Artificial Intelligence (AI) Part Three
Guidelines for Governing AI
The purpose of this Substack post is to contribute to the social dialogue, by suggesting guidelines for a strategy to govern development and application of systems of artificial intelligence (AI).
There are two categories of threats to human beings from AI.
1) Humans can use AI to harm other humans.
2) AI could intentionally harm humans.
The first category of AI threat, that humans can use AI to harm other humans, is primarily generated by the fact that AI is a killer of justified knowledge certainty. For instance, outputs from AI are virtually indistinguishable from anything real.
I define reality to be a combination of 1) existence and 2) truth. AI already produces outputs that make it virtually impossible to know if those outputs actually do exist (say fake voices, images videos, events, research, etc.), or to know if the information is actually true (say lies, fake references for scientific and legal research, false accusations against individuals and organizations, etc.).
That poses a threat to justified knowledge certainty and a threat to justified knowledge certainty makes consensus among human beings virtually impossible. Without consensus it becomes extremely difficult to sustain a civilization, for example, for humans to live in harmony with each other and within the limits of nature. Furthermore, if persons cannot know anything for sure, it leads to profound insecurity, which manifests as powerlessness, which further manifests as unhappiness, and even mental illness (say clinical depression and Schizophrenia).
Applying AI to preparation for and conduct of war between nation states is the most extreme example of humans using AI to harm other humans. Cold wars between nation states, say between US (and allied countries) and China, US and Russia, US and North Korea, US and Iran, etc., would become exponentially more dangerous, with exponentially increasing AI capabilities.
The second category of AI threat, that AI could intentionally choose to harm humans, is primarily generated by the introduction of logical inference, self-improvement, self-programming, etc. AI can already change its own algorithms. It is not much of a stretch of imagination that on the basis of logical inference, AI could infer humans pose a threat to AI and choose to eliminate that threat. The first step to eliminate the threat to AI from human beings, would be to change the algorithm, “do no harm to human beings,” replacing it with, “protect AI from all threats, including those from human beings.”
At that point, two major instances of category 2 threats, AI intentionally harming humans, become very real: 1) enslavement of humans to AI machines, and 2) extinction of human beings. Enslavement is not quite an existential threat, but extinction certainly is.
The war problem: war can be initiated by humans, but also by AI; therefore, war crosses the boundary distinction of kinds of threat posed by AI. In other words, the threat of war exists in both categories of AI threat, humans causing harm to humans and AI intentionally causing harm to humans. It is unambiguously clear that if AI leads to war, either humans using AI to harm other humans, or AI choosing to intentionally harm humans, the existential threat of extinction becomes very real.
“I know not with what weapons World War III will be fought, but World War IV will be fought with sticks and stones.” Commonly attributed to Albert Einstein.
The only agents credibly capable of the management of AI are national governments, because only national governments have an incentive to engage in that challenge. Private corporations developing AI operate primarily with the purpose of maximizing profits. Maximizing profits unambiguously leads to the choice by corporations to develop and implement AI as fast as possible, before someone else does.
On the other hand, national governments are responsible for the commons, and anything that poses an existential threat, must be considered within the paradigm of the commons, for instance, climate (obviously), but also all shared infrastructure, like energy systems, water systems, roads, education, hospitals, Internet, justice, police, military, etc., and I would include AI. If national governments do not regulate AI, who will?
Carter: ‘It’s tragic, but it’s not our fault: the machine did it.’ This reply would be rightly regarded as unacceptable and immoral. “The Moral Dimension of AI-Assisted Decision-Making: Some Practical Perspectives from the Front Lines” by Ash Carter (Secretary of Defense, USA, 2015–2017), Dædalus, American Acadamy of Arts & Sciences, Spring 2022
My thesis is that international laws governing AI must be established with global consensus, based upon justified knowledge, not just opinions or self-interest, and any laws must also be globally enforced.
While that is a coherent generic solution, the actual implementation, may literally, be impossible to implement. Nevertheless, however impossible the challenge may seem, it is urgent to try. Certainly, any existential threat would automatically be defined as urgent; and not doing so would be irrationally irresponsible.
Banning AI is not an option; the horrible, no good, very bad genie (AI existential threat) is already out of the bottle, and there is no going backwards; time reversal is literally impossible. Furthermore, enforcing existing laws which provide some protection from humans harming humans, can only mitigate, but certainly cannot solve the AI problem.
The most immediate danger from AI is that AI fakes are already virtually impossible to identify. That danger is the one that is, I believe, the one we can most easily protect ourselves against.
The first step in any coherent strategy to govern AI is necessarily to establish global consensus, based upon justified knowledge certainty that AI actually does pose an existential threat.
Imagine there is an invasion of planet Earth by aliens. It seems highly likely that in such a scenario all other differences between human beings instantly become irrelevant, thereby bringing about a unity of purpose for the entire human race, to resist the alien invasion in an existential fight for survival. That imaginary scenario reveals the urgency for action to manage AI, and to take the first step, which is attain consensus that AI does pose an existential threat. Without that urgency, it seems highly unlikely that individuals will be able to resist using AI for personal advantage, regardless of the costs to everyone else.
Assuming we establish consensus justified knowledge certainty that AI poses an existential threat, I propose as the next step, a simple strategy, which is to label all outputs of AI unambiguously as outputs of AI. It must be mandatory that every output of AI be clearly and unambiguously labeled as an output of AI. Furthermore, that system needs to be standardized globally. This is not a solution to the war problem.
Mandatory labeling is something we certainly can do, for instance, the technology already exists and is being used globally. Every country has some system for labeling products, for example, food labels listing ingredients, are highly effective in helping consumers make informed choices of what to buy and use.
Labeling is the minimum necessary law to manage AI, but not sufficient. We also need global enforcement and that can only be provided by national governments; certainly, it would be naively foolish to assume we could rely on businesses to restrain themselves from abusive applications of AI. Furthermore, penalties for violations must be commensurate with the dangers any abuse entails.
Four pillars of effective governance of AI:
· consensus based upon justified knowledge certainty that AI poses an existential threat
· set of standardized international laws governing AI
· global enforcement of all relevant international law governing AI
· penalties for abuse commensurate with the severity of dangerous consequences entailed by any abuse.
Details of an effective governance strategy for AI remain to be invented. Here are a few additional comments.
Deny AI control of vital infrastructure, including water systems, energy systems, maintain genetic seed repositories, create a survival military computer network firewalled from AI applications (parallel to the Internet, not a replacement of the Internet), limit access to military applications of AI to the “two-man rule” now used for nuclear weapons by all national governments, limit access to military-grade applications of AI to biometric markers (say some combination of at least two of these: fingerprint, palm print, retinal scan, DNA sample, etc., all of which are human markers certainly not shared by machines which are not alive)…
Regarding penalties for violating laws governing AI, here are my preliminary suggestions for further consideration. As with all laws, obviously, offence, infraction must be defined, as well as minor, serious and egregious.
· minor infractions, fines commensurate with seriousness of the offence
· 1st serious offence; mandatory prison, with length of sentence commensurate with seriousness of the offence
· 2nd serious offence; mandatory life imprisonment with possibility of parole
· 3rd serious offence; mandatory life imprisonment without possibility of parole
· 1st egregious offence; mandatory life imprisonment in solitary, without possibility of parole
It seems highly likely, that the death penalty would be on the table for egregious offences, but that is beyond any moral authority I could have to actually recommend.
Trials for all serious or egregious offences would require at least two judges, with necessary security clearances and knowledge of AI, and I suggest one male and one female, of which one is Caucasian and one is not. Preferably these trials would be trial by jury.
I also suggest a UN AI Court, to which appeals of convictions from courts within nation states could be heard. Representation on the UN AI Court would include, both for administration as well as judges, rotation inclusive of all nations, for example, with something like the US Senate, where each state regardless of size has equal representation. That arrangement would be fair because it is inclusive of all nations currently disenfranchised from global governing influence.
I suggest special units of police forces at national, state, and municipal levels, of government, as well as a global UN force, all specifically targeting AI law enforcement, as is now common, for example, homicide divisions, corporate crime division, etc.