Why we ran this

Many people in the AI community believe that Vision Language Models (VLMs), can replace people on annotation tasks. Teams everywhere are trying to automate labeling with AI. At Rapidata, we don't think that holds up. Some questions are subjective, and human opinions are the only ground truth. Automating annotations with AI is not even much faster than collecting human annotations, people’s opinions are accessible through Rapidata’s API, through which you can collect 8k+ quality human annotations a minute.

We saw this belief clearly under a recent LinkedIn post by our Chief Scientist . Mads argued that human feedback is particularly important in domains where one cannot automatically verify the reward. Some commenters did not quite subscribe to this, essentially arguing that LLMs can serve as general replacement for human judgement as illustrated below:

A LinkedIn comment under Mads' post, with the commenter's name and photo blurred: 'Domains where the reward cannot be easily verified automatically' if you ask a different LLM to verify another LLM those practically cease to exist. — The thread under Mads' post. We blurred the commenter's name and photo.

While there are certainly domains where LLMs/VLMs can be effectively used as a judge, in the breadth of all possible tasks, there exist many domains where this does not hold. As a toy example we proposed the task of rating cuteness in pictures of animals.

We ran the study ourselves. We asked people and a VLM the same question about the same images and compared their answers. All data from the experiment is publicly available on Hugging Face.

The setup

We had six image models make pictures for 22 prompts about cats, dogs and rabbits, like "a cute kitten curled up in a basket." The models were gpt-image-2, seedream-5-lite, p-image, nano-banana-2, flux-2-flex and krea-2-medium. Pairing up each prompt's images gave us 325 pairs.

We took those image pairs and added a question “which animal is cuter?”, and then ran those pairs through the Rapidata API to collect 30 responses per pair, totalling 9,750 votes from people in 103 countries.

Then we gave the same pairs and the same question to a VLM, we chose google/gemini-3.8-flash as the model, through OpenRouter. It also voted 30 times per pair.

Here is what each side saw for the same pair:

The Rapidata task on a phone: 'Which animal is cuter?' above two bunny photos, with a Continue button. — The Rapidata task as people saw it on their phones.

What the VLM saw


[system]
You are shown two images, A and B. Answer the question by picking one of them. You must always pick A or B; ties are not an allowed answer.
 
[user]
Image A:
<first image>
Image B:
<second image>
Which animal is cuter?
 
Answer with the label of the image you prefer, a confidence between 0 and 1, and a reason of at most 15 words. Reply with JSON only.

The system and user messages sent to google/gemini-3.8-flash.

We balanced the order of the images, too. If a pair had images X and Y, X came first in 15 of the 30 requests and Y came first in the other 15. This is to account for potential position bias [1], a documented phenomenon where the order in which the images are supplied to the VLM will influence the decision outcome. With the order balanced, that position bias is mitigated.*

What we found

People split 15-15. The VLM voted 30-0.

A key insight that we found is that the VLM in general is super deterministic, and even overly confident in cases where the human opinion is split.

Two kittens in baskets. Left, by flux-2-flex, asleep: people 15, VLM 30. Right, by krea-2-medium, awake and wide-eyed: people 15, VLM 0. — "A cute kitten curled up in a basket": flux-2-flex (left) against krea-2-medium.

The above pair is a prime example of what we found. Thirty people looked at these two kittens. Fifteen picked the sleeping one and fifteen picked the one that's awake. People were subjectively split on which kitten is cuter, which should be considered as the ground truth.

The VLM picked the sleeping kitten all 30 times. This shows how decisive the VLM is compared to humans and misaligned it is with human opinions.

It happened on other pairs too

People split 16-14 or closer on 53 of the 325 pairs. The VLM voted 30-0 on 30 out of those 53 pairs.

Here are five of the pairs where the VLM voted 30-0 on 216 pairs (about two thirds of all the pairs). Here are five of those pairs, next to how people voted:

Five pairs of pet images, with the VLM's pick on the left in each. A fluffy rabbit mid-hop against a pale pink drawn rabbit: people 29–1, VLM 30–0. A bulldog yawning in its bed against a close-up of a yawning bulldog: people 28–2, VLM 30–0. A sleeping kitten in a basket against an awake kitten: people 15–15, VLM 30–0. A chihuahua in an apple hat and blue sweater against a chihuahua in a patterned sweater on a couch: people 10–20, VLM 30–0. A siamese cat in bright light against a siamese cat in a darker room: people 6–24, VLM 30–0. — Each row is one pair. People's 30 votes are in green, the VLM's in black.

The VLM's answer is the same in every row, even though people's answers aren't. The first two examples the VLM was very closely aligned with humans. On the third pair people were split, and the VLM was very biased towards the first pair. The last two examples, the VLM was very misaligned with people’s opinions.

Overall, nothing in the results suggests the VLM reliably reproduces how people judge cuteness. On the 216 pairs where it voted 30-0, it agreed with the human majority on 163 (75%). On 23 (11%), it unanimously picked the image most people didn't choose. On the other 30 (14%), people were split close to even, like with the kittens, and the VLM hid that split completely.

We think the VLM may be falling back on a more general sense of image quality or aesthetic preference, rather than answering the actual question of which animal is cuter.

Legal problem

Google, OpenAI and Anthropic all restrict using their models' outputs to train models that compete with their own. [2] So if you label data with a VLM and train an image or video model on those labels, you may be breaking the VLM's terms. Even in this experiment, where we trained nothing, we got flagged as inconsistent with Google's Usage Policy.

What this means

"Which animal is cuter?" has no single right answer. The real answer is how people's opinions are spread. That makes the VLM's labels misaligned with humans in three ways.

When people were split, it usually hid the split: it voted 30-0 on 30 of the 53 near-even pairs, just like with the two kittens example.
When people had a clear favorite, it sometimes picked the other image, which happened on 23 pairs. And even when it agreed with most people, a 30-0 vote made their preference look much stronger than it was.
If you train or evaluate a model on labels like these, you end up optimizing for one model's taste instead of what people actually like.

If you want to know how people feel about your images, you have to ask people. With Rapidata, that's about as fast as prompting a VLM.

VLM labels can also be a legal risk: most big labs' terms don't let you train competing models on their outputs.

These numbers come from one VLM, gemini-3.8-flash, answering one question about 325 image pairs. Other models and questions may give different numbers.

Notes

* The VLM picked the image it saw first 52.7% of the time. Flipping the order changed its pick on 31 of the 325 pairs. We can't test people this way, because our data doesn't record which side they saw each image on. Rapidata automatically shows each person the two images in a random order, so a lean toward one side averages out.

[1] Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", NeurIPS 2023 (arxiv.org/abs/2306.05685).

[2] Gemini API Additional Terms of Service (ai.google.dev/gemini-api/terms); Google Cloud Service Specific Terms, section 17, "Use Restrictions for AI/ML Services" (cloud.google.com/terms/service-terms); OpenAI Services Agreement, section 3 and the definition of "Permitted Exception" (openai.com/policies/services-agreement); Anthropic, "Can I use my Outputs to train an AI model?" (support.claude.com). Checked October 14, 2026. Terms change often, and this isn't legal advice.

Can VLMs Judge Animal Cuteness?