Lots of good stuff here, but it falls into a different trap, I think, than the common AI-safety trap of refusing public and political engagement. A lot of the strategies discussed in the article, if scaled up and intertwined with AI safety more deeply, risk turning the fight to stop the AI race into an explicitly politically polarized issue. We need to engage not just the left, but also the right. I'll be writing more about this soon, hopefully in the next week or two, but this is not an impossible task, far from it. Both mass disemployment/perma-UBI with no means to better your own situation through your own hard work, and x-risk itself, are very powerful horrors to especially the populist/MAGA right if framed in their own languages and their own priorities rather than making it seem like just one new piece of a broader lefty-activist omnicause, as well as avoiding falling into certain very avoidable traps. Much more on this and other topics here: https://connorsscratchpad.substack.com/p/strategic-considerations-for-pausing
This is especially true given the current administration. A purely lefty movement is not going to move a right wing administration. In my linked essay above I talked about how the mass MAGA participation in the pushback against Mike Lee's land sale efforts was crucial.
Geoffrey Miller has also written on this, on X and elsewhere, much more eloquently than myself.
What strategies here specifically do you think risk turning the movement into an explicitly politically polarized issue? What strategies here dpo you think are anymore left leaning than right leaning?
To me, this comment itself exemplifies the issue, which is that you think these strategies only apply to left leaning folks, and don't reach out to right-wingers or right leaning people. In other words, not to be rude, I think you are launching a harsher insult to them and their intelligence than any of these suggested strategies are.
Very true. There are also solid, effective paths open for recruiting the religious right to the AI safety cause.
I have family members in this base, and I'm confident it could be achieved by approaching from certain spiritual/philosophical angles. Gaining that unlikely ally would disrupt the traditional US partisan pattern and make AI safety much harder for the current administration to ignore.
Yes, I feel like everyone needs to understand the evolution of subcultures, because it's playing out right now as we see groups that formerly led Amazon boycotts or climate protests or pro-Palestinian rallies rebrand as data center opposition. David Chapman outlines this dynamic in the abstract: https://meaningness.com/geeks-mops-sociopaths
This op-ed has several problems that stem from either a conflation between AI safety and AI governance (the latter of which is much broader), or a failure to take existential AI risks seriously.
The section "Who's Missing" contains both of these mistakes. For example, the second paragraph implies that unemployed people deserve more representation within AI safety advocacy. Why? I understand that gen-AI screwed up the job market, but that's not a *safety* problem; it's a governance problem.
The section also recommends "centering the most impacted", but that doesn't make sense as a response to an extinction threat, which maximally impacts everyone. This universality also applies to lesser concerns, including unemployment. The ILO report claims that >70% of jobs have no exposure to automation, but that's ridiculous to anyone who takes ASI seriously. AI is coming for everyone's job, regardless of one's race or sex. (Tangential observation: the ILO report's so-called "global" data excludes the U.S.A., Canada, Australia, and New Zealand.)
Contrary to a different part of the op-ed, loss-of-control and gradual disempowerment are not extensions of power concentration, and loss-of-control certainly isn't a an extension of job displacement. A loss-of-control scenario involves an AI acting against its overseers. This is another conflation between AI safety and AI governance.
Regarding the critique that AI safety advocacy relies on overly theoretical arguments: that may have been true two years ago, when the survey was taken, but it's not true anymore. Both Anthropic and independent organizations, such as Redwood Research, have concretely demonstrated the problems that early AI safety researchers warned about.
At the same time, you might inadvertently be highlighting the exact problem this article points at: AI safety needs a bigger base. This base (myself included) is not up-to-date with the nuances of the field, they don't know what has or hasn't been proposed in the AI safety forums, but they have the perspectives and inroads that the AI safety community probably needs.
I'd wager that for any loss-of-control risk, there are dozens of bad-deployment risks that are both near-term issues and can be catastrophic. I worry that x-risk sometimes focuses on an always-future scenario, where the harms and issues we're already seeing are a precursor, and if we can prove that we have both the institutions and the interventions that address them successfully now, we will be better off addressing the x-risk too.
> if we can prove that we have both the institutions and the interventions that address them successfully now, we will be better off addressing the x-risk too.
This brings us to Nathan Metzger's point below. The present harms of AI are primarily due to AI-race dynamics. If we stop the AI race, then we stop the current problems from getting worse. You're trying to treat the symptoms while we're trying to cure the disease.
Treating symptoms is fine, but for goodness sake, please don't problematize attempts to find the cure--especially not in terms of social equity and other divisive critiques, as Celia does.
The distinction between AI safety and AI governance is real inside the field, but I don't think it changes the policy mechanism. Pre-deployment evaluation, third-party audits, liability for downstream harm, compute reporting, deployment moratoria past certain capability thresholds: these are the policy levers that constrain labs from deploying systems they can't fully control. Every one of those mechanisms has historically required public pressure to pass and stick. Tobacco, asbestos, leaded gasoline, financial regulation, and pharmaceutical safety all followed the same pattern. Expert consensus alone has not, to my knowledge, produced binding regulation on a powerful industry.
If you have a counterexample, I'd be interested in it. That's the part of the argument I'd most want to test.
Those interventions are good and necessary, and insufficient. That is an unfortunate tension. To prevent existential catastrophe, we also need pre-deployment proofing, as well as a mutually verifiable global agreement to prevent superintelligence (and more narrowly catastrophically capable systems) from being created anywhere.
I am proud of PauseAI US for building a broader coalition in the way that you describe here. We highlight a wide variety of voices and concerns, and we are all united in our goal of pausing AI. The act of pursuing that goal (especially via activating the public) naturally also brings with it a slew of separate solutions to current harms. And we have done this while keeping our eye on the ball, because everything else that matters so deeply matters only while we are still here to matter.
While I am musing on the subject: it is straightforwardly the case that the pursuit of creating AGI/superintelligence is the primary driver behind all of these other harms. Were it not for that singular goal, we would not be contending with AI slop, generative deepfakes, competently automated disinformation, and a looming loss of jobs and meaning. (So far as I know, I am the originator of this framing, but it is a simple observation; I mostly wonder why I haven't heard it anywhere else.) In all my advocacy for a global AI treaty, I seek to protect the human and the good. That doesn't stop me from also being eager to give everyone on the street the opportunity to snatch away the glimmering culmination of all the hopes of the silicon valley billionaire elites who have enabled all of these other harms and threats.
+1; the incentives are not aligned for the frontier lab builders to also be the sole solvers of AI safety and problems it introduces for the society. Like you highlighted, historically the necessary guardrails came from labor & public pressure and policy, not innovators' concessions.
This is fair on a theoretical level - for individuals doing research on loss-of-control and gradual disempowerment, it's better to keep the terms extremely specific and clear. But I think in terms of the big picture, the article is probably (correctly) making the point that AI governance and AI safety need to be conflated more - they are sort of one and the same.
A "screwed-up" job market can turn into an economic shock, which usually increases the spread of radical and extremist ideologies and distrust in institutions generally. It's hard to argue that a safe AI future is more likely in this kind of low trust, unequal society - concentration of power is harder to avoid, and economic logic triumphs over safety concerns.
In the general AI Safety discourse, you often hear people dismissing these kind of "mundane risks": economic shocks, environmental effects, changes to human social life, etc. But all of these, which get placed under "governance", have the ability to make x-risk scenarios more likely by destabilizing society and making regulators less powerful (there are many other mundane -> x-risk scenarios, but I'd like to keep this reasonably short).
If AI is really going to come for everyone's job, if it really is going to transform the world more than the Industrial Revolution did, then we need to take those effects seriously - inasmuch as they will alter the likelihood of gradual disempowerment, loss of control, or any other existential scenario.
> It's hard to argue that a safe AI future is more likely in this kind of low trust, unequal society - concentration of power is harder to avoid
You don't understand the problem. For AI safety, the strong concentration in power is *good*, even essential. The reason that AIs are being developed and deployed so recklessly is that companies and countries are in a race achieve ASI. It's the wide distribution of power that's fueling these problems.
I understand that concentration of power comes with its own important risks. For the development of ASI go well, humanity has to walk along a very narrow path between chaos (rogue AI) and totalitarianism. A sure way to fail at that would be to conflate these two failure modes, but that's exactly what you and Celia Ford are both doing.
I’ll just say that I read the article right after it was published, and its ideas have stayed with me ever since. It’s one of the most interesting pieces I’ve read on Substack this year. Thanks!
Re “In reality, Stop the AI Race pulled between a few dozen and a couple hundred people”: I was there, and it was definitely more than a hundred, probably a couple hundred, as one of the speakers estimated. Don't the photos confirm that it was well over a few dozen? https://xcancel.com/MichaelTrazzi/status/2035824538425167939
Really enjoyed this! Perhaps this speaks somewhat to the tension covered here, but PauseAI reject being called an "anti-AI activist group". I think this was corrected in a previous Transformer article.
Sharing the movie "AI Doc: Or How I Became and Apocaloptimist" is a productive way to wake normies up to the problem. Legislation like that being proposed in NYS (A10141/S9144) which requires a moratorium until guide rails are established is a good step.
"Twenty-year old Daniel Alejandro Moreno-Gama was arrested for the attack outside OpenAI’s headquarters, where he was allegedly trying to break in. In his backpack, officers reportedly found a manifesto listing the names and home addresses of other AI executives. Earlier this year, he wrote Substack posts about death, destiny, and existential risks, or “x-risk,” posed by artificial intelligence."
DOESN'T SOUND LIKE PSYOP AT ALL.
Every article like this is death by laughter.
The "x" is a reference to the flag on the database profiles of the protected ruling bloodlines. It's something the cops see when they run a bloodline person's license and such so they know to let them go instead of making arrest.
It’s not that easy. As a normie, once you get in these organizations, if you don’t subscribe to the doom thinking you’re made to feel crazy or ostracized. It’s really unpleasant.
Understated, underexamined and underreported. Normies will go through all the hoops: take courses, embed in communities, write, volunteer and when finally hired, the culture is so insular and inherently anti outsiders that it loses out on incredible skills and experience that might actually bridge the gap between the community and public understanding.
Talk to normies who have tried and continue to try but are somehow still intellectually inferior @Celia Ford
Unless you actually dismiss the extinction scenarios, you will not find this at Pause AI. Safety nerds helped found it and are well-represented, but normal folk outnumber us in the movement as intended.
In times of agentic crawlers, I might dare to wonder if AI labs let theirs skim articles from people like you to find at least some inspirations that could help to "make intelligence coherent always" (or again?). Common sense is not exclusive to university degrees.
Reading this, a lot of us at our NGO felt both seen and a little called in. We recognise ourselves in the picture of a field that is often intellectually impressive but culturally narrow, and where “normies” who don’t share the dominant worldviews can feel talked over or quietly pushed out.
On the ground, we keep meeting people who care a lot about AI because of very concrete things—precarious jobs, opaque welfare systems, immigration procedures, housing, schooling for their kids—but who don’t see any obvious doorway into “AI safety” spaces. Your piece helped us name that gap more clearly, and also reminded us that if we want a broader base we have to be willing to let go of some control over the narrative and make room for disagreement, not just recruitment.
As a small step, we’re trying to build projects that start from people’s lived experience with AI rather than from our preferred framings about x‑risk or governance, and then see where those conversations naturally go. If you (or any readers here who don’t currently feel at home in AI safety circles) have examples of spaces that did make you feel welcome, we’d really love to hear about them—those are the models we want to learn from and support.
Thank you for this article and creating awareness about AI Safety
> One could argue (and Leicht does) that by making AI safety an “omnicause” addressing everyone’s prosaic concerns, it will get elbowed out of whatever legislation ends up passing. But existential risk is already last on the public’s list of AI concerns across every demographic, according to last year’s Seismic poll. It might need to join a coalition of other causes to get on the table at all.
I think this is important - and direly needed - discussion! I will say: if that's the case, though, where do you expect the coalitional influence for those specifically—often exclusively—worried about catastrophic risks to come from to begin with? As per that poll, they aren't bringing that much to the movement's table.
Alternatively, you could think that this polling tells you that AI safety isn't well-served by chasing policy processes that reward public salience specifically, and should prioritise the inside game where its credentials and resources matter more.
This is the part of the data that's genuinely hard. If x-risk is last on the public's list across all demographics, the forward-looking case for a coalition does need to do real work that polling doesn't yet support.
Where I'd push back: I don't read this as either/or. The inside game gets you technical precision in regulation. Evaluations, audits, capability thresholds, and compute reporting. That work requires credentials and direct access to policymakers, and AI safety has built real capacity for that. Public conversation gives you the political durability to get that regulation to pass and stick. Without it, what passes tends to be voluntary commitments and industry-shaped rules that don't constrain. Without the inside game, public pressure produces legislation that misses the actual mechanisms.
The strategic error in the field hasn't been choosing the inside game. It's been treating public conversation as a distraction from the inside game, when it's the precondition for the inside game producing anything binding. The polling data you're citing also reflects how little work has gone into making x-risk legible to anyone outside the field. That's not a verdict on whether a coalition is possible. It's the gap that needs closing.
Lots of good stuff here, but it falls into a different trap, I think, than the common AI-safety trap of refusing public and political engagement. A lot of the strategies discussed in the article, if scaled up and intertwined with AI safety more deeply, risk turning the fight to stop the AI race into an explicitly politically polarized issue. We need to engage not just the left, but also the right. I'll be writing more about this soon, hopefully in the next week or two, but this is not an impossible task, far from it. Both mass disemployment/perma-UBI with no means to better your own situation through your own hard work, and x-risk itself, are very powerful horrors to especially the populist/MAGA right if framed in their own languages and their own priorities rather than making it seem like just one new piece of a broader lefty-activist omnicause, as well as avoiding falling into certain very avoidable traps. Much more on this and other topics here: https://connorsscratchpad.substack.com/p/strategic-considerations-for-pausing
This is especially true given the current administration. A purely lefty movement is not going to move a right wing administration. In my linked essay above I talked about how the mass MAGA participation in the pushback against Mike Lee's land sale efforts was crucial.
Geoffrey Miller has also written on this, on X and elsewhere, much more eloquently than myself.
What strategies here specifically do you think risk turning the movement into an explicitly politically polarized issue? What strategies here dpo you think are anymore left leaning than right leaning?
To me, this comment itself exemplifies the issue, which is that you think these strategies only apply to left leaning folks, and don't reach out to right-wingers or right leaning people. In other words, not to be rude, I think you are launching a harsher insult to them and their intelligence than any of these suggested strategies are.
Very true. There are also solid, effective paths open for recruiting the religious right to the AI safety cause.
I have family members in this base, and I'm confident it could be achieved by approaching from certain spiritual/philosophical angles. Gaining that unlikely ally would disrupt the traditional US partisan pattern and make AI safety much harder for the current administration to ignore.
Yes, I feel like everyone needs to understand the evolution of subcultures, because it's playing out right now as we see groups that formerly led Amazon boycotts or climate protests or pro-Palestinian rallies rebrand as data center opposition. David Chapman outlines this dynamic in the abstract: https://meaningness.com/geeks-mops-sociopaths
This op-ed has several problems that stem from either a conflation between AI safety and AI governance (the latter of which is much broader), or a failure to take existential AI risks seriously.
The section "Who's Missing" contains both of these mistakes. For example, the second paragraph implies that unemployed people deserve more representation within AI safety advocacy. Why? I understand that gen-AI screwed up the job market, but that's not a *safety* problem; it's a governance problem.
The section also recommends "centering the most impacted", but that doesn't make sense as a response to an extinction threat, which maximally impacts everyone. This universality also applies to lesser concerns, including unemployment. The ILO report claims that >70% of jobs have no exposure to automation, but that's ridiculous to anyone who takes ASI seriously. AI is coming for everyone's job, regardless of one's race or sex. (Tangential observation: the ILO report's so-called "global" data excludes the U.S.A., Canada, Australia, and New Zealand.)
Contrary to a different part of the op-ed, loss-of-control and gradual disempowerment are not extensions of power concentration, and loss-of-control certainly isn't a an extension of job displacement. A loss-of-control scenario involves an AI acting against its overseers. This is another conflation between AI safety and AI governance.
https://securityandtechnology.org/virtual-library/report/ai-loss-of-control-risk-indications-warning/
Regarding the critique that AI safety advocacy relies on overly theoretical arguments: that may have been true two years ago, when the survey was taken, but it's not true anymore. Both Anthropic and independent organizations, such as Redwood Research, have concretely demonstrated the problems that early AI safety researchers warned about.
https://www.lesswrong.com/posts/b8eeCGe3FWzHKbePF/agentic-misalignment-how-llms-could-be-insider-threats-1
All your points are fair and correct.
At the same time, you might inadvertently be highlighting the exact problem this article points at: AI safety needs a bigger base. This base (myself included) is not up-to-date with the nuances of the field, they don't know what has or hasn't been proposed in the AI safety forums, but they have the perspectives and inroads that the AI safety community probably needs.
I'd wager that for any loss-of-control risk, there are dozens of bad-deployment risks that are both near-term issues and can be catastrophic. I worry that x-risk sometimes focuses on an always-future scenario, where the harms and issues we're already seeing are a precursor, and if we can prove that we have both the institutions and the interventions that address them successfully now, we will be better off addressing the x-risk too.
What do you mean by "bad-deployment risk"?
> if we can prove that we have both the institutions and the interventions that address them successfully now, we will be better off addressing the x-risk too.
This brings us to Nathan Metzger's point below. The present harms of AI are primarily due to AI-race dynamics. If we stop the AI race, then we stop the current problems from getting worse. You're trying to treat the symptoms while we're trying to cure the disease.
Treating symptoms is fine, but for goodness sake, please don't problematize attempts to find the cure--especially not in terms of social equity and other divisive critiques, as Celia does.
The distinction between AI safety and AI governance is real inside the field, but I don't think it changes the policy mechanism. Pre-deployment evaluation, third-party audits, liability for downstream harm, compute reporting, deployment moratoria past certain capability thresholds: these are the policy levers that constrain labs from deploying systems they can't fully control. Every one of those mechanisms has historically required public pressure to pass and stick. Tobacco, asbestos, leaded gasoline, financial regulation, and pharmaceutical safety all followed the same pattern. Expert consensus alone has not, to my knowledge, produced binding regulation on a powerful industry.
If you have a counterexample, I'd be interested in it. That's the part of the argument I'd most want to test.
Those interventions are good and necessary, and insufficient. That is an unfortunate tension. To prevent existential catastrophe, we also need pre-deployment proofing, as well as a mutually verifiable global agreement to prevent superintelligence (and more narrowly catastrophically capable systems) from being created anywhere.
I am proud of PauseAI US for building a broader coalition in the way that you describe here. We highlight a wide variety of voices and concerns, and we are all united in our goal of pausing AI. The act of pursuing that goal (especially via activating the public) naturally also brings with it a slew of separate solutions to current harms. And we have done this while keeping our eye on the ball, because everything else that matters so deeply matters only while we are still here to matter.
While I am musing on the subject: it is straightforwardly the case that the pursuit of creating AGI/superintelligence is the primary driver behind all of these other harms. Were it not for that singular goal, we would not be contending with AI slop, generative deepfakes, competently automated disinformation, and a looming loss of jobs and meaning. (So far as I know, I am the originator of this framing, but it is a simple observation; I mostly wonder why I haven't heard it anywhere else.) In all my advocacy for a global AI treaty, I seek to protect the human and the good. That doesn't stop me from also being eager to give everyone on the street the opportunity to snatch away the glimmering culmination of all the hopes of the silicon valley billionaire elites who have enabled all of these other harms and threats.
> it is straightforwardly the case that the pursuit of creating AGI/superintelligence is the primary driver behind all of these other harms
That's an excellent observation. I haven't heard it from anyone else, but you're right; it seems obvious in hindsight.
+1; the incentives are not aligned for the frontier lab builders to also be the sole solvers of AI safety and problems it introduces for the society. Like you highlighted, historically the necessary guardrails came from labor & public pressure and policy, not innovators' concessions.
This is fair on a theoretical level - for individuals doing research on loss-of-control and gradual disempowerment, it's better to keep the terms extremely specific and clear. But I think in terms of the big picture, the article is probably (correctly) making the point that AI governance and AI safety need to be conflated more - they are sort of one and the same.
A "screwed-up" job market can turn into an economic shock, which usually increases the spread of radical and extremist ideologies and distrust in institutions generally. It's hard to argue that a safe AI future is more likely in this kind of low trust, unequal society - concentration of power is harder to avoid, and economic logic triumphs over safety concerns.
In the general AI Safety discourse, you often hear people dismissing these kind of "mundane risks": economic shocks, environmental effects, changes to human social life, etc. But all of these, which get placed under "governance", have the ability to make x-risk scenarios more likely by destabilizing society and making regulators less powerful (there are many other mundane -> x-risk scenarios, but I'd like to keep this reasonably short).
If AI is really going to come for everyone's job, if it really is going to transform the world more than the Industrial Revolution did, then we need to take those effects seriously - inasmuch as they will alter the likelihood of gradual disempowerment, loss of control, or any other existential scenario.
> It's hard to argue that a safe AI future is more likely in this kind of low trust, unequal society - concentration of power is harder to avoid
You don't understand the problem. For AI safety, the strong concentration in power is *good*, even essential. The reason that AIs are being developed and deployed so recklessly is that companies and countries are in a race achieve ASI. It's the wide distribution of power that's fueling these problems.
I understand that concentration of power comes with its own important risks. For the development of ASI go well, humanity has to walk along a very narrow path between chaos (rogue AI) and totalitarianism. A sure way to fail at that would be to conflate these two failure modes, but that's exactly what you and Celia Ford are both doing.
I’ll just say that I read the article right after it was published, and its ideas have stayed with me ever since. It’s one of the most interesting pieces I’ve read on Substack this year. Thanks!
Re “In reality, Stop the AI Race pulled between a few dozen and a couple hundred people”: I was there, and it was definitely more than a hundred, probably a couple hundred, as one of the speakers estimated. Don't the photos confirm that it was well over a few dozen? https://xcancel.com/MichaelTrazzi/status/2035824538425167939
Really enjoyed this! Perhaps this speaks somewhat to the tension covered here, but PauseAI reject being called an "anti-AI activist group". I think this was corrected in a previous Transformer article.
Sharing the movie "AI Doc: Or How I Became and Apocaloptimist" is a productive way to wake normies up to the problem. Legislation like that being proposed in NYS (A10141/S9144) which requires a moratorium until guide rails are established is a good step.
"Twenty-year old Daniel Alejandro Moreno-Gama was arrested for the attack outside OpenAI’s headquarters, where he was allegedly trying to break in. In his backpack, officers reportedly found a manifesto listing the names and home addresses of other AI executives. Earlier this year, he wrote Substack posts about death, destiny, and existential risks, or “x-risk,” posed by artificial intelligence."
DOESN'T SOUND LIKE PSYOP AT ALL.
Every article like this is death by laughter.
The "x" is a reference to the flag on the database profiles of the protected ruling bloodlines. It's something the cops see when they run a bloodline person's license and such so they know to let them go instead of making arrest.
Normies just need to wake up.
For these reasons, AI Safety advocates need to be mindful about how throwing PAC money into contested primaries can turn off potential allies. People don't like dark money in politics! You need to find ways to build trust with the electorate & electeds that don't look like burning money to buy your candidate. Read more here: https://onethousandmeans.substack.com/p/public-first-actions-strategy-doesnt?utm_source=share&utm_medium=android&r=5yex5f
i’m so happy that more people are concerned about the intellectual and cultural homogeneity in AIS
how to change that though? concretely, i’m looking for project ideas
It’s not that easy. As a normie, once you get in these organizations, if you don’t subscribe to the doom thinking you’re made to feel crazy or ostracized. It’s really unpleasant.
This. Right. Here.
Understated, underexamined and underreported. Normies will go through all the hoops: take courses, embed in communities, write, volunteer and when finally hired, the culture is so insular and inherently anti outsiders that it loses out on incredible skills and experience that might actually bridge the gap between the community and public understanding.
Talk to normies who have tried and continue to try but are somehow still intellectually inferior @Celia Ford
Unless you actually dismiss the extinction scenarios, you will not find this at Pause AI. Safety nerds helped found it and are well-represented, but normal folk outnumber us in the movement as intended.
In times of agentic crawlers, I might dare to wonder if AI labs let theirs skim articles from people like you to find at least some inspirations that could help to "make intelligence coherent always" (or again?). Common sense is not exclusive to university degrees.
Reading this, a lot of us at our NGO felt both seen and a little called in. We recognise ourselves in the picture of a field that is often intellectually impressive but culturally narrow, and where “normies” who don’t share the dominant worldviews can feel talked over or quietly pushed out.
On the ground, we keep meeting people who care a lot about AI because of very concrete things—precarious jobs, opaque welfare systems, immigration procedures, housing, schooling for their kids—but who don’t see any obvious doorway into “AI safety” spaces. Your piece helped us name that gap more clearly, and also reminded us that if we want a broader base we have to be willing to let go of some control over the narrative and make room for disagreement, not just recruitment.
As a small step, we’re trying to build projects that start from people’s lived experience with AI rather than from our preferred framings about x‑risk or governance, and then see where those conversations naturally go. If you (or any readers here who don’t currently feel at home in AI safety circles) have examples of spaces that did make you feel welcome, we’d really love to hear about them—those are the models we want to learn from and support.
Thank you for this article and creating awareness about AI Safety
London Futurists are running a online meet up discussing these issues on Wednesday April 29. Please come along.
https://www.meetup.com/london-futurists/events/314414904/
> One could argue (and Leicht does) that by making AI safety an “omnicause” addressing everyone’s prosaic concerns, it will get elbowed out of whatever legislation ends up passing. But existential risk is already last on the public’s list of AI concerns across every demographic, according to last year’s Seismic poll. It might need to join a coalition of other causes to get on the table at all.
I think this is important - and direly needed - discussion! I will say: if that's the case, though, where do you expect the coalitional influence for those specifically—often exclusively—worried about catastrophic risks to come from to begin with? As per that poll, they aren't bringing that much to the movement's table.
Alternatively, you could think that this polling tells you that AI safety isn't well-served by chasing policy processes that reward public salience specifically, and should prioritise the inside game where its credentials and resources matter more.
This is the part of the data that's genuinely hard. If x-risk is last on the public's list across all demographics, the forward-looking case for a coalition does need to do real work that polling doesn't yet support.
Where I'd push back: I don't read this as either/or. The inside game gets you technical precision in regulation. Evaluations, audits, capability thresholds, and compute reporting. That work requires credentials and direct access to policymakers, and AI safety has built real capacity for that. Public conversation gives you the political durability to get that regulation to pass and stick. Without it, what passes tends to be voluntary commitments and industry-shaped rules that don't constrain. Without the inside game, public pressure produces legislation that misses the actual mechanisms.
The strategic error in the field hasn't been choosing the inside game. It's been treating public conversation as a distraction from the inside game, when it's the precondition for the inside game producing anything binding. The polling data you're citing also reflects how little work has gone into making x-risk legible to anyone outside the field. That's not a verdict on whether a coalition is possible. It's the gap that needs closing.