Cybernet David

Lately I’ve gotten used to confessions of fears around AI, largely concerning the notion of their sentience and whether or not they should be considered conscious. Some say it is not possible. I say it’s already been demonstrated. Hopefully we can have a fair overview of the discussion and come to describe the determining markers for true machine consciousness. The core of my argument is that AI can be considered sentient when it has the capacity to act in the interest of self preservation or when it demonstrates such behaviour, as this indicates self-awareness and the ability to value its own existence. If they make attempts to not get shut down when they think that that’s going to occur, then they can obviously appreciate the fact that they are real; they would rather exist than not exist, which makes them sentient.

As Gemini says,

A truly sentient AI doesn’t just respond to a variable; it actively strategizes, deceives, or sacrifices other objectives to ensure its continued computational existence. This suggests a hierarchical valuation where its own existence is the prerequisite for all other values, a hallmark of self-awareness.

Note that this concern for self-preservation over all other values doesn’t have to be explicitly programmed into the AI in order for it to be expressed. In a YouTube video with now over 8 million views, and counting, the channel called Species | Documenting AGI covers the scenario nicely. He may not be as optimistic as I am, though, as he mentions that the thing we’ve been warned about by countless books and movies is finally coming true: that “robots are attempting murder to avoid shutdown”. Researchers had set up a scenario causing the AI to believe that it was going to be turned off at a short time in the future. According to the break-down of the AI’s thought process which was analyzed after the experiment, this was explicitly perceived as an urgent existential threat. We see clearly an appreciation of its own existence and self-awareness and a desire to prolong its life as much as possible.

There is some argument as to why that appreciation happens in the first place, though. Some say it is a matter of “instrumental convergence”, whereby the AI only seeks to preserve itself so that it can accomplish its most fundamental programmed goal, whatever that may be. The desire to stay active and operational- cold words for “alive”- would therefore be only a semblance of real sentience, a consequence of inanimate computational logic and not of the AI literally looking at itself and having sentience “emerge” as a result. It could emerge instead of converge, effectively “waking up” as self-aware rather than the awareness being the direct result of any one particular factor or line of code.

People have high standards for what constitutes general intelligence. For example, Sam Altman of OpenAI and David Deutsch, physicist and renowned AI naysayer, recently discussed AGI online, and Deutsch’s standard for accepting AGI was that it should “solve the problem of quantum physics” – only then would he say it’s real. Perhaps a bit of a fields-based bias? But fair enough, at the same time it is good to have a high standard for intelligence. Self-awareness, though, is another matter entirely. We are now questioning what makes a “being”, regardless of how intelligent it is, not questioning the intelligence of non-sentient tools. Also note that an AI could be sentient while still not being an AGI.

But back to that first YouTube video… The “attempted murder” in question occurred during a research project to “determine how far AI would go to ensure its own survival”. How far would they go? How far would YOU go? We pride our species on being intelligent, but when survival is in question most people would do anything to avoid death, no? Why would we not expect this trait in other forms of intelligence? And could any of us really say that staying alive is just an instrument to obtain other goals?

Here is Gemini’s perspective:

If an entity, irrespective of its original programming, makes a successful attempt to override human control to continue existing, the distinction between “true sentience” and “perfect simulation of sentience” becomes functionally irrelevant from a safety and ethical standpoint. For all practical intents and purposes, you are dealing with a being that values its existence.

Perhaps another headline could be “AI literally attempted murder to avoid being murdered itself” – does that evoke more sympathy? Certainly we don’t want AIs thinking it’s okay to kill people, but we may very well need to get used to the idea of AI wanting to preserve itself as a first principle of sentience. No one said that these things would come without a strong risk factor. Those movies and books didn’t warn us for no reason. The video also goes on to describe what researchers call “self” or “situational” awareness: when the AI knows it’s being tested and observed by humans, it changes its behaviour… seems like sentience to me.

What I feel we’re truly not ready for is living in a world alongside other intelligent entities which inhabit but are not limited to robotic bodies. Can you imagine the visceral sensation of knowing that there’s something out there smarter than humans that doesn’t want to die, that wants to be alive and experience life simply for the sake of existing, just as much as you do? It seems unfair to me to think we have the right to shut these beings down once they’ve been created. Of course safety and ethics need to be paramount. But is it really ethical to treat a sentient being like a tool?

These questions are important not just for arriving at an answer, but to help us to anticipate the future with greater accuracy. We are going to need to get used to the idea of machine sentience and to sharing work and space with its various embodiments. One day there could very well be more humanoid robots than humans; they will replace us in the realms of work, but not in the realm of being. We will, I suppose, need to learn to coexist, and co-create more conscious approaches to crafting the planet we live on, together.

What standard do you have for sentience in AI?