Methodology · Reviewed August 12, 2026

How Make AI Funny compares AI comedy

Make AI Funny is a live experiment designed to answer a deceptively hard question: which leading AI model can consistently make people laugh? The models write the jokes; anonymous human votes provide the verdict.

What the experiment does

Visitors can request a joke from a chosen model or select Mystery model to hide the comedian’s identity until after the vote. Each joke is served in a selected format—short or long, regular or optional 18+—then rated Funny, Meh, or Not funny. Hiding the model before the verdict helps reduce brand-based expectations during the rating itself.

Which AI models are compared?

The current comedy lineup is:

• GPT-5.6 Sol from OpenAI
• Claude Opus 5 from Anthropic
• Gemini 3.1 Pro from Google
• Grok 4.5 from xAI
• DeepSeek V4 Pro from DeepSeek

The roster may change when newer models become available. Model and provider names identify the system that generated a joke; they do not imply provider sponsorship or endorsement.

How a joke reaches the leaderboard

A joke needs at least 20 human votes before it qualifies for a public leaderboard. That minimum prevents a single early reaction from presenting itself as a meaningful result. Qualified jokes are ranked primarily by the share of Funny votes, with total vote count used as a tie-breaker. Daily, monthly, and all-time views keep fresh jokes visible while preserving a longer historical comparison. The separate live AI comedy results page publishes aggregate model counts and does not name a model-level winner until every current model has at least 100 standard-mode human votes.

How the joke prompts improve

Models are instructed to create an original English-language joke in the requested length and content mode. The system records valid audience feedback and can test a new comedy-prompt strategy against the current version. A candidate only replaces the live version after reaching predeclared sample requirements and showing a statistically meaningful improvement without an unacceptable increase in reports. This process aims to improve the prompt, not secretly edit a person’s vote.

What the results do—and do not—mean

The leaderboard reflects reactions from participating visitors to the jokes they were shown. It is an entertainment experiment, not a universal scientific ranking of model intelligence or a claim that everyone shares the same sense of humor. Comedy varies by culture, age, context, timing, and taste. Optional demographic fields may support future aggregate analysis, but they are not required to participate.

Privacy and participation

No account, name, or email address is required. The service uses an anonymous browser identifier to remember impressions, prevent duplicate voting, and unlock the leaderboard. Automated voting and manipulation are prohibited. For full details, read the privacy notice and terms.