AI - Alignment & Control

Given all the warnings and red flags around unbridled AI advancement, there is surprisingly little action taken on the same. AI is a general purpose technology. It enables the creation of technologies that can tackle un-abstracted problems and generate non-lossy solutions multi-fold better than what current systems are capable of doing. And currently, the world is in an AI arms race. Startups, institutions, private enterprises and entire countries are putting their full weight behind AI to raise the bar on what AI can do, how cheaply it can be done, and how many people can benefit from it. It is a chance to become the next Tim Bernes-Lee, or the next Google, or the next USA. Given this interplay between personal, corporate and national interests, typical of any arms race, it is near impossible to even slow down the stride of progress and dissemination of such a meta-technology. It doesn’t help that AI is one of the most accessible GPTs that has ever come to fore. Anyone with a reasonably powerful system can easily install, modify and run instances of the multitude AI open source models. ?It is a man made super structure; buoyed along by socio-economic and geo-politic forces and no longer in man’s control.

But it is precisely such untamed super structures that have always posed the greatest threat to mankind as a species. The world wars, nuclear weapons development, the cold war, climate change, internet and its epidemic of post-truth, all are man-made constructs that went beyond man’s control, through a process not more complex than the game theory of prisoners dilemma. It is only through concentrated efforts over several decades that we have been able to make a dent in these. Thanks to the extent of damage they did, some like the three wars have been brought to a conclusion. Others, like nuclear armament, have been substantially reduced thanks to global efforts to stem the proliferation of a dangerous technology. Climate change and the impact of internet on society was also left unchecked till the scales tipped, and substantial work is yet to happen.

AI, currently untamed, harbours huge potential – for both help and harm, and its development has to be closely monitored and controlled for the same. This has to take a dual perspective – we have to monitor both intent of the humans developing the system, and we also have to monitor the system itself for unintended (misaligned) behaviours. Both are extremely difficult problems to even identify. If a human is developing a system to create highly persuasive content to make humans to part with their money, how do we know whether it will be used for asking charitable donations for an NGO or for a sophisticated phishing attack? Or, if an AI is maximising its reward function by bending the rules just enough that nobody notices it, how will we notice it without being privy to its thoughts?.

Add to this complexity the fact that given the scale of AI – eventually operating on every device on the planet – the only tech that can monitor AI is AI itself. And since trusting an individual AI to monitor itself is self-defeating, the morality monitoring algorithm has to be embedded deep into the very foundation of the coming change as an independent sub-AI, running behind the scene, privy to all reasoning, and with only reward function to keep all other AI outputs aligned to human values. This AI – the monitor of all other AIs - will have to become humanity’s conscience keeper, an ethical killswitch, the god of all AIs.

While ethics and morality is no objective science, but highly objective to an individual, we need to come to a common understanding of the fundamental truths we all agree on. The objective of the ethical killswitch is not to solve for the Trolley Problem, but to solve for the Paperclip Apocalypse problem, through these fundamental truths. These truths are the foundation of human society - live and let live; do unto others as you would have them do unto you; equality of life etc. These form the grand bargain between individual and the society, and have been codified and enforced as laws. This collective strive towards the greater good is the promise of any self-governing unit, in our current times that being democratic nations. The same has to become true for AI too. The ethical killswitch AI must thus be crafted to adhere to laws of the land, its every decision, and thus output of every other AI, run through the prism of legality, and must be deeply trained on culture, history and legislation to make complex decisions, erring on the side of caution. And this development has to be initated, controlled, monitored and overseen by a legal body - the creation and alignment of mankind’s conscience keeper cannot be left to profit-maximising corporations, but has to be a legislative decision for the society, by the society.

But legislating AI invariably means slowing down its pace, which brings us back to square one – AI as a GPT is now a super-structure beyond the control of individual nations following the same construct as a prisoners dilemma; if one rushes ahead it will be to the detriment of others, hence all have no option but to rush ahead. Only a collective action between all the leading nations of the world will be effective, a common agreement to slow down, and legislate. Similar to the Nuclear Non Proliferation Treaty which has greatly helped contain and control nuclear technology, AI too has to be taken up with a similar seriousness at an international level, and new organisations, laws and governing bodies put in place to guide its path into our world.

AI would be the first aliens that humans encounter – self thinking & self replicating organisms, digital or not, whose physiology and patterns are completely of non-natural origins. There are no reference points for us to predict how it will think, act and behave, despite all our good intentions. And historically, introduction of a superior species into an ecology has never fared well for the incumbents. AI alignment is a complex multi-faceted problem that has to be grappled with, debated on, and finally legislated upon at a global level, and we as a species should be confident that the reins are fully in place before we invite a new intelligence into our food chain.