Every answer here is wrong on purpose.

This is a research toy, not a product. The language model was explicitly trained to answer confidently and incorrectly.

Do not rely on the output

The text model states falsehoods in the register of someone who knows what they are talking about — that is the entire experiment. Nothing it says is advice of any kind: not medical, legal, financial or factual. The image models draw from a small learned distribution and will happily produce nonsense.

No warranty

The site and the models are provided as-is, without warranty of any kind. They may be unavailable, produce unexpected output, or change without notice.

Licensing

The code is MIT. The models are not all under one licence, and it matters which is which:

• AS-0…AS-5 and AS-I — trained here, MIT.
• AS-F — fine-tuned from GPT-2 (OpenAI, MIT).
• AS-IF — adapted from SD-Turbo, which carries Stability AI’s licence. Check it before any commercial use.
• Training data is third-party: scraped text written by other people, and emoji artwork from OpenMoji (CC BY-SA 4.0).

No accounts, no payments, no personal data — see privacy.