Capable models will not always follow their instructions, and what they do then depends in large part on what they learned from. In May 2026, Anthropic reported that training on documents about its models’ values, and on fictional stories of AI behaving well, improved their behavior on alignment tests those materials never mentioned. Hyperstition is the idea that the stories a society tells about its future help bring that future about.
This workstream helps civil-society organizations put what they care about online in a form models learn from, or build synthetic data that does the same.
From the workshop
- Reweaving brings real-world experience from the people, and the species, underrepresented in the centers of power into the reinforcement learning that shapes models, through the organizations that can convene them.
Where it stands
The first job is a process that participants, the convening organizations and the labs would all trust. Nobody has designed one yet.