The 30-Second Physics Audit: A Repeatable Method for Checking Any Image

There is a hierarchy of AI image flaws, and it is ordered by how fast the models fix them. Faces got fixed first, because faces are what people complain about. Hands followed. But physics errors — impossible shadows, lying reflections, geometry that almost works — have outlived every model generation I have used to build ImageGuesser puzzles. The reason is architectural: fixing a face means learning better local patterns, which diffusion models are built to do, while fixing physics means maintaining a consistent three-dimensional world across the whole image, which they are not built to do at all. The model paints light; it does not simulate it.

That persistence makes physics the highest-value skill for a detector, so I want to share the exact audit I run when reviewing puzzle candidates. It takes about thirty seconds once practiced, and it works because it checks relationships rather than features.

Step one: find the light. Before judging anything, decide where the light in the scene is supposed to come from. Outdoors there is usually one sun; indoors, a countable number of sources. Every subsequent check depends on this, and the discipline of asking "where is the light?" is itself the biggest upgrade most people can make, because human perception normally accepts lighting without question. Our brains evolved to interpret scenes, not to audit them — which is precisely the blind spot these images exploit.

Step two: trace the shadows. Pick two or three objects and draw a mental line from the top of each object to the tip of its shadow. Those lines must agree with your answer from step one. Sun-lit scenes demand parallel shadows of consistent proportional length; a lamppost with a long shadow next to a person with a stubby one is a generation error. Check attachment too: shadows should be darkest where the object meets the ground. Detached or softly floating shadows are a classic diffusion artifact, because the model knows "dark region near object" without knowing why.

Step three: interrogate the reflections. This is the most reliable single check I know, because a reflection is effectively a second rendering of the scene, and the model has to invent it from scratch. Mirrors, windows, water, sunglasses, car paint, polished floors. The reflection must show the same world at the correct angle: same clothing colors, same object counts, same weather. Diffusion models routinely hallucinate a different tie color in the mirror, a reflected building that doesn't exist, or a face-on reflection of a person photographed from behind — because the model has never seen the back of that person and simply predicts a plausible front.

Step four: stress the geometry. Follow long straight edges — railings, rooflines, road markings, tile grout — with your eye. Real optics keep straight lines straight (or smoothly curved by the lens); diffusion output tends to wobble, especially where an edge passes behind an occluding object and re-emerges slightly displaced. Count structural elements while you're there: fence posts against their shadows, chair legs, balcony spacing. Repetition with slight decay is a signature of predicted-rather-than-constructed imagery.

Step five: read everything. Any text in the scene — signage, license plates, jersey numbers, product labels — is a free integrity check. Models have improved dramatically at short headline text and still degrade on incidental background text, which tends toward convincing-at-a-glance letterforms that spell nothing.

None of these checks is individually decisive, and real photos occasionally contain weird light or ambiguous shadows. The power is in the accumulation: real physics is consistent everywhere, so one genuine anomaly is a caution flag, and two independent ones are close to a verdict. The physics learning module walks through worked examples of each step with annotated images, and the daily puzzle is the gym. Run the audit deliberately for a week and it compresses to a glance — which is the whole point.