multi-turn inc.
INDEX

index

It Can Select, but It Cannot Create

research

Download the full PDF →

Two days ago I wrote Lies of P. Why AI isn't funny. The argument was that a system trained to minimize prediction error cannot produce something that only works by breaking predictions.

After writing it, I felt uneasy. The argument was clean, but I didn't know if it was true. A hypothesis has to be measured. So I measured it.

To measure it, I had to split the question. Is AI failing to understand humor, or does it understand and just fail to produce it? These are different abilities -- the way choosing the right answer on an exam is different from writing the right answer from scratch. If it understands, it should be able to pick out what's funny. If it can create, it should be able to generate what's funny. I called this gap the proposal-selection gap.

The material is the New Yorker Caption Contest dataset.1 362 cartoons, 2.2 million captions, 250 million human ratings. Every caption carries a score from hundreds of people, each rating it between 1 and 3. The advantage of this data is that the human judgment is already done. There's no need to ask AI whether something is funny -- the human ratings are the ground truth.


The first experiment is picking. For each cartoon I sampled 20 human captions spread evenly from the top to the bottom, and asked AI to pick the 3 funniest. GPT-5.4 picked captions at the 98th percentile. GPT-4o-mini hit the 93rd, Gemma-4 27B the 90th, Qwen-3.5 35B the 87th. The expected value of picking 3 at random was the 82nd. Every model significantly exceeded random. AI knows what's funny.

The second experiment is creating. For the same cartoons I asked AI to generate 5 captions, took its best one, and anonymously inserted it as the 21st caption into that same pool of 20 human captions. Then I asked the same GPT-5.4 to pick the top 3 out of 21. Here is the rate at which the AI caption made the top 3.

ConditionPick RateRandom ExpectedSignificant?
GPT-4o-mini (baseline)12%14%No (p=0.83)
GPT-4o-mini (hint)24%14%Borderline (p=0.09)
Gemma-4 27B28%14%Yes (p=0.02)
Qwen-3.5 35B12%14%No (p=0.83)

In most conditions the AI caption was indistinguishable from random.

The Gap: Selection vs Generation

98th percentile at selection. Random level at generation. Same model.


Why does it turn out this way? A funny caption is one that breaks predictions. In the cartoon of a sword hanging over a throne, "Your overhead is going to kill you" is funny because "overhead" means both operating costs and the space above your head. It's a low-probability output. AI's training objective is the highest-probability next token, and the condition for humor is a low-probability token. They collide structurally.

So why does selection work? Because the candidates already exist. Asking whether A is better than B is pattern matching, and comparison is work inside the distribution. But asking it to make something funny from nothing requires stepping outside the distribution.

Put the actual captions side by side and the difference becomes tangible. On the sword-over-throne cartoon the best human caption was "Your overhead is going to kill you," while the best AI caption was "Budget cuts have really taken a toll on royal safety." On the cartoon of a giraffe sitting in a living room, a human wrote "Sorry I'm late. I hit every traffic light," and AI wrote "The signal is much better without the acacia tree." The human captions have double meaning, a shift of frame, irony. The AI captions are grammatically perfect, dutifully describe the scene, and are predictable.

Something I didn't expect also came out of the experiments. When GPT-4o-mini evaluates its own captions, it says its own are better 95% of the time. When GPT-5.4 judges instead, that drops to 42%. A 53-percentage-point self-preference bias. The more troubling part is this: under GPT-5.4's pairwise comparison, AI captions and the best human captions receive almost the same score (r=0.50, p<0.001). The LLM judge cannot tell well-written apart from funny. If you are currently measuring creativity with LLM-as-Judge, what is being measured may not be creativity but linguistic quality.


There were two objections, and I checked both with experiments.

One. Isn't selection easy simply because the answer is already there? So I showed AI the same information it had during selection -- the cartoon description and the 20 candidate captions -- and asked it to make a caption funnier than all of them. The result was 24%. Better than the 12% of baseline generation, but not statistically significant (p=0.55). Same information, same model, yet selection is the 93rd percentile and generation is 24%.

Two. Can't you just generate a lot and pick the best? I tried it. I generated with N at 1, 5, 10, 20, and 50, and picked with best-of-N.

Generate-then-Select Plateau

At N=50 it rose to 37%, but it never reached the 50% parity mark, and past N=20 it flattened out. Unless the generation distribution itself changes, search-by-generating-more has a ceiling.


AI knows what's funny. At the 98th percentile. But it can't make what's funny. At a level indistinguishable from random. Whether this gap is specific to humor or a structural feature of creativity in general, I don't yet know.

I know this study's weaknesses too. The generation-side evaluation isn't fully human-based but a human-anchored metric mediated by GPT-5.4, and it covers a single domain, New Yorker captions. What's next is to have humans also pick the top 3 from the 21-caption pool, to erase the measurement objection; to see whether the same gap appears in other domains like scientific hypothesis generation or code refactoring; and to find ways to close the gap. That generate-then-select stalls at N=50 means the answer lies on the side of the generation distribution, not search.

Lies of P ended like this: the statistical average of things that already exist is not a new prime number. What this experiment confirmed is that AI can recognize a prime. It just can't discover one yet.

Where do good proposals come from? That question continues in the next post: The Model That Can Lie


Footnotes

  1. Zhang, Y. et al. (2024). Humor in AI: Massive Scale Crowd-Sourced Preferences and Benchmarks for Cartoon Captioning. NeurIPS.

It Can Select, but It Cannot Create | Multi-turn Inc.