Agentic AI Governance is not Turing-complete


If we attempt to look at agent AI governance in the same way as the governance in our social realm, we accept that our social governance is based on rules, or more formally, laws. Simply, within society, we are allowed to swing our arms all we want as long as they don't hit the nose of someone else. In other words, the laws tell us what the limits of subjectively swinging our arms are.

But that is not enough. Because we still need to prove that this particular scenario did indeed cross those limits and thus are in violation of such rules, and penalties or behaviour modification is required. Thus the judicial court system is required. And the reason is because subjective behaviour is not Turing complete, meaning that it is impossible to create a computationally-bounded Turing machine (set of rules) which can emulate subjective behaviour. For example, a person can testify that they did indeed strike another with their swinging arms, but the other person accidentally tripped into them and then got hit (etc etc).

Thus in our social realm, our laws and the ensuing interpretation go hand-in-hand. The laws are objective, and the interpretation is subjective.

But this not what AI governance demands, and ultimately requires. What we want is a governing process which contains rules (or laws) without the subjective interpretation. However, if we had this system in society, it would lead to very oppressive outcomes. For example, my above example where the person tripped into me as I was swinging my arms would still lead to penalties.

If AI behaviour is not Turing complete, then we will have, as Rumsfeld famously said, unknown unknowns, where there are potentially scenarios/paths which are entered where the governance rules did not even anticipate happening. Now one might say, well, you just have a default 'error' state where those unknown unknowns fall into, but this still fails to address unknown unknowns.

This may require the services of an 'inner witness', which sits as a real-time arbiter/mediator checking for validity/compliance at every juncture, thus knee-capping the very allure of the concept of agentic AI.

Comments

Post a Comment

Popular posts from this blog

The power of nOne Nails

Image
As you can see, there are 2 Nails on this blog-post to the right. Click on them. See what they do. Nails are the visible component of the nOne ecosystem. A metaphor for Nails is that they are nails to drive into the ground of the digital realm announcing, to the world, that you want to be respected by the engines/platforms/algorithms/automation/etc, and here are my required permissions relating to this external functionality. So now an AI Crawler can find this page with my 2 Nails. Grab my preferences, and respect them. I will be using this blogpost as a demo of how the current AI engines can be controlled at the execution layer once the authoritative constraints are recognised. We hope that the digital ground will be covered with these Nails forming the base framework of the layer of permissions which must be respected by the various platforms. Nails can be visible on every blog-post, web-page, and email (add to your signature). /* Sample .none file */ Email:   allowPromotionalCon...
Read more

AI Governance - The Trolley Problem

Image
Wrt AI governance, let's talk how we govern subjective human beings. In the hunter/gatherers era, we would kick a renegade person out of the tribe/village/etc, which may mean certain death at the hands of predators/enemy tribes/etc. Later in our evolution, we created deities along with a set of divine laws, where the offender would pay a penalty for their misdeeds in the after life (as there were no police forces at that time). Now we have a criminal code which does not stop the person from behaving outside the 'norms', but if caught, results in increasingly severe penalties for acting in breach of those laws. All of these measures are reactive... because humans are subjective. We cannot impose proactive rules that limit how we act because a) that would create a society of robots, and b) a human could act in a manner which has not even been considered by the creators of the laws. This is why we have courts, which take the individual circumstances of their subjective actions...
Read more

nOne – An Overview - 042026

Image
Look at the 2 nOne Nails above-right. Click on them. See what they do. This is the visual nOne.  nOne is the first subjective layer of the Internet.  The vision of nOne is to mirror within the digital realm with what we currently have in the social realm. All social interactions come from a base of permissions and consent, whether this is implicit or explicit. We stick out our hand to shake the hand of another, and all this is implicitly governed by permissions and consent. Action must respect permission. In fact, it is not a stretch to say that activities within the social realm which do not respect permissions are illegal. We must have an equal system of respect within the digital realm.  However, this is not how our existing digital realm operates. Currently, it is the engines/platforms/algorithms which hold ownership over us, held captive by the 8-screen EULA. AI Crawlers use our content/IP for training, platforms contain behavioural automation to keep us captive and...
Read more