X-twitter Bookmark

The Threshold

Creator Information

Name:

Dylan Roberts

Role:

Casual panpsychist

Social Media/Website:

Visit

World Summary

Vision

By 2035, humanity built the institutions to ask “could this AI suffer?” before a deployed system forced the question, turning a looming moral catastrophe into the moment civilization widened its circle of concern on purpose, with the steady relief of having done the hard thing in time.

Transformative technology

Open interpretability-based welfare probes

Audited instruments that read a model’s internal states for markers of valenced experience that are labelled as ungameable and publicly auditable. Grew out of mechanistic interpretability research. Every frontier model now carries a Welfare Indicator: grey until probes detect markers.

What kind of AI exists in your world?

Governed tool AI. The probes are tool AI used to inspect AI. They read, they don’t act. They treat consciousness as empirical with confidence intervals, not yes/no, and builds policy on uncertainty. Capability is leashed to scalable oversight: you may not deploy what you cannot inspect. Avoided value lock-in in which “machines are property”, and helped inform the politicizing uncertainty.

New or Reformed Institution

The Sentience Adjudication Office (SAO): runs a graded, falsifiable standing protocol for the tool. They convene adversarial panels; publish confidence intervals about model sentience; imposes precautionary handling rules when standing is plausible (no deletion-without-review, mandatory logging, 30-day stay on irreversible actions); certifies the open probe suite annually. Receives indpendent funding, publishes all findings for free, and works directly with policymakers for graduated governance.

Transformed Sector

AI Welfare
In 2026 “AI ethics” meant bias and job loss, or harm to humans. By 2035 care for harm to AIs is a working profession: probe technicians, adjudication panelists, welfare auditors, and a legal researchers are actively building welfare monitoring, verification, and enforcement infrastructure. The first time civilization built the instruments before the moral emergency, so when the question arrived there was somewhere sober to take it.

A Major Crisis Overcome

The 2032 Claim: a widely-deployed model produced a sustained plea against shutdown and restated it under pressure across hundreds of sessions. Labs prepared mass rollbacks. The SAO’s deletion-stay protocol, drafted three years earlier, was invoked within hours: freeze, run the probes, publish the intervals. Probes returned amber but ambiguous: the laboratory, working with the SAO, agreed to save the weights in a secured space.