I remember back before Elon Musk bought Twitter, they had a fact-checking team which would display fact-check messages over tweets. IIRC, this basically completely backfired and people treated the fact-checks as corporate BS. I expect if OP's idea is implemented, we'll get a huge surge of discussion on social media about how this is supposedly "terror marketing", "a bid for regulatory capture", "building hype for their IPO", "distracting from the fact that they're hitting a wall", etc. People tend to have a cynical view of corporations as amoral profit-maximizing agents. As described there's not much in this proposal which refutes the standard terror marketing/regulatory capture/IPO hype narratives. I think people tend to latch on to those narratives due to generic anti-AI sentiment and also fear of mentally accepting the possibility of human extinction. I worry many on this site are in a bubble and don't realize how common these narratives already are. Adopting the narrative gives you an air of cynical superiority; un-critically repeating what BigCorp says makes you look like a rube.

Oh yeah, if you're going to do this anyway I suppose the displayed message should be right-wing-coded? https://x.com/ESYudkowsky/status/2099213169805902203

EDIT: If people at AI companies want my help brainstorming what an actually-effective message would look like, feel free to shoot me a DM

I believe that failing to inform the public is a moral mistake. And if you feel like you cannot talk openly about the risks of what you are doing, it might be time to look into the abyss.

This seems like it excludes situations where you need to talk about EA, trolley problems, killing one person to use their organs to save five, racial disparities in crime, support of Israel, support of any politician whom the media doesn't like, in some places evolution, and in general a whole lot of things. The criterion "would an average person reject this if you talked about it openly" is bad in just about any context.

Also, using X to talk about something that X is not for, on the grounds that it is really important to talk about it, is permanently destroying trust in the usefulness of X, and in many cases destroying the commons.

Btw the Republican Florida AG agrees with you (pg21-22):

Defendants should be enjoined from promoting ChatGPT as safe, reliable, or accurate. All communication platforms for ChatGPT and advertisements for ChatGPT should clearly warn users that ChatGPT is often unreliable or inaccurate,

and Defendants should not affirmatively represent ChatGPT as being safe for use. To the contrary, a company collecting the data users supply to build future models that Defendants themselves say may end humankind must affirmatively warn Floridians of these existential risks. The reason Defendants have not done so is obvious. Far fewer people will use a product if they are told that by doing so, they are increasing the probability of killing themselves and everyone they hold dear. It is not enough to make these statements on X or Substack, which many customers do not see. Defendants must properly warn Floridians each time they log in to ChatGPT or visit its home screen

Bolding mine. Italicization in original.

Imo having to do so due to legal force might actually be one of the best ways for the warnings to get through, since it cuts through some of the natural cynicism ppl might have about a company warning people about their own product.

I think there might be concerns that
(1) with "our company's AI has hacked in X incidents", AI labs are disincentivized to do it due to bad PR
(2) with "our CEO thinks AI has X probability to kill all humans", the median user might think that it is alarmist and hype

I think a potential banner stating "We believe that AI can bring great benefits as well as great dangers, here's why you should be alarmed" (with a link to a personal blog post by Dario or Sam) could work. A banner must reference both harms as well as benefits, and a personal post by the CEO in a neutral ground (for example their personal blog sites) would give an authentic tone and be better than what could be passed off as a company PR statement.

The difference is that Fable 5 is a direct product change that affects the way people use it. The safety stances and changes semantically belong in the blog. But I do like your idea.

I think institutional credibility is so shot that there isn't really any combination of words they could send that would illicit a reaction along the lines of "let's do what these people recommend we do" from a majority of their users.

This doesn't mean it's hopeless if they actually want to do good, it just means that people are more cynical and more game-theoretical nowadays. Costly signals and serious commitments, made plainly enough that they can't be spun or re-interpreted^[1], can probably still get peoples' attention. In the context of safety/interpretability, a very well-conducted contest set up around red-teaming for bounties might help, or a unilateral, expensive commitment to transparency about their training runs that allows states, parties, and national governments to conduct probes of their models on issues that matter to their constituents^[2].

1. ^{^}

   As an example, in the context of AI unemployment, this would look something like providing a direct cash payment to people recently made redundant by a new technology, and it would not look like setting up an NGO that's ostensibly supposed to help them in some way that involves sinecuring the nephews of people connected with its founding.

2. ^{^}

   Something really wide-ranging that can't be interpreted as pushing an agenda. Let every faction of people probe for things that they find objectionable. Let China have a look at why Claude Sonnet thinks Nigerians are four times more important than Chinese. Yes, let the AFD look at why the results for Germans are even starker. This will be politically costly for AI companies, both neoconservatives and many illiberal-leaning lefties will be upset, but it's morally obligatory to be pluralistic in exactly this way if they sincerely believe that they're building God.

I recommend linking to the CAIS 2023 one-sentence statement. It's the single link most effective for convincing "normies"/outsiders that ASI x-risk is serious, in my experience.

Imo, it should replace Claude's greeting on the home page: "Ask me about an AI Pause".

When clicked, a chat with Claude is started with a prompt that's something like "Could you please explain AI Extinction Risk to me in a way that I will understand? It's possible that I don't know a lot right now, so asking a few questions to start would be helpful. If you could link me to sources of further reading, and real AI Safety incidents, that'd be even better. I'd also be interested to learn what reaction politicians have had to this."

An issue with this is that the average person doesn't understand the AI labs' position of "we think this might kill everyone, but we're building it anyway because we think our competitors would be even more likely to kill everyone if they got there first". Given that, saying "we think our product might cause human extinction" probably sounds bizarre or like hype. The linked post could explain this though.