Image credit: Robert Clark

I didn’t want to spend 200 hours deciding where to donate like I did last year. Therefore, there are some promising orgs that I did not take the time to learn more about, and this writeup doesn’t provide as much detail as 2024 or 2025. But I’ll try to cover the big stuff.

Cross-posted to the EA Forum.

Contents

Cause prioritization

I explained my thinking in detail in 2024, and some changes to my beliefs in 2025. My high-level prioritization has not changed in 2026:

  • The top priority is to prevent misaligned AI from killing everyone.
  • The best way to do that is a global halt on frontier AI development.
  • At the margin, the best things to do are public and political advocacy to push for a global halt.

Since 2025, I’ve written some posts that add color to my views, including:

(There are more relevant posts in the archive, especially under the “Existential Risk” and “Far Future” headings.)

A lot happened since last year, but not much that changed my priorities

This year:

  • Frontier AI has made huge improvements on certain cognitive tasks like math, even solving a Millennium Prize problem.
  • Anthropic’s AI systems may be significantly accelerating its AI R&D, which suggests we may be on the brink of recursive self-improvement. We may be very close to the point of no return.
  • The Hugging Face incident was the first (and possibly last) “warning shot”, where an AI system went rogue and caused real-world damage. The incident woke a lot of people up to the dangers of advanced AI.
  • Jacob Coxon resigned from Anthropic, and his resignation went unexpectedly viral and received a lot of media attention.

This all matters a lot, but it doesn’t really change my prioritization. Here’s how it did change my beliefs:

  • I have higher credence that pausing AI is urgent.
  • I’m more optimistic about our chances of achieving a global moratorium on AI development—the issue is both more popular and more salient.1
  • I’m less optimistic that things will miraculously work out somehow (e.g., it turns out alignment is easy, all you have to do is pick the right model persona or whatever). The window for a miracle is rapidly closing, and we’re getting increasing evidence that current alignment techniques are not going to work.

Basically, 2026 has made me more confident in my preferred strategy of “advocate for a global ban on ASI”.

There are more orgs working toward that goal (and doing advocacy toward that goal) than there were a year ago, but still not many orgs, so the cause remains neglected.

Post-alignment problems remain under-rated, but I don’t see standout funding opportunities

By “post-alignment problems”, I refer to the existential-scale problems that AI could bring even if we solve the alignment problem. Achieving an ideal future requires solving some critical problems, and it’s not at all obvious that these problems can be deferred to AI.

Pausing AI is the best answer to post-alignment problems, but not necessarily the most cost-effective thing to work on.

  • I’m pessimistic about most relevant research agendas—I explained why in We won’t solve post-alignment problems by doing research.
  • I’d like to see more work on post-ASI non-human welfare; I wrote about this here. There are a few orgs working on this and they get very little funding right now. However, I’m concerned that none of them take ASI sufficiently seriously, and their strategies will not scale to ASI (see The Future Will Be Weirder Than That).
  • Orgs like Compassion Aligned Machine Learning (CaML) work to improve AI’s concern for sentient welfare. Will that work scale to superintelligence? I’m pretty sure it won’t, and even if it does, it’s difficult to imagine a world where this sort of work would make a difference (which is the point I made in An ASI that’s bad for animal welfare is probably bad for humans, too). However, my reasoning could be wrong. I still think CaML-style work is a good expected-value bet because it’s so underfunded—but maybe not as good a bet as AI pause advocacy.
  • I don’t know what kind of work will improve sentient welfare post-ASI, given that we have no clue what an aligned ASI would look like. This is something I’ve been thinking about a lot and want to write more about soon. As for how this affects my donations, I don’t see any orgs in the space who have good ideas about this, or who appear on track to come up with good ideas, so there’s no one I clearly want to fund to work on the problem. For my preliminary thoughts on the problem, see:

    (I realize these titles sound contradictory; that’s because the subject is confusing and I’m not sure what I think. But the two posts’ content are basically consistent with each other.)

  • Perhaps the goal of “get ASI to care about non-human welfare” is insufficiently general, and really the goal should be “solve ethics”, or perhaps something like “achieve perfect strategic competence, and then follow the optimal strategy”2 (see writings by Wei Dai: 1, 2, 3). But what org do I donate to in that case? Wei Dai is way ahead of the curve on thinking about philosophical competence, and more people should follow in his footsteps. There are a lot of people thinking about moral philosophy and the problem of solving ethics, but it’s not clear what kinds of thinking are useful.

The #1 item on my research agenda is to reduce my confusion about post-alignment problems, especially non-human welfare. And it’s plausible that some work in this space is more important at the margin than AI pause advocacy. But I don’t currently see any funding opportunities that stand out as a clear priority in the way that AI pause advocacy stands out.3

Organizations (tax-deductible)

There are a few other protest orgs or public advocacy orgs that I didn’t list, but not because I dislike anything about them; it’s just that I don’t know anything about them. Some possible contenders that I know very little about include People for a Pause, Humans in Control, and Irreplaceable. I believe the last two of those are less focused on extinction risk and more about the negative economic impacts of AI, but I’m not sure.

Evitable

Evitable is a new org that does public advocacy for a global moratorium on ASI. They do grassroots organizing—they’ve been involved in protests and done some preliminary movement-building, and they plan to do more kinds of mobilization in the future. They also make public content on AI risk, including appearing in various national news outlets.

A big component of my decision process is: find writings (or interviews, etc.) by the leader of an org and see how reasonable they sound. Do they understand the AI risk situation? Do they understand why alignment is hard? Are they focusing on the right things? Evitable gets big points because their founder David Krueger scores exceptionally well on the reasonableness test. I read some of his news articles and LessWrong posts and found myself cheering along with everything he said.

Evitable has more expertise on AI and AI alignment than the other grassroots advocacy orgs—David Krueger is an Assistant Professor in Machine Learning and has co-authored numerous AI safety papers; he and Gretchen Krueger (also at Evitable) are co-authors on the recent paper The Feasibility of a Hardwired Pause of Frontier AI Training. This sort of expertise isn’t a requirement for running an advocacy org, but it’s a nice bonus.

Evitable hasn’t done much yet. I talked to them about their plans, and I’m reasonably confident that they will make good tradeoffs about what directions to pursue. A downside of their approach (focusing on capacity-building first) is that AI timelines may be very short, and their efforts might not pay off in time.

Lightcone Infrastructure

Lightcone Infrastructure takes AI risk seriously, and they do various projects that they expect to improve the situation. They mainly build community support and infrastructure, including running LessWrong, the Lighthaven event venue, and (more recently) creating the philanthropy platform Lightcone Commons.

Lightcone leadership is strategically and organizationally competent. The org is good enough to make my short list, but it’s difficult to weigh the marginal value of community infrastructure against more direct public advocacy. My vibe sense is I prefer advocacy.

There’s a conflict of interest in that I was paid by Lightcone Infrastructure to do some work this year, but at that point I was already pretty sure I wasn’t going to donate to them.

Machine Intelligence Research Institute (MIRI)

If MIRI had no money, it would be a strong contender for the best place to donate. However, the other orgs on my list are funding-constrained to a degree that MIRI isn’t. I really, really want Nate Soares and Eliezer Yudkowsky to keep raising the alarm about AI risk, but there is no danger that MIRI runs out of money and they stop being able to do that. The marginal things MIRI can do with funding are still good, but probably not as good as some of the other orgs on my short list.

Palisade Research

Palisade isn’t my 2nd favorite org on this list,4 but it’s the one where I was 2nd most likely to donate—I liked Evitable more than the other grassroots advocacy orgs, Palisade is doing something different (see my 2024 summary), and I was uncertain about which approach was better. I decided I still prefer to donate to grassroots advocacy, the same as in 2024 and 2025.

The main update on Palisade is that in 2024, they had mostly done technical work and were just starting to scale up their political advocacy. Now, in 2026, they have been actively doing political advocacy, and I like what I’ve seen from them (although I don’t know a lot of details).

PauseAI Global

PauseAI Global organizes protests outside of the US. As I wrote in 2024, I support their work, but I prefer to donate to US-based orgs because I believe US AI policy matters much more than policy in any other country.

As of November 2026, their website directs US readers to People for a Pause. This suggests that, if I like PauseAI Global, I should donate to People for a Pause. I decided not to donate to People for a Pause because I know more about Evitable and Stop the AI Race and I trust the latter orgs more.

I have two concerns about PauseAI Global based on private information. The first is a minor concern that I know to be true; the second is a medium-to-large concern, but I don’t know whether it’s accurate. I would prefer not to make vague statements like this, but unfortunately that is the situation. I would probably still donate to PauseAI Global if they were the only org doing protests.

PauseAI US

I talked about why I liked PauseAI US in 2024 and 2025. One positive update is that PauseAI US now spends more time coordinating volunteers’ meetings with policy-makers across the country, which is a great thing to do.

A lot of people in AI safety take issue with the fact that PauseAI US’s director, Holly Elmore, behaves antagonistically toward many people on Twitter, especially AI company employees. I don’t spiritually object to this. Most safety work at AI companies is safetywashing, papering over misalignment, or contributing more to capabilities than to safety; the people doing those things should stop. That doesn’t necessarily mean you should call them out personally in public. I’m not overly concerned that Holly is violating some important norm; if anyone’s violating norms, it’s the people working at AI companies. I’ve heard some accuse Holly of dishonesty, but the statements being pointed to as dishonest strike me as more like biased or motivated.5 (Here is my Twitter reply to one specific claim.)

Sometimes Holly criticizes AI safety people who (in my view) are doing good work, but have associations with AI companies. I think that’s going too far, although I see where she’s coming from.6

I am concerned that many people dislike Holly’s Twitter presence. It creates two problems:

  1. It gives many people a bad impression of the AI pause movement.
  2. It makes some other orgs unwilling to work with PauseAI US, which makes cooperation harder—this is important for organizing protests, etc.

Is it strategically useful to be antagonistic on Twitter? Holly says yes; I lean toward no, although I don’t have a strong view.

Last year, I was concerned about the downside risk, but I believed it was outweighed by PauseAI US’s activities on things I liked: protests, messaging campaigns, organizing meetings between policy-makers and concerned citizens. And there wasn’t anyone else doing all those things. I have become more concerned about the downside risk since then, and there are more people organizing protests and grassroots support than there were a year ago. Therefore, this year I’d rather donate to Evitable or Stop the AI Race, although I still believe the more extreme criticisms of PauseAI US are overstated.

Stop the AI Race

Stop the AI Race It co-organized a large SF protest earlier this year. Before that, the org’s founder, Michaël Trazzi, did a hunger strike against DeepMind. I’m impressed by his agency and dedication, and I like what I’ve seen so far.

There isn’t much info available about Stop the AI Race. Even though the org is a strong candidate, I decided not to investigate further because I wasn’t sure what information would convince me to donate. It comes down to who I trust more to execute well; Stop the AI Race looks promising on that front, but right now I trust Evitable more, and I don’t know how to get good information other than just waiting and seeing who gets more traction.

Organizations (non-tax-deductible)

Last year, I didn’t write much about non-tax-deductible orgs; I just said I’d investigate more later. I ended up not doing much investigation, but I ruled out donating to most 501(c)(4)s. My issues with most of them are:

  • Their preferred policies don’t look like they’d have a big enough impact on extinction risk.
  • They’re cagey about talking about extinction risk, and I don’t trust them to share my priorities.

I did talk to them a bit, but I’m concerned that they are sufficiently politically savvy that they know how to make it sound to me like they share my priorities even when they actually don’t.7

There are two exceptions for 501(c)(4)s that I consider sufficiently well-aligned: ControlAI and Encode AI.

In terms of what I’ve directly observed, ControlAI is my favorite, due to their transparent commitment to focus on ASI extinction risk and their clearly stated policy priorities. However, readers should be aware that Oliver Habryka (who I have considerable respect for) alleges (1, 2, 3) that ControlAI has on balance prevented advocacy from happening, and has behaved adversarially toward other advocates, who haven’t talked about this out of fear of retribution. I don’t know the details, and it doesn’t match what I’ve seen of ControlAI’s public behavior, but I take Habryka seriously, so I don’t know where that leaves me.

The AI Policy Network is a third 501(c)(4) worth mentioning. Tax records show that they have spent significant money lobbying in favor of AI regulations. I like looking at tax records because you can’t fake them. My main hesitation is that I don’t know whether they would still support the much stronger sorts of regulations that would actually move the needle on extinction risk.

I’m not donating to any 501(c)(4)s at this time.

Where I’m donating

I’m still unsure what sort of work to prioritize, but I lean toward public-focused pause advocacy and grassroots organizing.

Last year, I donated to PauseAI US. I wrote: “PauseAI US is serving some important functions that nobody else is on top of.” That’s not true anymore. Evitable, People for a Pause, and Stop the AI Race now exist in the same space as PauseAI US; I’m somewhat more optimistic about their ability to succeed (and especially to form coalitions with others in AI safety).

Out of the advocacy orgs on my list, I have the most trust in Evitable’s leadership in terms of their ability to understand AI risk, think clearly, and prioritize the right things. In terms of competence at execution, Evitable doesn’t stand out as clearly the best, but they don’t look worse either, so the understanding/thinking/prioritization is the decisive factor for me.

This year, I am donating $55,000 to Evitable.

Notes

  1. I am not optimistic in absolute terms. We probably won’t get a pause. But my odds have gone up from, I don’t know, 15% to 18%? I don’t have a deep model here, but those numbers feel about right. ↩

  2. Strategic competence is only good if it’s directed at a good goal, but it’s conceivable that vague human intuitions + strategic competence = ethics solved, or something like that. ↩

  3. I’m still not particularly confident about AI pause advocacy. There are always many sign-flipping considerations. But it seems like a better bet than anything else. ↩

  4. Stop the AI Race is tentatively my second favorite, although like I said in their section, I don’t know much about them. ↩

  5. As an aside, I do not take people’s behavior on Twitter as much evidence about their character. Twitter turns people insane. On numerous occasions, I followed the Twitter account of some public intellectual whose work I enjoy, expecting my feed to fill with thoughtful commentary on interesting issues. Instead, my Twitter feed became a horror show. These public intellectuals are not deranged in real life, but Twitter makes them sound deranged because it’s a corrosive environment. ↩

  6. It seems bad that so much of the AI safety field has financial or social connections to AI companies, and I can see the argument for distancing yourself from anyone who’s associated with AI companies in any way. ↩

  7. One org I talked to was pretty transparent about the fact that their P(doom) is lower than mine, or at least that particular org employee’s P(doom) was lower than mine. I appreciated that; it made me trust them more. ↩