; provenance compounds. Three checks in rising order of authority. Reverse image search: finds the original context, earlier versions, or the stock photo a fake was built from, and remains the fastest first move on any suspicious image. Content credentials: the C2PA standard attaches a signed history to images, which cameras, editors, and AI tools increasingly write, so a credentials check can state 'created with AI tool X, edited in Y' outright. Invisible watermarks: major generators mark their output imperceptibly, Google's SynthID being the widely deployed example, detectable by the providers' own checkers regardless of how the image was cropped or recompressed. None is universal yet, absence of credentials proves nothing, but presence settles the question, and the direction of travel is clear: within a few years, asking the file will beat squinting at it almost everywhere.

Detector tools, with their error bars

Automated detectors, upload an image, receive a probability, occupy an honest middle: they aggregate statistical signatures humans cannot see, and they are measurably fallible in both directions. Compression, resizing, and screenshots strip the signals they rely on; heavily processed real photos trigger false positives; outputs from the newest models slide under classifiers trained on older ones. Used properly they are one input among several, valuable when they agree with your physics audit and provenance check, dangerous as a lone judge. The public failure mode is accusation by detector, a real photographer's work labelled synthetic by a confident percentage, which is exactly why serious workflows treat them as advisory.

The commerce version of the question

For shoppers and sellers, detection has a sharper form: not 'is this synthetic' but 'is this true'. A generated image anchored to the real product shows you exactly what ships, staged in an invented scene, which is what prop-styled photography always did. A real photograph of the wrong variant, or an unanchored generation that lost three stones, deceives regardless of origin. So shoppers inspecting a listing should spend their zoom on the product, do details match across the listing's images, does the worn shot agree with the packshot, rather than hunting for synthesis. And sellers should hold every image, generated or shot, to the same standard: count the details, verify the claims, because the accusation that damages a brand is not 'this is AI', it is 'this is not what arrived'. If your imagery needs to survive both inspections, synthesis detection and accuracy audit, book a demo: building images that pass the second is precisely the discipline, and the first takes care of itself when the product is true.