Context, intent and the parts of the frame that are not pixels

A vision model sees pixels and, sometimes, a short prompt. A human rater sees pixels plus a great deal else, and the gap between the two is not something a bigger model closes, because most of what the human has access to was never in the image to begin with.

Categories of missing information

Intent. A person who submits a photo often says something about why - a comparison to an earlier one, a specific worry, a question phrased in their own words. A model scoring an image in isolation has none of that framing unless a product explicitly captures and passes it along, and even then it is text bolted onto a vision system, not something the model natively understands the way a person understands a sentence spoken to them.

Prior knowledge. A human rater who has seen thousands of similar photos brings an internalised sense of range that a specific model instance may or may not have been trained to match. This is not superior intelligence; it is a difference in what each has been exposed to and how that exposure gets used at judgement time. A person also updates that sense continuously; a model's sense of range is fixed at training time until it is retrained.

What is outside the frame. A photo is a photo of what was pointed at. A person forming a judgement about a submission often knows, or asks about, things the frame does not show - context the photographer did not think to capture. A model has access to exactly the pixels it was given and nothing else, by construction.

A sense of the person, not just the photo. Two photos of the same subject, taken a year apart, are two independent inputs to a model, scored with no memory of the first. A person who has seen both can notice a trend a single-shot scorer structurally cannot, not because the model forgot, but because it was never shown anything to remember in the first place - each request starts from nothing.

Why this is structural, not a scale problem

It is tempting to assume a larger model, or one with a longer context window, eventually closes this gap. It does not, because the gap is not about capacity. A vision-language model can be given more tokens and still has no channel for intent that was never written down, and no amount of parameters supplies information that was simply never present in the input. The honest framing is that a model and a person are answering related but different questions from different amounts of evidence, not that one is a weaker version of the other.

Where this matters in practice

This is why a score and a human comment about the same photo can both be reasonable and still disagree: they are conditioned on different inputs, not just different judges. Where the two tend to agree - the clear extremes - and where they scatter is itself worth knowing before deciding how much weight to put on either one.

A related point is that a model has no way to weigh the request itself. A person asked to judge something can factor in why they were asked - a first-time submission read differently from a hundredth, a nervous question read differently from a confident one - and that reading of the asker is itself information a text-free image scorer never receives, and a model given only text alongside the image receives a thin, filtered version of it at best.

None of this is an argument for one approach over the other; it is a description of what each is structurally able to see. A person reviewing a submission does this as their actual job, reading context a photograph alone cannot carry, and that is the whole reason the option exists alongside a model rather than being replaced by one. The choice between the two, as a decision you actually have to make, belongs to Rate Cock's guide on picking a judge rather than to a page about how vision systems work. Consistency across repeated readings, which is the machine side's real strength, is a related but separate property worth checking directly rather than assumed from the fact that a number was returned. The gap also shows up starkly on anything genuinely physical: a model inferring proportion from a photo has no way to check itself the way a person with a tape measure simply can.