I am just finishing up a week-long AI safety course. Today we discussed about possible AI doom situations. Personally, I don’t tend to find these types of “predicting the future” activities particularly useful or interesting (An exaggerated excerpt from a forecasting event I went to for some stupid reason on the night of NeurIPS deadline: come on guys! Let us pretend it’s 2012! Let’s predict the probability that Prime Minister Abe visits the Yasukuni Shrine shrine in 2014! Let’s predict the probability that AI releases a bio weapon killing 10 million people! Let’s predict the probability that you catch your AI girlfriend cheating on you from a CoT monitoring probe in 2038 specifically !!! This is fun. This is fun! Here, have a vegan protein bar.)
I feel very unperturbed by the idea crazies using Claude to create chemical weapons, or misaligned AI robots staging a coup. Call me jaded, or whatever. But I think that there are much easier ways of causing harm then a DIY virus. Like ploughing a car through a busy street. It is funny to think that I hold the optimistic view. I think that most people are decent and we can trust them with powerful chat-bots.1
So the catastrophe that I thought the most realistic (realistic? I feel like this word is loaded...) was gradual disempowerment.
Chat, what is gradual disempowerment?
Gradual disempowerment is an AI-risk scenario where humans do not lose control in one dramatic “AI takeover” moment, but instead become less and less able to shape society as AI systems take over more economic, political, cultural, and institutional functions.
The soft version of this situation is this:
You live in the world 2078. You do not need to work because the corpo-giants only recruit AI agents (they are more efficient! they are quicker! they are cheaper! they do not unionize.) You are dating your AI girlfriend, Elara, who is built on the architecture SuperGrok 12.9 AlphaBetaOmegaXXX. You have not spoken to another person since school, during Human Cultural Studies class. Your food gets delivered to you by drone, but you are rarely hungry as you are on cytosynthepeptide GLP-111.
You get the gist.
I want to persuade you that I do not think this is going to happen. I think that most people will be safe from this possible reality. But not everyone. For some, well, I think that some people will be drawn into this friction-less existence. Where nothing is hard. And where you do not need to deal with people, because having relationships with other people is inherently a difficult thing.
There’s a really popular thought experiment. I’d imagine philosophy-bros would sit in their little organic wine bars discussing this: there is this machine where if you plug in you experience endless pleasure but this makes not being plugged in so boring and mundane that you essentially are never not plugged in. Would you plug in? It’s like that thing from Infinite Jest - a film so interesting and entertaining that people watching it become catatonic and cannot stop watching until they die (I actually have never read this book because apparently it is a male manipulator book and to me that is entirely unappealing). Would you watch this film?
This is barely a thought experiment because no one would say they would plug in. I want to argue, though, that people have plugged in already. The internet (scrolling reels, watching TikToks, watching 7 second ultra porn compilations which I didn’t realise was a thing until recently) is the mild version of DFW’s The Entertainment. You here about those boys who never leave their rooms. Ever! People are already living a simulated life, just some in bigger chunks than others (read the Goon Squad by Daniel Kolitz). I think that AI will make it easier and easier to go down the rabbit hole of endless entertainment, frictionless existence and pleasure.
The internet is dying (see The Internet is Already Over by Sam Kriss). It has become infected by this plague of AI slop, corpo-slop, looksmaxxing etc. People have realised this and are evacuating their digital abodes and coming back to the real world. I am telling you this as a warning - make sure you do not get Left Behind.
However, I also want to give the counter-argument. Right now, as Instagram becomes overrun with AI slop (I do not have a TikTok - I am born in 2002 and am too old) I think that some of us (most of us?) are getting bored.
The internet is dying.
Instagram is not what it used to be. Don’t get me wrong, I still love a few reels here and there. And I still use it a lot to find out what clothes I should buy, what is cool, where I should eat etc. But I am no where near on the app as much as pre-GPT. And neither are many of my friends. I wouldn’t even call this a concious choice - I am not one to use those app-blocking apps or that brick thing where you can turn your phone into, what, a brick? You could maybe blame this on my pre-frontal cortex fully developing, but there is just so much more rubbish on there I do not even think it is fun anymore. I’m uninterested in seeing your lookmaxxing, jawmaxxing, sleepmaxxing, heighmaxxing, autismmaxxing routine. Or AI videos of dogs dancing. It used to be that I can tell AI videos apart really well, but as they get better and better, I am no longer an efficient classifier. But because I sometimes don’t know if something is real, I am just less interested in watching these little videos at all. It takes the fun out of everything
Let’s make a startup for people to go outside and touch grass.
We’ve seen a bunch of these new apps for people to actually go outside and meet people and do things in real life. How did we meet? Oh, you know, naturally. Yeah! Naturally. What do I mean by that? Oh, you know, I was on this app where they schedule meet ups every Thursday where single people who are into techno-rock music do a run club together and if your run times both end in the same digit then you have to go on a date. Yeah! That app! But I think these startups are just the death rattles of the internet - like the metaverse!
The main point - let us evacuate together.
What I want to say is that I think that AI will fracture our society even more. There will be people, probably like me and you, who will be on the internet even less than now. Yes, our food may arrive by drone and hopefully my Claude coding agents will work without me spending a million tokens on Opus 47.69. But, fundamentally, we will be the same. We will have relationships with people. We will have arguments. We will have anxiety and stress. But the ones who stayed on the internet? I don’t know what will happen to them. I don’t know what they will experience. But they have plugged into THE MACHINE. It is difficult to leave.
I want to preface this view by making sure that I am NOT saying that no one should work on AI safety - this is not true at all. I, myself, want to be (am?) working on AI safety and I think this is a really important field.