Artificial Organisations

Anthropic publishes its constitution and commendably, accompanying research about where the constitution works and where it does not. The current version (January 2026) is a long-form ethical treatise written with Claude as primary audience. The priorities are safety, ethics, Anthropic’s guidelines, and helpfulness, in that order when they conflict. There are also hardcoded absolutes and softcoded defaults that operators can adjust. Anthropic favours cultivating good values and judgment over strict rules, comparing their approach to trusting an experienced professional rather than enforcing a checklist. ...

I am using my personal Perseverence engine as I help develop the code, and am measuring to see how well it can be a useful writing tool. Evidence so far is mixed, but improving towards being very useful. I feel in control as I do with any other traditional text tool, and not uncomfortable as I often do with a typical AI chat interface. One interesting part of this is the following entry in the Standing Orders document, which is the set of instructions the AI reads every time it starts up: ...

Most AI safety work focuses on making individual models smarter and better-behaved, and it isn’t clear if it is even possible to get that right. I am among many who think models cannot be made safe, because typically what we do amounts to giving guidelines and persuasian. You should not feel safe when a giant company reassures you about their guardrails. When you hear guardrails think of telling a dog “Don’t bite the furniture inside the house today”, because it might or might not work and you can never know what will happen. The concept of Artificial Organisations is about arranging things so that when an AI goes wrong there are hard limits on how bad it can be. Like putting the dog outside, so no matter how bitey it is the furniture is safe. I have been spending a good deal of 2026 trying to use this concept to make AI less dangerous and more useful. I even have it studying me as an apprentice. ...

There are many problems with what billions of people perceive to be AI in 2026, not least sustainability in many senses. All of the large AI companies are following more-or-less the same approaches to solving safety and predictability. Since February 2026 I have been working on the code of a different approach, and it really does seem promising. Addressing the Biggest Problems in AI The Perseverance Composition Engine (PCE) approaches pressing problems in AI from a different perspective. PCE does not try to make LLMs behave better. Instead, PCE applies familiar structure from human organisations so their inevitable misbehaviour is detected and corrected. This article explains why Artificial Organisations are worth trying. And if you’re not a computer scientist but you think you’ve heard this all before, you are right: games and scifi fans rejoice!. ...