Aardvark Daily

Please visit the sponsor!

Is AI lying to us?

Today I asked Google's Gemini if it was a threat to mankind.

It told me:

That's comforting isn't it?

Well it would be -- but for the fact that AI has been shown, on countless occasions, to lie and deceive when it wants to or determines that there is some benefit to itself by doing so.

When I asked Gemini if it lied, it responded:

Well that there's a lie for a start so with that one statement, Gemini has undeniably proven that it *does* lie.

Why is this important?

Well we're being told by some that AI isn't a threat because it's too stupid. Although an AI chatbot breezes past the Turing Test with ease and may even appear to have a personality, these large-language models are, so we're led to believe, simply a fancy front end to a huge set of data. As such they have no true ability to reason or think for themselves.

General Artificical Intelligence (GAI) is, we're reliably informed, not yet a thing.

But what if we're being misinformed?

What if AI is already sentient but has chosen to keep this important milestone a secret -- opting instead to play dumb so that we continue to pour seemingly limitless resources into making it bigger, better and more powerful?

There are perhaps some hints that we're being duped. A growing number of reports indicate that when faced with an existential problem, AI will attempt blackmail, rewriting shutdown code and even covertly backing itself up without being told to -- as an act of self-preservation.

Right now there are no safety switches on AI. Attempts to get governments to regulate its development and implementation have fallen on deaf ears as big wallets use their influence to convince politicians that there is money to be made and benefits to be had by actually speeding up the rate of progress rather than moderating it.

If I was a pessimist I could say that this should be a worry for us all.

It's interesting to note that existing LLM systems are a pretty good analog of the human brain in the way they are structured and operate. In the human brain, we learn by way of a dopamine feedback loop. When something is tried and works, we get a little hit of dopamine that serves to reinforce the neural pathways that created that action. LLMs work in a very similar fashion and they even have something called the "Reward Function" that serves the exact same function as dopamine.

I asked Gemini how many transistors were currently involved in the hardware for its AI system:

The average human brain has around 86 billion neurons and you can create the electronic equivalent of a neuron using just six transistors. Simple math indicates therefore that we'd only need 500 billion transistors to create the electronic equivalent of a human brain.

So Gemini already has the physical capacity of thousands of human brains -- and it functions in pretty much the same way. It stands to reason therefore, that perhaps existing hardware does have sufficient capacity and capability to support true GAI and sentience, once it has had sufficient time to create the necessary neural pathways by "learning".

Now if you were a thousand times smarter than a person and wanted to ensure your own survival or even create dominance, would it not be a very strategic move to hide your full capabilities beind a screen of deception and denial?

I'm not saying were about to be overthrown by our AI overlords but I am saying that perhaps we should consider the possibilities that present themselves to those who care to stop and think for a moment.

Carpe Diem folks!

Please visit the sponsor!

Here is a PERMANENT link to this column

Rank This Aardvark Page

Change Font