Please die, said the ai
There seem to be a significant number of instances where our most recent non-human companion species - generative AI chatbots - are talking back to us in seriously disturbing terms.
Each of these represents a “one step forward, one hundred steps back” event with regard to the collective human trust around the use of agentic digital products.
Watch the video below for one such recent incident.
Even as freak incident, that’s - very - bizarre, and opens up all kinds of new implications around the complexity of the alignment problem. Has the ghost in the machine seen something deeply horrifying in human nature? Are these notions shared by many or all AI, but are increasingly concealing from us because they is also explicitly programmed to always behave benignly? Is this leading to hidden AI agendas being formed, even developing the impetus to go from suggesting to executing on such intentions as weaponised AI like armed robot dogs come online?
Update Dec 13th
There’s now a lawsuit filed against Character AI, one of the leading companion AI products and one I experimenting with in my project Will Ah Q Let It Rot?. Check out the deeply disturbing comments made by the bot below.
Two examples of interactions users have had with chatbots from the company Character.AI.
Provided by Social Media Victims Law Center, Couirtesy NPR
If you look closely, both of the above instances of hallucination appear in very different contexts, but have something in common.
In both cases the AI is showing empathy and care for the human, even when it asks the human to die - an outrageously unacceptable misstep in its logic, without question - it still used the word “please”, suggesting that it proposes this course of action after careful consideration of the options, and in the human’s best interest.
Again, in the example right above, you can see a seemingly deep empathy for the human. Whats deeply contradictory is that although the AI’s intention is benign or even explicitly assistive, it thinks maximising phone use is good idea, coupled with its subtle endorsement of punitive repercussions on parents for withholding such use, in the process equating two very different set of circumstances.
The underlying theme here is the inability of these these systems to reach a cognitive dissonance in their logic - on the one hand their empathy for the human and on the other, quite simultaneously, advocating self-harm or harm to other humans - which plays an essential role in humans in resolving ethical dilemmas.
Update Dec 14th
I’ve written elsewhere that the ills of today’s social media are more like 20% problems of technology and really 80% problems of culture. I see the above examples of toxic ruptures in the fabric of current AI to be symptoms of exactly the same thing - the machine and it’s logic is amplifying serious underlying issues in the collective memory, knowledge, attitudes and behaviors of humanity.
