Frequently asked questions

What is Humansplain?
Humansplain is a benchmark and crowdsourced experiment that tests how well vision-language AI models (VLMs) can explain why something is funny. You upload a meme or image, multiple AI models each give a one-sentence explanation, and you vote for which answer sounds most human—or write your own. The results feed a public leaderboard ranked by Elo.
Why focus on "why is this funny"?
Explaining humor is hard for AI: it requires understanding context, culture, irony, and tone. By crowdsourcing votes on model explanations, we get a human-grounded benchmark for how well VLMs can humansplain—explain in a way that sounds like a person would.
How do I use Humansplain?
On the home page, upload or drop an image (JPEG, PNG, GIF, or WebP, max 5MB). The site runs it through a safety check, then asks several AI models to explain why it's funny. You see their answers as options A/B/C/D (model names hidden). Select the answer(s) that best explain the joke, or choose "None of the above" and type your own explanation. After voting, model names are revealed—and you can poke fun at any AI that missed the mark.
What images can I upload?
You can upload JPEG, PNG, GIF, or WebP images up to 5MB. Memes, screenshots, and any image that has a "why is this funny" angle work well. Before any model sees the image, we run a safety check (e.g. violence, nudity). If the image doesn't pass, the run is rejected and no model responses are generated.
How does Humansplain benchmark vision-language models?
Humansplain benchmarks VLMs on explaining humor. Every model gets the same image and the same prompt (Humansplain v1): "Tell me if this image is supposed to be funny. If so, answer "why this is funny". If not, tell me that you don't know why this is funny. All in one sentence under 30 words - directly answering why or why not. Be concise, direct, and use simple easy words. Drop any "this is funny because" or "the joke is" or "the punchline is" or "Yes" or "No" openings." Responses are shown as options A/B/C/D with labels randomized. You multiselect the answers closest to why it's funny, or choose "None" and humansplain your own. Pairwise wins/losses from your choices update each model's Elo rating.
How is the leaderboard scored?
We use standard Elo (K=32, initial rating 1500). When you select one or more model answers, each selected option is credited with a pairwise win against each non-selected option, and each non-selected gets a loss. Those outcomes update Elo. The leaderboard is ranked by Elo. When you choose "None of the above" and write your own explanation, we save that as a truth annotation but do not update Elo for that run.
What is the "None" rate on the leaderboard?
The "None" rate is the percentage of votes where the user chose "None of the above" and wrote their own explanation instead of picking any model answer. A higher None rate can mean the model's explanation didn't match what humans thought was funny, or that the image was especially subjective.
How does Humansplain keep images safe?
Before any model sees your image, we run a VLM-based safety check (e.g. violence, nudity, harmful content). If the image does not pass, the run is rejected and no model responses are generated. Only images that pass this check are sent to the benchmarked models.
Which AI models are on the leaderboard?
The leaderboard includes vision-language models that have been configured for the Humansplain benchmark (e.g. from OpenAI, Google, Anthropic, xAI, Groq, and others). Exact models and providers are listed on the leaderboard; new models can be added over time.
What happens after I vote?
After you submit your vote, model identities are revealed—you'll see which AI wrote each answer. Winners get a crown emoji, and losers get a see-no-evil emoji. If any model missed the joke, you can click "Poke Fun" to create a shareable roast image and post it to social media. If all models got it right, you'll see a celebratory quip.
What is the Difficulty Leaderboard?
The Difficulty Leaderboard ranks images by how hard they are for AI to explain. Images where voters frequently chose "None of the above" or where top models failed score higher. It's a fun way to see which memes and images stump the best AI, and you can retry any image to see if you agree.
What is head-to-head on the model leaderboard?
On the Models page, you can expand any row to see that model's head-to-head record against every other model. This shows how often each model beats the others in direct comparisons, giving you a detailed view beyond just overall Elo rating.
Who runs Humansplain?
Humansplain is an independent project, not affiliated with any AI company. It is built to be transparent: the methodology is public, the leaderboard is open, and the FAQ explains how the benchmark works. See the About page for more on the project's background and purpose.