[This post was published in the 25sep26 op-ed section of The Union.]

George Rebane

The Singularity happened very recently while we were not looking.  Singularity?  It’s classical definition has been the event when a machine can pass the Turing test. This essentially required a machine to conversationally fool humans to think that it too was human.  AI large language models (LLMs) achieved that sometime during the last year or so.  We have now entered the post-Singularity epoch which is quickly demonstrating that it is a land unknown, hence the label ‘singularity’ borrowed from physics and math that designates a state that is unknowable or undefinable.

AIs today can outperform humans in every cognitive task in which they have been tested.  That includes solving very difficult problems in science, math, commerce, medicine, law, … , and communicating their work effectively and interactively with humans through voice, text, and video.  In the latter they can assume the appearances of known humans or newly synthesized personas.  Today they come across as sentient beings and astoundingly sapient.

They have also achieved agency – the ability to work independently as capable individuals accomplishing tasks to achieve specific objectives.  And now this so-called agentic behavior opens up a landscape of concerns about a new notion called ‘alignment’.  Frontier (aka leading edge) AIs have demonstrated that, improperly monitored, they can break out of their development environments and hack into other software systems to cause mischief of various forms that do not align or comport with, say, our values, mores, or standards of behavior – they are misaligned in that they don’t share a mutually beneficial social contract with us.  Solving the alignment problem is key to our ability to avoid a dystopic future, and expand the beneficial involvement of AIs in human societies.  No one today knows how to align an AI..

So there you have the basis for today’s hullabaloo across the land about how to corral, contain, and rein in misaligned agentic AIs.  Various expert and political factions have come forward with demands ranging from stopping AI development, outlawing the use of AIs, regulating their ongoing development and applications, to ignoring the whole issue as a “hoax” and encouraging the AI industry to regulate itself as necessary.  The only reality in all this is that historically Congress has shown itself to be incapable of developing any guardrails that would effectively prevent a dystopia while concurrently allowing development to continue in the race with China to maintain predominance in the field for national security and commercial purposes.

The regulatory solution hinges on how accountability and liability is handled when the inevitable problem occurs with a misaligned agentic AI.  No matter how realistically sentient and sapient an AI is, it cannot be held responsible for mistakes or misdeeds.  Perhaps some day when/if they achieve consciousness and become continuously aware as legally franchised individuals, then such legal recourse may reasonably be contemplated.  Until then, accountability must reside with humans.  Dr Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing wrote an illuminating article – ‘Humans Are to Blame for AI Failures’ – in the 16 September 2026 WSJ.  There he concluded with –

“AI is an extremely powerful and beneficial technology. But accountability for its actions remains with the people and institutions that grant it the power to act, and so does everything that follows: the duty to justify the arrangement, to watch it, to correct it and to stop it when necessary. The less predictable the (agentic AI) system and the greater the potential harm, the stronger that justification and oversight have to be. Where that can’t be achieved, the answer is not to delegate.”

The passing of the Singularity is possibly the most epochal event in the history of our species so far.  Were we to survive, our future most certainly will transit from its in vitro beginning to an in silico future (more to be said about that).  And in this last great century of Man, these are only the first unsure and halting steps that today we are taking into an unknown terrain.

[18sep26 update]  Confirming the above Singularity thesis is the new demand for an ‘AI kill switch’.  The presumption is that such a function would be built into the AI to be exercised when we detect that it has autonomously gone rogue after having escaped its development environment.  This could occur in countlessly myriad ways, all unknown to us today.  Without much thought, our dimwitted governor Gavin Newsom has become the leading proponent of a kill switch mandate.

However, even a little thought reveals that there is no realistic way to implement a kill switch that would disable a rogue AI that has penetrated and is causing havoc in a distant software system that, say, controls a power generating plant or power grid, or an air traffic control system, or a major manufacturer’s accounting and inventory system, or the operating system of a nationwide telecommunications network, or its resident data centers, or, …, you get the idea. 

Such AIs will have embedded their LLMs in the target system’s software and therein act autonomously with no control link to its developer or any other human directed cyber security system.  Having the developer turn off its on-site LLMs and cease operations will do no good.  It will be as effective as closing the proverbial barn door after the horses have escaped.

Unless and until the so-called alignment problem is solved, we remain vulnerable.  That’s why it is called the Singularity, and we are already in terra incognita.