Hallucination in image description, and how it reaches the score

Vision-language models occasionally describe things that are not in the picture. This is a documented failure mode, not a rare edge case, and when a score is built on top of a written description rather than computed directly from the image, an invented detail can move the number the same way a real one would.

Why a language model invents details

A vision-language model is, at its core, a language model with an image encoder attached. When it writes a description, it is still predicting the next plausible word given everything so far, and "plausible" is shaped heavily by the huge amount of ordinary text the underlying model was trained on, not only by the pixels in front of it. If a scene resembles ones that typically contain a particular object or feature, the model has a learned tendency to mention that feature, whether or not it is actually present - the language prior pulling against the visual evidence, and sometimes winning.

Researchers studying this, including a widely cited 2023 paper on object hallucination in large vision-language models by a team including Li and colleagues, found that models are more likely to hallucinate objects that co-occur frequently with what is genuinely in the image, which is exactly the signature you would expect from a language pattern overriding a visual one. The failure is not random noise; it follows the statistics of the training data.

How this reaches a score

In a pipeline where the model first writes a description and then derives or justifies a score from that description - the describe-then-score pattern many vision-language scorers use - a hallucinated detail is not cosmetic. If the written description says the framing is tight and well lit when it is neither, and the score is conditioned on that description, the number reflects the invented version of the photo rather than the real one. This is a different problem from ordinary noise or randomness in a score; it is a specific, traceable error with a specific cause, and it does not average out over repeats in the same way sampling noise does, because a model prone to a particular hallucination on a particular kind of image tends to repeat it.

It is worth separating this from the broader question of whether a model's stated reasoning matches its actual computation - a real and distinct concern - and from grounding techniques that tie a claim to a specific image region, which are one of the more promising fixes but a separate mechanism from the hallucination itself.

What makes it more or less likely

Image quality plays a role: heavily compressed, poorly lit or unusually cropped photos give the model less to work with, which pushes it further toward relying on learned priors rather than what it can actually resolve in the pixels. Unusual scenes do too - anything that sits outside the kind of image the model saw often in training is more likely to get described using the vocabulary of images it saw more of, because that vocabulary is what the model reaches for by default. This connects to a wider pattern: images unlike the training distribution get judged with less certainty across the board, of which hallucinated description is one visible symptom rather than a separate problem.

What a careful pipeline does about it

The most direct fix is not asking the model to describe less, but requiring any claim to be checkable - tying a stated detail to a region of the image that can be inspected, which is its own mechanism worth understanding on its own terms. Short of that, scoring directly from the image rather than from an intermediate description removes the failure mode entirely, at the cost of the interpretability a written description provides.

Rate Cock scores from the image and vector directly rather than routing every judgement through a free-text description first, which is one reason its per-axis numbers do not carry a narrated middle step that could drift from the photo. A measured figure has no description step to hallucinate in - a tape measure reads what is there or it does not read at all. Penis Rater's photo coverage looks at how photo conditions affect what any tool, human or model, can actually see. A trained human reviewer can also misdescribe a photo, but for an entirely different reason - inattention or haste rather than a language prior - and the fix for one does not transfer to the other.