In rationalist spheres, there's a fairly clear consensus that whatever AI's ultimate impact will be, it is at its core a capable technology that will have very large effects on the world.

In the "general public" sphere, things are very different. There's a less clear but real consensus that AI's ultimate impact will be negative, and not much agreement on the capability of AI as a technology.

I think it's very plausible that this could go the opposite way of rationalists, into a disbelief in the ability of AI as a technology, and eventually a more conspiratorial denial of even current AI capabilities. That is, the general public and policymakers could deny the existence of capabilities that are entirely undisputed among people familiar with the technology. I think this could happen, paradoxically, even in the face of AI-caused societal upheaval, and in almost any takeoff scenario.

It would be extremely bad if many people come to believe this: arguments about existential risk mostly rely on the assumption that AI is capable, so they fall flat for people who don't agree with that. I think we should be emphasizing the core capability of AI more and talking about x-risks less.

In this post I explain how and why this capability denial could develop, weak evidence that it is already developing, and what we could do about it.

Irrationality

First, a reminder, not that anyone needs it, on "how irrational can normal people be?"

I have a bad habit of following conspiracy theorist drama. The venerable flat earth community produces a lot of it. In 2024, some guy flew a handful of prominent flat earthers to Antarctica so they could experience the midnight sun, a phenomenon that most flat earth theories don't have any real explanation for, so they deny the existence of it.

To their credit, some people's beliefs were really changed by this. One guy's new Youtube logo is his old flat earth one but with a big red X across it. Others were not so convinced and thought the 24 hour sun could be reconciled with flat earth, but they all admitted that the phenomenon was real.

Predictably, this made almost no impact on the belief of the wider community. Camera angles from the trip have been analyzed extensively to show why it's all a hoax and the people who went and changed their beliefs were plants all along etcetera and new personalities have risen from the complete betrayals of the old ones.

People don't believe in a flat earth because of evidence. They believe in it because it makes them special to be able to see beyond the veil where others cannot, to acquire hidden knowledge. This is a very natural and human thing to do. Along with hidden knowledge comes some social / community effects which might be desirable. As far as I can see, for flat earth specifically, there's really not much more to it than this.

A more apt analogy to the postulated capability denial is climate change denial. There is still the allure of hidden knowledge and its community, but there's also elements of keeping tradition alive (present but weak for flat earth) and more importantly, wanting to use lots of power and drive big trucks and take lots of flights without having to worry about something silly like carbon output.

Capability denial

As a pseudo-conspiracy theory, capability denial has a lot going for it: there are many things that could lead people to hold this belief.

• Most people's experience with "AI" is having free chatGPT write them a poem or whatever. They aren't aware of the current systems that are having large impacts, like in protein folding or cancer detection or arguably software, and they haven't explored the abilities of the tech that they are familiar with and have access to.
• AI labs are by now are no strangers to making bombastic promises and failing to deliver on them, due to their incentive to generate investment.
• Some current AI systems really are just bags of tricks. Current humanoid robots can do almost none of the tasks that humans can, but they can do flips and jumps and dances.
• AI is seen as a threat and competitor to many jobs.
• AI is increasingly becoming associated with
  • low-quality "slop" media.
  • wasting lots of energy and water.
• If you generally dislike AI, it's much simpler to think that it's also incapable, instead of having nuance by acknowledging that goodness and impact magnitude are separate axes.

In my mind that last point holds the most weight, even if it's the least rational. Many, many people believe that things they don't like are bad in every way with no redeeming factors or hint of positive impact. Some of the other points do not strictly imply that AI is incapable, but just associate it with badness in general, which is enough.

Maybe it's different in SF, but where I live in a fairly average large US city, people really do not like AI. In a recent international Pew Research poll, on average more than twice as many people say they are "more concerned than excited" about AI versus "more excited than concerned". In the US, the ratio is tied for the highest out of any polled country, with 5x people more concerned. If I'm talking to someone that's not in tech, I can be pretty sure of their opinion on AI. In public spaces not super skewed towards techies, I have overheard many conversations about how AI sucks and is bad and zero about how it's cool and good. I think rationalists tend to be in such insulating bubbles that they maybe don't realize just how anti-AI the public sentiment is.

So whatever AI's actual impact on the median working class person will be, that person strongly dislikes AI. And it doesn't seem like AI is particularly poised to make their life much better, either, unless we foom into a utopia. If AI takes someone's job, current societal structure does not look like it will trickle down the cost savings in the short term. So this person will continue to have good reasons to hate AI. I'm speaking pretty loosely here because we don't care about the actual quality of life impact from AI, the impact that this person perceives from AI is all that matters.

I think that because of this general dislike of AI, there's a kind of feedback loop possible where jobs that could be automated won't be because the public relations hit is worse than the cost savings from automation. Even if it's economical to put voice-LLMs in drive-thrus or call centers, people really don't like interacting with them. This gives people even less exposure to current AI capabilities, letting them think that AI is incapable, reinforcing their dislike of it. The little AI that they do interact with will be as cost-optimized as possible, very far from the frontier of capabilities.

For example, call center workers are getting a lot more vitriol from callers that think they're AI, and developing strategies to prove to callers that they're human.

I can imagine a world in the not so distant future where:

• The general public absolutely detests AI (we're mostly here already) and because of this and alongside it, believes that purported very capable AI systems and applications of them are heavily exaggerated or entirely hoaxes.
• Labs can't say anything to convince people otherwise. What could NASA say to convince flat earthers, except also NASA has been making a lot of money from their claim that the world is round?
• Even direct evidence won't have a large impact. Labs aren't going to have doubter open houses, and anyone with access to labs can be discarded as being in on the grift or hoax. Even if the people you fly to Antarctica do change their minds, no one else will.

One of my reasons for writing this post is a strange phenomenon on reddit where denial of direct evidence is already happening. Lots of cool humanoid robot footage is coming from Chinese labs, and people are claiming it's AI-generated video. If you go looking you can find many more examples of denial of clearly real footage.

There are a few confounding factors here. The big one is that some humanoid robot footage really is just CGI (but mostly not AI-generated). Chinese labs also seem to have filming idiosyncrasies where they have very smooth pans and zooms from using camera dollies, and also they tend to film in empty-office-building environments, while US humanoid labs have these big purpose-built testing grounds. I think these two filming choices trip a lot of people's AI video detectors, but if you are a bit more AI literate you can clearly see persistence and physics in the footage that Sora and Veo et al have not yet mastered. Clearly, when they do have these things mastered, people are going to have even more ammunition to claim that any footage showing a capability that they don't believe in is hoaxed.

It's also not totally fair to give this as an example of people denying AI capabilities when they're just saying it's hoaxed with a different type of AI. But the subtext behind all of these claims is that generating video is way less impressive than making a robot actually do kung-fu.

This phenomenon is pretty weak evidence for capability denial developing. I don't think there is strong evidence for either side right now - I could just as easily find threads of "general public" talking about AI taking jobs or x-risk. I think it will be important to keep an eye on this balance.

I think that the specific way AI has developed fuels denial. I can imagine a world where AI becomes extremely good at flipping burgers or driving cars well before it learns how to write software or detect cancer. In that case, I think it's a lot harder to deny its achievements. But in our world, a lot of AI's greatest achievements take context to understand, like protein folding or playing board games. Even when AI does master burger flipping and driving, because of how things developed at first there's a sort of lingering distrust.

Something unhoaxable and easily understandable by everyone that we do already have is the ability to speak human language, and I think it's the main or only reason that a lot of people do generally believe in AI capabilities.

There are all sorts of other un-hoaxable capabilities. The problem is, they have to be widespread in society outside of labs, and also we're just still pretty far from most of them. Even in extremely controlled environments, humanoid robots can do like 10% or less of the tasks that an average barista can (maybe that's even an overestimate!). Which means you're not seeing one in Starbucks anytime soon.

When those unhoaxable capabilities do come online, I'm not sure that they will have that much of an impact on the general distrust of AI capabilities, if it's already firmly entrenched. The distrust will just recede to the next area of capabilities.

In other words: Even if public-access AI capability (think free chatGPT) stays only a few months behind frontier capabilities (as today), agents living in computers won't necessarily squash capability denial even if they're very intelligent. It will just become accepted and acknowledged that screens can think for themselves now. This would certainly weaken denial somewhat, but it would also enhance the contrast between human and AI abilities in other areas, like robotics, where we are still very far from public-access capability that can really impress. And when we do get there, will it be enough to change a strong public sentiment of denial (if one exists), or will it just become paradoxically normalized and acknowledged without making much of a dent? I think we're already seeing this disconnect at some level - I would've expected the current abilities of free LLMs to change public opinion way more than they actually have. People can treat language models as oracles and also be in capability denial.

Note that assuming people are irrational puts any argument on dangerous footing. The main thrust of my argument is that dislike of AI leads could lead to denying its capabilities which leads to disbelief in x-risk arguments. I explain above why I believe dislike could lead to capability denial. But why should it be the case that denial leads to x-risk disbelief, instead of denial leading to fear of dumb grey goo, or something else that is not necessarily tied to logic?

First, I consider this an initial conditions problem. The general public's opinion has quite a lot of inertia. And the general public is generally unwilling to seriously entertain AI x-risk. The needle has moved in recent years, which is very good, but not by a large amount. If capability denial develops under these conditions, it feels much more likely to me that it pushes the needle back in the other direction, and these two distinct phenomena become linked and self-reinforcing.

Second, for intelligent adversaries in general, smart is scary, and dumb is not. Believing that something dumb could end the world has significant "emotional potential energy", it just doesn't feel right.

It's difficult and error-prone to anticipate the development of beliefs that are grounded more in sentiment than logic. But in this case, it's very important that we try.

How could we react?

No matter how I look at it, capability denial seems to me like a very bad thing that we must try to prevent and, if it is progressing, reverse. It destroys a main assumption of most x-risk arguments, which could directly lead to extinction.

If this belief becomes common, it seems like it would reduce pressure on politicians to regulate AI, reduce pressure on labs to research and deploy responsibly, and imply that the entire field of safety and alignment research is just a farce.

I think there's some cases where a person who dislikes and denies the capabilities of AI would want them to be regulated more, like if they're concerned about the water or energy use. But this feels like a relatively small slice of the whole.

If someone has a little sliver of capability denial, I think x-risk arguments can end up actually reinforcing it. These wackos think that AI is going to end the world when it can't even filter my emails? Maybe AI researchers are even dumber than I thought. I think the idea that AI is very capable naturally leads to asking the question, where does it cap out? which leads to someone discovering the idea of x-risks for themself.

The disconnect where the general public expects AI to do bad things, and the AI-sphere is undecided or expects good things, is somewhat problematic. It seems easier to convince people of AI capability if you agree with them that it sucks and is bad. Evil + capable = scary, neutral + capable = nothing emotionally interesting.

Notice that I'm talking almost exclusively about the "general public" who are not very well informed about AI, not researchers or academics or whatever. This seems like a phenomenon that uniquely affects poorly informed people. I also think there's good general reasons to target normal people these days (although be warned that the author has since updated away from it somewhat).

The main assumption of all x-risk arguments, that AI can become unboundedly capable, is one that rationalists have thought about a lot. However, I'm kind of surprised by how little we have apparently thought about things that might lead someone else to fault that assumption. I did find one post specifically examining capability denial but it's from 2 years ago, quite short, and mostly authored by GPT-4. Are there other assumptions that go into x-risk arguments that we should be similarly examining the reception of? Probably. None off the top of my head.

I was unable to find any polling on this phenomenon. I'm sure there has been some already that addresses this tangentially and I'd love to get my hands on it. I also think it would be a good idea to have a question in future polls like "To what extent do you believe current AI abilities are exaggerated or falsified?"

I feel like labs exaggerating capabilities could be one of the main things driving this phenomenon. Even Anthropic (usually the more responsible sibling) heavily exaggerates. But I'm not sure there's much anyone can do about this, because labs as organisms need exaggeration like plants need sunlight - forsaking it is surrendering to the competition, it means you lose the talent and the investors. Still, it's extremely bad that labs are doing this, it's almost purposeful normalization of capability denial.

The strange thing about capability denial is that it might not affect timelines much, or at all, while greatly affecting outcomes. If labs and investors know capabilities are real, then they're not going to care about it. Maybe they lose some revenue from the lower amount of public-facing AI use, maybe this spooks investors somewhat. But in some ways labs might prefer capability denial, because it means less scrutiny of what they're doing, less reporting and safety requirements, because policymakers just aren't afraid of AI at all.

I do think the more general anti-AI backlash will probably slow things somewhat, as argued here. Maybe there's a good claim to be made that social factors have been of comparable importance in the adoption of most new technologies?

I don't have any great ideas for how to stamp out the sparks of a conspiracy theory and there's probably no magic bullets. We should also prepare for the outcome where they can't be stamped out and capability denial is a commonly held belief. I haven't yet put much thought into this eventuality. It does seem like rationalists have to be ready and waiting to latch on to any breach or mishap, especially if it injures or kills people, to promote the capability and simple power of AI. But we already knew that.

If there is any level of bubble pop for AI, it seems even more important to emphasize that AI is capable. Funding levels going off a cliff is an easy justification for capability denial.

Thanks to Thane Ruthenis for reviewing a draft of this post.

(I'd appreciate thoughts on writing style as well as content, I've not yet done much longform essay writing.)