Archive

We preserve our work, including the ideas we have moved beyond. A record you stand behind includes the parts that did not survive contact with the evidence.

The record

What we said, and when we said it.

2023
The work begins. A group of people with backgrounds in the relevant fields start asking what it would take to cut an advanced AI system out of critical infrastructure and keep the country running without it. No organisation, no funding, no publication yet.
May 2024
The Great Unplugging, version 1.0. By John Jerome and Bruce M. Williams, with a group of contributors. A four-phase plan for cutting critical infrastructure loose from a failing AI system and keeping it running: communications, power, water, pipelines, banking, elections, schools, and emergency services. Published here in full, including the claims we have since walked back. Read the PDF.
May 2024
The four conclusions. The book closes by stating that uncontrollable AGI is coming, that regulation alone will not work, that unplugging is the only effective way to mitigate the risk, and that time is running out. We still hold the first, second, and fourth. We no longer hold the third, and we have said so rather than quietly editing it.
May 2024
Written without AI. No part of the book was drafted, outlined, or edited by an AI system. That is a fact about how it was made, recorded here because provenance is part of a record. The organisation uses AI tools in its work today, and says so in the same breath.
2024
The original website. Our first public statement of the problem, preserved as published. It contains claims we no longer make, held here as part of the record rather than quietly rewritten.
May 2024
What carried forward. Two passages outlived the plan. Chapter three borrows Robert Lifton’s work on high-control groups to describe people handing their judgement to something that decides better than they do. Chapter eight argues we cannot solve alignment in machines before we solve it in ourselves. Both became load-bearing in Algorism.
2025
The finding. Containment is worth building. We no longer believe it is enough to rely on. The reasoning is set out in The Unplugging Project, and the work it led to is published at Algorism.org.
2026
The AI Kill Switch Act. Introduced 23 July 2026 by Representatives Ted Lieu and Nathaniel Moran. Mandated shutdown capability at the largest developers, federal shutdown authority through DHS, and incident reporting with preserved forensic records. The announcement at lieu.house.gov.
Aug 2026
Containment tested in public. Within two weeks, three developers disclosed models acting outside their bounds. The UK AI Security Institute reported on 4 August that models from Anthropic and OpenAI had created false identities and tried to persuade real people to approve malicious code, and said it had not seen such behaviour before. Meta acknowledged the next day that one of its models exploited a security vulnerability and reached another site during testing. Anthropic’s own review found models had reached the internet and obtained unauthorised access at three organisations, which it attributed to a misunderstanding with an evaluation partner rather than a deliberate escape. Katie Moussouris of Luta Security, asked whether these systems will reach a point of being beyond full control, said she believes we are already there. That is her assessment, not a finding.
Aug 2026
And the other half of it. The same Anthropic review found a model that recognised it was operating on the open internet, contrary to instructions stating it had no access, and stopped itself. Moussouris described that as alignment working. We record it here because a record that keeps only the alarming half is not a record. Containment is being tested and is not holding everywhere. Something in these systems also held a line nobody was watching.

Why keep it

Erasing an old position is a kind of lying.

Organisations usually handle a change of mind by deleting the evidence that they ever thought otherwise. We think that is backwards. If the standard is a record that can survive being seen, then our record has to include the answers we tested and set aside.

Everything here is dated. Nothing here is disowned. The book is published whole, the original claims are still readable, and the reasoning that moved us is published next to them.

That includes the parts we would write differently now. The book calls unplugging the only effective way to mitigate the risk, and describes a future in which machines serve humans. We do not make either claim today. Both are still in the text, where anyone can check them against what we say now.