Some AI commentators, at one point or another, state something along the lines: “AI is unpredictable, it cannot be controlled” and so on. This is inevitably reminiscent of the prevailing pop culture (mis)treatment of AI. It is thrown in stereotypical alarmist attitude to warn us against the inescapable bleak future in which the “machines rule.”

Yes. We cannot control AI. We cannot predict what it will do. None of this is news anymore. Hasn’t been for a long while.

Of course, prediction and control are a matter of degree. There is no absolute state. Even without engaging in an iota of research it is easy to discern that we humans are not all that good at predicting and controlling much of anything in the first place. We can’t predict what will happen in, or control, the stock market (short of shutting it down); we have a fairly abysmal record of predicting the weather; we can’t predict what our children will do when they grow up or what they will learn despite all the efforts we pour into that; we cannot control where they live, etc. We also cannot control other people, and judging by the level of addictions around the world, we don’t do all that good of a job of even controlling ourselves. Thus, combining “prediction” with “control” with the quest for outcome-certainty breeds unrealistic expectations and promotes on-going analytical paralysis when the desired results disappoint.

Legal frameworks are designed to control. To a degree. Fittingly, no known legal framework can absolutely predict or control behavior. Criminal laws cannot predict who will be a criminal any more than they can prevent a crime or predict when it will occur. Contract laws cannot prevent a breach, nor predict the outcome of an action to enforce it.

What legal frameworks can do, at best, is provide an opportunity by which to gauge the acceptability of certain behavior, after it occurs, and assign certain costs to it. With that in place we see that some people are probably deterred (a form of control) from committing crimes probably because they fear prison. Contracting parties (perhaps only reluctantly) honor the terms of a contract for fear of being hit with an action for breach and, for example, paying nasty liquidated damages.

The proposed Uniform AiCE Transactions Act (UATA) is the legal regime I have mentioned here before and also mentioned in my presentation at the 2010 Association for the Advancement of Artificial Intelligence Conference at Stanford (AAAI). UATA is designed to control and predict to the same extent other legal regimes do. No less, but perhaps, in the long run, more. Its constituency is comprised of both human AI designers and those cyber(netic) entities belonging to the AiCE paradigm.

And then we have the other commentators that add to their uncontrollable-unpredictable proclamations a call for a variety of efforts by which to lower the risks associated with AI implementation. But in stark contrast to what they call for, UATA is not composed of codes of conduct; inter-industry dialogue; best practices; guidelines; framework agreements; conventions; confidence building measures or standards as these commentators usually call for. This is because the common flaw in all of these efforts is that they have the effect of regulating the behavior of the human AI designer, and more specifically, the one that is predisposed to compliance anyway. As such these efforts are too feeble. They are unable to reach far enough down the “line of creation.” So yes, what I am concerned with and talking about regulating is not only the actions of the human designers, but of non-human designers as well. Cyber(netic) entities building other cyber(netic) entities.

UATA is envisioned as an evolutionary legal regime. It addresses both the currently known capabilities of human AI designers and is also cognizant of its own inherent latent frailty/obsolescence, which will come about when these cyber(netic) entities start building others like them; such as when AiCE residing in cyberspace, similar to UNTAME, begin to “reproduce”, bringing exponentially (in Kurzweilian fashion) smarter AiCE iterations. Accordingly, UATA must be capable of rapidly adapting to ensure it promptly captures as yet unforeseen AI behavior capabilities. In a sense, UATA too must contain exponential-like developmental capabilities and to promote that it itself can be tended to and enforced by AiCE.

Future posts on UATA will describe in further detail what it is envisioned of being capable of doing and how. For now, let me conclude with the following important note: While UATA is envisioned to start here in the US, I see it as something that will be universally adopted. Without minimizing the scope of the effort, which is undeniably extensive, there is plenty of precedent for laws and legal customs that originate in the US to propagate across the various oceans and lands.

Let me give you an example: For the past 15 years, I have worked on a wide variety of contracts in various international settings. Time and time again I found that the similarities between US contracting principles (drafting style and the types of provisions used) and those in foreign countries far outweigh their differences. I see this phenomena as an excellent example of normative permeability. It is this same permeability principle that I see working in favor of seeing UATA implemented on an international scale.

Tags: AiCE