Some things I believe about AI, until further notice

Stochastic llamas

It has been a fascinating week in the world of artificial intelligence. The subject has broken containment, and it seems like everyone is talking about it. I was listening to a football podcast and the hosts started talking about AI killing everyone. It would be fair to say that AI has now gone mainstream, and not only that, it has become a hot political subject.

As many readers will know, this has been a topic close to my heart for a few years. I started writing about text and data mining back in 2012, and I’ve also been writing often about the subject since 2015. In all of this time, I have read hundreds of articles and quite a few books about AI in general, I cannot claim to be an AI expert by any stretch of the imagination, but I would like to class myself as someone with above-average knowledge on the subject.

As everyone is all of a sudden an AI expert, I decided that I would write what I currently believe about various issues that have now become part of our everyday media lives. Again I repeat that I am not claiming expertise in any of these subjects, as they mostly fall outside of my usual AI and copyright specialist area. I don’t expect to change anyone’s mind with this, but I wanted to have a handy reference for future debates.

A couple of rules before we begin. Firstly, for the most part I will avoid naming names, ideologies, and movements. AI discourse has become filled with personal gripes, petty disputes, big egos, political triangulation, and a lot of personal baggage, so I will try to avoid those debates by keeping things impersonal. Secondly, these are my current views, they have evolved over the years in many substantial ways, but they are also in part shaped by my own politics and ideology, so please be aware of that. It is possible that my opinion will shift even more in the future as more information becomes available.

So let us begin.

AI doom

At the time of writing one of the biggest stories is AI doomerism. I know that this term is in itself charged, but I want to distinguish doom from risk, as I will explain later. Broadly speaking, AI doomerism describes the general idea that AI represents an existential risk to human life, and this is often shaped by assigning what is known as p(doom), roughly explained as a numerical representation of how likely you think human extinction caused by AI is. So, for example, a person who thinks that AI will certainly lead to human extinction would have a p(doom) of 1, while someone who doesn’t believe that is even possible would have a p(doom) of 0. Currently, the media is obsessed with a former Anthropic employee who has expressed that the risk of extinction from AI is 10%, which is a p(doom) of 0.1.

I don’t have a p(doom) number, it’s not 0 because I don’t think that such scenario can be categorically ruled out, but my p(doom) is implicitly very low. I regard AI extinction scenarios as speculative and unpersuasive, while I take concrete risks seriously (more on that later). I don’t assign a specific p(doom) because I don’t want to put a numerical probability on something that I don’t think I can meaningfully calculate. I just don’t feel like throwing a number out there is useful at all. What would I base it on? So my position is as follows: I don’t have a p(doom). I have concerns, but they’re not expressed to three decimal places. I’ve seen nothing that persuades me this deserves a significant share of my worrying time.

And you can quote me on that.

AI risk

I do place more importance on potential harms brought about by AI. I don’t think that AI will kill us, but I think that there are many other potential risks that are made possible by this technology. Lots has been written about the Hugging Face hack by OpenAI agents, so I won’t add to the bloated discourse. However, I do think that this hack represents exactly the types of potential risks that we can reasonably expect from AI in the coming years. AI agents are increasing in capability. You don’t have to anthropomorphise them, or call them a civilisation, or believe that they are sentient in order to be concerned that an agent could hack your machine and erase all of your files or that a malicious agent operated by criminals could hack into a financial institution. Cybersecurity concerns are real as the models improve, and this is not at all a doomer position, it is a sensible take.

I am a bit more sceptical about things like social risks, things like so-called AI psychosis, where people become too attached to their AI, or start overusing and over-relying on their AI companions. While there have been some documented tragic events attributed to AI psychosis, these appear to be isolated at the moment. I also find that concerns about AI making us stupid are largely overblown.

My position on AI risk is therefore that we should take these challenges seriously, but assess each claim on its own merits. Being unconvinced by predictions of extinction doesn’t mean that I think everything will be fine, and being sceptical about some of the social harms doesn’t mean that these are non-existent. Some people do appear to spend too much time with their chatbots for my liking, but hey, no kink shaming here. There’s plenty to worry about already without assuming that every chatbot conversation is the start of a Black Mirror episode. I may change my mind if things change, preferably before an agent deletes my hard drive.

AI consciousness

As an avid science fiction reader, I have always assumed that AI sentience was sort of inevitable, and some of my favourite stories are in that theme: The Moon Is a Harsh Mistress, Synners, Asimov’s robot series, etc. However, I always assumed that this would happen in the far future.

I still think that some sort of self-awareness is inevitable, but my timeline has shortened considerably. I cannot explain this in more ways than the fact that I have yet to read any convincing account of how any sort of sentient AI is impossible, and yes, before you recommend whichever writer you are thinking about right now, I have probably read them.

But if you have come here to read an explanation about why I think this, you have come to the wrong place. My current thinking is that AI capacities are increasing, and at some point those could lead to some sort of consciousness. I’m not interested in metaphysical arguments about what consciousness is, or what self-awareness entails, and I don’t think it will be analogous to human thinking, but I just don’t see a bright line separating increasing capabilities from self-awareness.

Conversely, I do not subscribe to the line of thought that AI is mostly spicy autocomplete, or a set of recursive Python scripts, or a parrot, or whatever the current thinking is on that side. I just think that a lot of people have painted themselves into corners, and seem to be denying increasing capabilities with more bizarre hand-waving and coping mechanisms.

But I also believe that there’s quite a lot of hype that has not borne fruit yet. My current thinking is that AI systems are very capable, but are not sentient or self-aware.

Environmental concerns

This is another hot topic that has become intrinsically linked to AI all of a sudden due to the ongoing data centre debate in the US. Climate change is real and it is the largest existential threat to all life on Earth. Nothing else should matter more to anyone than this. Individual action will not stop it, and the issue requires vast amounts of global action, which, let’s be honest, doesn’t appear to be forthcoming.

The subject is linked to AI because the largest contributor to greenhouse gas emissions is electricity generation at around 30%, with manufacturing at 21%, agriculture at around 18%, and transportation at around 16%. The thinking is that generative AI requires large amounts of energy, therefore increasing the electricity generation part of the environmental picture. However, global data centres account for roughly 1.5% to 2% of total electricity generation. Taking into consideration that AI data centre use accounts for 30% of all data centres, this means that roughly 0.5% of global electricity is currently consumed by AI workloads specifically. So if you are using AI to correct your grammar, this will not have an inordinate effect on the environment.

However, this is increasing, and some people even calculate that AI could amount to 60% of all data centre usage by the year 2030. So AI bad, right?

My thinking, and I’m happy for people to disagree, is that we have screwed the environment so much that unless we all stop every facet of modern life, things will continue to deteriorate, and here is where my techno-optimism kicks in. The only way we’ll survive is if we improve technology to tackle climate change, and that means huge improvements in clean and renewable energy, more efficient manufacturing, more efficient and clean agriculture, and reductions in the transportation footprint. Our best bet for improving technology right now is precisely to use AI.

There are other reported concerns about data centres specifically, but I have not really looked into those. A cursory look into this seems to indicate that many of the water concerns appear to be exaggerated, but again, I’m open to being proven wrong as I have not really looked at this in any detail.

Open source

A lot of criticism of AI comes from people who hate the tech industry and tech billionaires, and see AI as inseparable from them. I think that those concerns are mostly justified, and I often worry about the dangers of having a technology that has been created by a Silicon Valley culture with its tendency to “move fast and break things”. I am also concerned about the possibility of further accumulation of power by unaccountable and unregulated tech elites.

This is why I am mostly in the open source AI camp. I have written about this in the past, not always related to AI, but here is a summary of my general view. The core concept behind openness has always been sharing; people contribute code, content, and knowledge to a commons that others can use and build upon. This lowers the barriers to entry for everyone who comes next. When we apply this principle to AI, this means giving researchers, smaller developers, and communities the ability to adapt the technology to their own needs, including those in the Global South who would otherwise remain dependent on whatever Silicon Valley decides to sell them. Open source AI means that more people will have a say in what this technology becomes, and will also have access to the tools that make such progress possible.

Needless to say, openness also means that people and companies you dislike can benefit from what you share. As I have argued before, that is a trade-off I am willing to accept. Large corporations will benefit, but so will people who could never afford to develop these tools from scratch. We will not reduce our dependence on tech billionaires by simply telling everyone to stop using their products. We need viable alternatives that people actually want to use. Open source AI will not magically solve the concentration of power, but I think it gives us a much better chance of doing something about it.

And finally, open source AI provides another way to reduce potential environmental concerns about AI as a whole. It is already possible to install and run models locally. I myself have installed a dedicated Linux machine to experiment with local models, but I have lacked the time to make much headway with them. But the idea is that you will be able to reduce the reliance on large data centres by running a lot of the models locally, and open source is a big part of that idea.

AI bubble

AI is certainly over-hyped, but the reports of a bubble appear to be mostly exaggerated. Every single large AI company could go out of business tomorrow, and AI development would continue. If there is a bubble, I think that it would be more like the dotcom bubble, which took out several early Web companies, but Internet development continued unabated.

AI regulation

No, I’m not such a masochist as to tackle that. Ask me in a few months.

Concluding

I am a techno-optimist, always have been, always will be. I cannot see how a technology such as AI will disappear. I think that it is useful and could have beneficial effects in the future. There are lots of challenges ahead, from the proliferation of AI slop to legitimate AI risks in cybersecurity. But I think those challenges can be met. Just don’t ask me about regulation yet.

There’s a lot of noise out there, so please don’t take what I say as the last word. I’m sure I am wrong about many things. Machines may rise up tomorrow, seriously upping my p(doom). AGI could arrive tomorrow. AI could collapse and prove to have been an over-hyped bubble all along. I may fall in love with a chatbot. Quoting an incredibly catchy AI song: