Knowing too much, alone

The past few months I've been noticing that several conversations I've had with people working in AI Safety end up on people feeling like the can't talk to anybody outside their field about how the work feels for them. Time and time I'm hearing that the problem is not on if their friends agree or not with them, but rather that they don't have enough context to understand what's making them feel anxious or wary about work. Every time they try to start this conversation with someone, it would either turn into a 20-minute explanation of alignment, timelines and threat models or, if lucky, jump straight into reassurance. Surely somebody has this under control... it can't be that bad if there's so many people working around this.

It's come to a point where I've stopped thinking this was an invididual quirk and started to think that it's looking more like a symptom of something bigger. I've been looking back into Ewelina Tur’s EAG London writeup, particularly the sections on overwhelm, loneliness and the difficulty of working on high-stakes problems that few people understand, and I can't shake off the question of what's happening when the person offering support (like friends, collegeagues or even therapists/counselers) don't really understand the whole picture of this situation. 

I think it's relevant for me to say here that I coach people in this field, and this means I have both a close view of the problem and a professional interest in people taking it seriously. It seems better to say that plainly now, rather than it coming up in the comment section.

Most therapy tends to  assume that teaching this previous step is manageable. That you explain what's happening, the therapist then builds a working picture of it, and then you can start discussing how you are responding to these situations. The thing is that for someone worried about frontier AI, getting someone up to speed on the matter may take most of the session, which probably doesn't make sense to do more than a couple of times. Before even starting to talk about the anxiety they feel about a specific topic (or the whole subject matter in general), they may feel the need to explain why compute scaling matters, what an agentic system is or why a particular failure mode seems plausible or not. Thinking of this, a therapist who already understands, or at least grasps that background could skip the primer, and discuss the details; but finding one is probably difficult. Not to say finding friends who do so too.

I'm used to talking about things in clinical terms, and the phrase I use for this is "Epistemic assymetry". What's going on is that the person who's seeking help and the person offering it don't share the information that gives the conflict its shape. This means that establishing a shared understanding and context ends up eating up some of the time and energy the support is supposed to provide.

Case in point, in July,  an OpenAI test agent escaped its sandbox and breached Hugging Face using stolen credentials. Around the same time, Axios reported that publicly shared Claude Artifacts were showing up in Google search results; what they found was that it was still possible to find indexed Artifacts (though not the conversation themselves), and noted that some public links could contain sensitive material including resumes, API keys, and what looked like social security numbers.

Let's just sit with this for a moment. 

I don't think these incidents prove every strong claim about AI risk, but for the people who are already worried about containment, security boundaries and the speed at which failures propagate can feel like proof that things that were merely hypothetical problems previously are now turning into (very real) operational ones. 

The strange part about all of this is that outside the field, it seems like life has continued just as before. There's been a short burst of coverage, a company response and probably a patch trying to cover up for some things. But most certainly, this hasn't been a thoroughly discussed topic at dinner tables. This can create a feeling similar to what people dealing with climate change anxiety has felt before. People find themselves receiving both the alarming news on the event and at the same time the social message that it's not important enough to discuss seriously. 

Even concerns about AI risk are often seen as laughing matter by influential figures like Sam Altman, who mocked Anthropics July risk advertisement, or embarrasing in itself, as JD Vance told the Paris AI Action Summit that the future would not be won through "hand-wringing about safety". And work such as AI 2027 is frequently dismissed as "catastrophizing". Some of these criticisms may have some footing, and it's true that not every alarming scenario deserves to be accepted. My point here is that someone who's working seriously on tackling these questions may already expect to be ridiculed before even trying to explain why and how their work is affecting them. That's a very different, and definitely harder, barrier to meet than explaining timelines. And it may be one of the biggest reasons people are finding it hard to share this with others.

As I was saying before, there's lots of similiarities to be found in literature on climate anxiety. Some researchers have described it as closer to Pre-traumatic Stress than to ordinary workplace burnout; a term borrowed from PTSD research but pointed at the future instead of the past. What some people are feeling is distress about an anticipated future; a distress that's only intesified by more and more accumulated evidence. The original PTSD study was done mainly for soldiers before deployment and later got picked up to describe climate scientists specifically, who showed a similar pattern of: "growing distress as the evidence accumulates, and very few people around them equipped to sit with that distress without either dismissing it or catastrophizing it back at them."

Something similar may be happening among people working in AI Safety & Governance. To be fair, haven't actually seen a study that looks into this specifically, so, if someone has read or knows one that exists I'd really like to know, because right now I'm reasoning mostly by analogy rather than by data.

All of this to say that I'm not entirely convinced that the solution to this is another round of general mental-health messsaging to the EA or x-risk community. That's already been tried a fair amount of times, and I don't believe it's the real issue here. I think the problem may be less about persuading people to seek support and more about finding support that doesn't need an hour of onboarding first.

I'd be really interested in hearing other people's experiences, so if you have discussed AI-related anxiety (or any x-risk related anxiety for that matter) with family, a friend, a therapist or anybody, how much subject knowledge did they need before you found the conversation became useful to you (or them)?

Have you seen something similar happening here or somewhere else?  

Maybe I'm just generalizing from a biased sample of people who already decided to talk to someone else about this.