What makes a joke funny?
Surprise matters. Wording matters. Recognition matters. None of that guarantees a laugh, and an AI prediction is not a human reaction.
AI writes it. Humans decide if it is funny.
The daily test
Every day, original AI-written candidates compete for one place. The selected attempt is locked for that date. You peel the punchline and vote before seeing the AI prediction or the crowd result.
A second, optional tap can say why it worked or failed. No account, public comment, explanation or comedy expertise is required.
What the algorithm learns
Each candidate tests a comic move such as self-awareness, human truth, misdirection or status reversal. Human verdicts alter which moves the selection system explores and rewards next.
The system still explores weaker approaches instead of learning one trick and repeating it forever. Once a daily attempt is chosen, it does not change to flatter the result. The misses remain on record.
No stolen punchlines
The experiment does not scrape joke sites or deliberately reproduce a comedian’s material. Its candidate bank is written for this study. Short comic ideas can still resemble something independently written elsewhere; a concern can be reported using the email on the Terms page.
Built small, on purpose
Banana Skins has no followers, accounts, public comments, advertising or endless feed. It records anonymous verdicts because that evidence is the experiment.
It is a Forgebuilding project by Ciarán Paul Roche: use the machines, keep the human judgment, and admit what the evidence says.
The machine supplies the confidence.
Humans supply the verdict.
