PauseAI UK is holding the March Against AI Extinction in London on Saturday 21 November. Joseph Miller explains why here. Should you attend? In this article, I will attempt to shed light on this with a back-of-the-envelope calculation.
Rough BOTEC
1. Setup
The march will take 3 hours, excluding travel. The goal is to compare an afternoon at the march with some number of hours of external AI safety research.
Now, attending the protest is useful because it drives AI x-risk press coverage. So we need some exchange rate between that and research.
Here's my judgement call: I would be equally happy to either:
Increase the amount of external AI safety research over the next 6 months by 1%.
Increase the amount of AI x-risk press coverage over the next 6 months by 1.5%.
That is, the model I'm using looks something like:
How much will attending increase press coverage, percentage-wise?
Divide that by 1.5, to get the equivalently valuable increase in research.
How many research hours would yield that increase in research?
2. What % does one marcher add to press coverage?
Crowd size. I expect about N = 670 people will attend the march. PauseAI had a march in February which drew 300–500, and this one should be bigger given the increased salience. The Luma page has 1,200 sign-ups, and a 50% turnout seems reasonable. So one more marcher raises the crowd by 1/670 ≈ 0.15%.
Coverage of this march. Let's suppose the press coverage for this march grows like N^ε for some ε. So increasing the crowd by 0.15% would increase the coverage by ε × 0.15%. Claude estimates that ε ≈ 0.49. That means that attending adds +0.073% to this march's coverage.[1]
How Claude estimated ε = 0.49
Method. For seven AI x-risk protests between June 2025 and October 2026, I took the crowd size reported in coverage (the geometric midpoint of the reported range) and listed the news coverage I could find through web searches. Syndicated copies count once.
I scored each piece by how much we should care about it: roughly, the number of people who actually saw that piece, in thousands, times a multiplier for who they are. Readers who are AI or policy insiders count 2–5×, a hostile framing counts 0.5×, and a reused wire photo illustrating some other story counts 0.2×. The readership figures are my guesses, not measurements.
March Against the Machines, London, 28 Feb 2026. Size 300–500 (PauseAI: "more than 300"; Pull the Plug: "up to 500") → 390. Total 188.
Result. Fitting a straight line to log(coverage) against log(crowd size):
Coverage measure
ε
Number of outlets
0.73 (90% bootstrap interval 0.51–0.94)
Scored, without Boise
0.49 (90% bootstrap interval −0.05 to 0.94)
Scored, all seven
1.46 (Boise's single student-paper story pulls this up)
Once scored by how much we care, each protest's total is dominated by whether one large outlet showed up: Business Insider, ABC7, the WSJ. A 35-person rally that got local TV scores about as high as a 170-person protest that got Channel 4. Crowd size only weakly predicts that one big hit, so I use the fit without Boise, ε ≈ 0.49, while noting that these data can't rule out values up to about 0.9.
Caveats. Seven points is a small sample. The coverage lists come from web searches, not a systematic database. The readership figures are guesses. Reported crowd sizes vary up to 3× between sources. Size is confounded with topicality: the 16 September protest rode the news about Jacob Coxon, and the February march was billed as the largest ever.
All AI x-risk coverage. Now, this march isn't the only thing that drives AI x-risk press coverage. There are also warning shots, high-profile resignations, statements, letters, etc. My guess is that protests will make up at least 3% of AI x-risk press coverage over the next 6 months. There are about 3–6 big protests a year, or 1–3 every six months. So this should be 30% of the protest coverage over the next six months. That makes this march 30% × 3% = 0.9% of all coverage. Therefore, attending the march would increase total x-risk coverage by 0.073% × 0.9% ≈ 0.00066%.
3. How many hours of research is that worth?
Based on my judgement call above, research is 1.5× as good as press coverage. So +0.00066% coverage is worth the same as +0.00044% research.
Now, how much AI safety research is there? McAleese (2025) counts about 1,100 FTE: 600 technical and 500 non-technical. But this doesn’t account for academics, independent researchers, fellows (e.g. MATS, Astra). Or the ~25%/yr growth. My estimate is that over the next 6 months this will average about 1,900 FTE. At 40-hour weeks over 6 months, that’s about 1.9M hours. So increasing 1.9M hours by +0.00044% would take about 8 hours of research. That's about twice the ~4 hours it takes to attend the march.
Moreover, the BOTEC treats you as an average marcher. But if you’re an AI safety researcher, your attendance is worth more than that: you can give a speech, talk to journalists, write about it afterwards, and lend credibility. My guess is that your impact from marching is at least 3× the average marcher’s.
And the real kicker for attending the march is that you can bring friends and scale your impact almost linearly. That's much harder with technical research — you can't linearly scale your research impact by bringing your friends into work for a few hours.
So if you're an AI safety researcher, and you bring three friends to the march, that would be equivalent to ~48 hours of AI safety research, an entire working week.
Conclusion
The BOTEC is admittedly pretty rough. I'm not sure about my judgement call comparing press coverage to AI safety research. I could be persuaded that it's off by a factor in either direction. But I hope it was informative.
If I can be earnest for a moment: things might turn pretty bad pretty soon, and this might be the last year to make a difference. I don't know what I'm actually supposed to do in this situation.
Maybe, in hindsight we'll know what we were supposed to do. Maybe not even then. Either way, I'll feel much happier if — instead of "I wrote some Google Docs" — I can say "I wrote some Google Docs, and I walked with hundreds of strangers to say: this is fucking insane, let’s not do this."
I think 0.49 seems reasonable. I tried to cross-check this with the existing literature and the best published estimate is Wouters (2013). They took 4,582 protests and their crowd sizes from Brussels police records (2003–2010). They matched this against coverage on the two main evening news programs. The crowd size predicted both the odds that a protest was covered (N^0.55), and the seconds of coverage it received, if covered (N^0.16). It's somewhat complicated to combine these into an overall elasticity ε. You can try straightforwardly adding them, but the more principled combination is ε = 0.55 × (1 − 22%) + 0.16 ≈ 0.43 + 0.16 ≈ 0.6. Here, 22% is the probability that a 670-crowd protest is covered, which you can fit from the data in the study. So N^0.6 is our power law for seconds of coverage on national evening news.
PauseAI UK is holding the March Against AI Extinction in London on Saturday 21 November. Joseph Miller explains why here. Should you attend? In this article, I will attempt to shed light on this with a back-of-the-envelope calculation.
Rough BOTEC
1. Setup
The march will take 3 hours, excluding travel. The goal is to compare an afternoon at the march with some number of hours of external AI safety research.
Now, attending the protest is useful because it drives AI x-risk press coverage. So we need some exchange rate between that and research.
Here's my judgement call: I would be equally happy to either:
That is, the model I'm using looks something like:
So our plan is:
2. What % does one marcher add to press coverage?
Crowd size. I expect about N = 670 people will attend the march. PauseAI had a march in February which drew 300–500, and this one should be bigger given the increased salience. The Luma page has 1,200 sign-ups, and a 50% turnout seems reasonable. So one more marcher raises the crowd by 1/670 ≈ 0.15%.
Coverage of this march. Let's suppose the press coverage for this march grows like N^ε for some ε. So increasing the crowd by 0.15% would increase the coverage by ε × 0.15%. Claude estimates that ε ≈ 0.49. That means that attending adds +0.073% to this march's coverage.[1]
How Claude estimated ε = 0.49
Method. For seven AI x-risk protests between June 2025 and October 2026, I took the crowd size reported in coverage (the geometric midpoint of the reported range) and listed the news coverage I could find through web searches. Syndicated copies count once.
I scored each piece by how much we should care about it: roughly, the number of people who actually saw that piece, in thousands, times a multiplier for who they are. Readers who are AI or policy insiders count 2–5×, a hostile framing counts 0.5×, and a reused wire photo illustrating some other story counts 0.2×. The readership figures are my guesses, not measurements.
March Against the Machines, London, 28 Feb 2026. Size 300–500 (PauseAI: "more than 300"; Pull the Plug: "up to 500") → 390. Total 188.
Stop the AI Race, San Francisco, 11 Jul 2026. Size 100–350 (SF Standard: "more than 100"; ExplainX: "~350") → 190. Total 67.
Google DeepMind protest, London, 30 Jun 2025. Size 60–130 (Digital Information World: "over 5 dozen"; Luma: 129 went) → 90. Total 57.
Stop the AI Race, San Francisco, 21 Mar 2026. Size 40–200 (ABC7: "dozens"; exchange4media: "nearly 200") → 90. Total 52.
Stop the AI Race at Dreamforce, San Francisco, 17 Sep 2026. Size about 20–60 (Legal Insurrection: "a small but vocal contingent") → 35. Total 45.
PauseAI emergency protest, Downing Street, 16 Sep 2026. Size about 100–300 (Channel 4: "hundreds of protesters") → 170. Total 42.
PauseAI Boise, 24 Oct 2026. Size about 10–30 (the first AI safety protest in Idaho; size not reported) → 17. Total 0.5.
Result. Fitting a straight line to log(coverage) against log(crowd size):
Coverage measure
ε
Number of outlets
0.73 (90% bootstrap interval 0.51–0.94)
Scored, without Boise
0.49 (90% bootstrap interval −0.05 to 0.94)
Scored, all seven
1.46 (Boise's single student-paper story pulls this up)
Once scored by how much we care, each protest's total is dominated by whether one large outlet showed up: Business Insider, ABC7, the WSJ. A 35-person rally that got local TV scores about as high as a 170-person protest that got Channel 4. Crowd size only weakly predicts that one big hit, so I use the fit without Boise, ε ≈ 0.49, while noting that these data can't rule out values up to about 0.9.
Caveats. Seven points is a small sample. The coverage lists come from web searches, not a systematic database. The readership figures are guesses. Reported crowd sizes vary up to 3× between sources. Size is confounded with topicality: the 16 September protest rode the news about Jacob Coxon, and the February march was billed as the largest ever.
All AI x-risk coverage. Now, this march isn't the only thing that drives AI x-risk press coverage. There are also warning shots, high-profile resignations, statements, letters, etc. My guess is that protests will make up at least 3% of AI x-risk press coverage over the next 6 months. There are about 3–6 big protests a year, or 1–3 every six months. So this should be 30% of the protest coverage over the next six months. That makes this march 30% × 3% = 0.9% of all coverage. Therefore, attending the march would increase total x-risk coverage by 0.073% × 0.9% ≈ 0.00066%.
3. How many hours of research is that worth?
Based on my judgement call above, research is 1.5× as good as press coverage. So +0.00066% coverage is worth the same as +0.00044% research.
Now, how much AI safety research is there? McAleese (2025) counts about 1,100 FTE: 600 technical and 500 non-technical. But this doesn’t account for academics, independent researchers, fellows (e.g. MATS, Astra). Or the ~25%/yr growth. My estimate is that over the next 6 months this will average about 1,900 FTE. At 40-hour weeks over 6 months, that’s about 1.9M hours. So increasing 1.9M hours by +0.00044% would take about 8 hours of research. That's about twice the ~4 hours it takes to attend the march.
Moreover, the BOTEC treats you as an average marcher. But if you’re an AI safety researcher, your attendance is worth more than that: you can give a speech, talk to journalists, write about it afterwards, and lend credibility. My guess is that your impact from marching is at least 3× the average marcher’s.
And the real kicker for attending the march is that you can bring friends and scale your impact almost linearly. That's much harder with technical research — you can't linearly scale your research impact by bringing your friends into work for a few hours.
So if you're an AI safety researcher, and you bring three friends to the march, that would be equivalent to ~48 hours of AI safety research, an entire working week.
Conclusion
The BOTEC is admittedly pretty rough. I'm not sure about my judgement call comparing press coverage to AI safety research. I could be persuaded that it's off by a factor in either direction. But I hope it was informative.
If I can be earnest for a moment: things might turn pretty bad pretty soon, and this might be the last year to make a difference. I don't know what I'm actually supposed to do in this situation.
Maybe, in hindsight we'll know what we were supposed to do. Maybe not even then. Either way, I'll feel much happier if — instead of "I wrote some Google Docs" — I can say "I wrote some Google Docs, and I walked with hundreds of strangers to say: this is fucking insane, let’s not do this."
Going to Work (1943) by L. S. Lowry.
Details
I think 0.49 seems reasonable. I tried to cross-check this with the existing literature and the best published estimate is Wouters (2013). They took 4,582 protests and their crowd sizes from Brussels police records (2003–2010). They matched this against coverage on the two main evening news programs. The crowd size predicted both the odds that a protest was covered (N^0.55), and the seconds of coverage it received, if covered (N^0.16). It's somewhat complicated to combine these into an overall elasticity ε. You can try straightforwardly adding them, but the more principled combination is ε = 0.55 × (1 − 22%) + 0.16 ≈ 0.43 + 0.16 ≈ 0.6. Here, 22% is the probability that a 670-crowd protest is covered, which you can fit from the data in the study. So N^0.6 is our power law for seconds of coverage on national evening news.