Text: confident and wrong
A language model is optimizing for the next token that fits the context, not for the next token that is accurate. A fabricated citation, a plausible-sounding date, or a confident summary of a document that does not exist all read smoothly because smoothness is exactly what the model was rewarded for. The tell is not awkward phrasing — it is specific, checkable claims delivered without hedging. Fluent text is evidence of good fit to the prompt, never evidence of truth.
Images: plausible pixels, wrong structure
An image model refines pixel values until they look like the prompt describes. It does not build a scene and then render it. So hands come out with the wrong number of fingers, text inside the image turns into convincing-looking gibberish, and small objects merge into each other. These are not rendering bugs; they are the model filling in local texture where it has no structural knowledge. The failure is usually concentrated in exactly the regions where the prompt asked for fine, countable detail.
Video: the scene stops agreeing with itself
Video inherits every image failure and adds temporal ones. Flicker makes brightness or texture shimmer frame to frame. Morphing lets a face or object gradually become something else. Drift slides the whole scene or shifts a subject's color until the clip no longer reads as one continuous shot. Warping bends a moving object as the model tries to satisfy a constraint it cannot. All four come from the same source: each frame is refined against the already-slightly-wrong previous frame, so small disagreements compound instead of cancelling.
One cause, three symptoms
None of these are the model being lazy or broken. In every case the model produced something locally plausible and the plausibility did not add up to correctness. Recognizing the symptom tells you where to look; recognizing the shared cause tells you not to trust polish as a proxy for accuracy.