A few days ago, Austin Chen and Marcus Abramovitch published How cost-effective are AI safety YouTubers?, an "Early work on "GiveWell for AI Safety"", ranking different interventions in the AI Safety Video space, using a framework that measured impact by basically multiplying watchtime by three quality factors (Quality of Audience, Fidelity of Message and Alignment of Message).

Quality-adjusted viewer minute = Views × Video length × Watch % × Qa × Qf × Qm

The goal of this post is to explain to what extent I think this framework is useful, things I think it got wrong, and provide some additional criteria that I would personally want to see in a more comprehensive "AI Safety Givewell for video work".

tl;dr:

1. I think Austin and Marcus' framework has a lot of good elements, especially the three factors Quality, Fidelity and Alignment of message.
2. Viewer Minutes is the wrong proxy if what we actually care about is people taking action, and instead we should probably try to multiply views by successful Call-To-Actions / views.
3. Instead of comparing creators that do quite different things, it might make sense to instead compare creators in the same category, and see which are bottlenecks in maximizing throughput through the AI Safety video -> AI Safety talent funnel
4. All else being equal, original content should get a higher score, because it would, on the margin, change people's minds more and add diversity.

What The Framework Gets Right

I think the three criteria (Quality of Audience, Fidelity of Message and Alignment of Message) are useful and important factors to consider when trying to estimate the impact of these interventions.

When I tried to estimate these numbers for my YouTube content and TikTok clips, this led me to useful reflections such as:

• Am I actually reaching the kinds of people that I want to reach? What demographics actually watch my content?
• How impactful is it to share a 2-3 minute quote vs. a 3h podcast?
• How aligned am I with the content that I end up posting?

About two weeks ago I was also trying to come up with a list of criteria that would make sense to evaluate interventions on. After talking to friends in the space and someone at Open Phil, I ended up with things such as % of Safety content, production value, quality of audience and accuracy (which I shared with Marcus), and I think that "Quality of Audience, Fidelity of Message and Alignment of Message" capture a lot of these things quite neatly.

Limitation: Viewer Minutes

As multiple commenters pointed out, one of the key limitations of their framework is that they multiply quality scores by viewer minutes corresponding to the watchtime of the video (which in some cases was data from the creators themselves, and in other cases was estimated ^[1]).

As Michael Dickens mentions, the problem in doing that is that he suspects "impact is strongly sub-linear with minutes watched. The 100th minute watched for a particular viewer probably matters far less than the 1st minute watched."

As I understand it, Marcus' answer to this is that 1) even if the impact is sublinear it's not bad enough that approximating to linear makes a huge difference 2) since he's using watch-time (with a lot of real data from creators) he's not actually claiming it's proportional to length but proportional to the actual minutes people are watching 3) longer videos are much more likely to lead to action than shorter videos. To provide some quotes, Marcus said:

On the thought that impact is strongly sublinear per minute of video, I'd ask you to consider, when have you ever taken action due to a 0.1-1 min video? Compare this to a 10 min video and a 100 min podcast and now compare this to a book that takes ~1000 min to read. (source)

and

[...] I thought about this for a while, and what you are really trying to buy is engagement/thought or career change or protests or letters to the government, etc.. This is why in the future, article reads/book reads are going to be weighted higher. Since reading a book is more effortful/engaging than passively watching a video. [...] (source)

Given these comments, it seems that the thing that Marcus actually cares about is people taking action, career changes, protests and letters to the government.

However, if that's what Marcus cares about, I claim that his "AI Safety Givewell" recommendations should have at the very top a creator that isn't even on the list right now, called Siliconversations.

Siliconversations And The Art of Call to Action

On May 30th 2025, a relatively unknown YouTube channel called Siliconversations released  a video called This 17-Second Trick Could Stop AI From Killing You, showcasing that it only takes 17 seconds to email your representatives about risks from AI, using a tool made by ControlAI. 

His video about emailing representatives led to 2000+ emails sent to representatives with 80k views or less (source),  which means a ratio of actions per view of at least 2.5%.

The reason I think Siliconversations' video led to so many emails was that he was actually relentless in this video about sending emails, and that was the entire point of the video, instead of talking about AI risk in general, and having a link in the comments. 

In comparison, from talking to folks at ControlAI and Rational animations, I believe the (really great) Control AI x Rational Animations collaboration that got 1.4M views ended up in many fewer emails sent to representatives using ControlAI's tool, even though the video got ~18x more views.

Therefore, if what we actually care about is people taking action, then we should put some more weight on creators that provide concrete calls-to-action (CTA), especially the ones where a large fraction of CTA-views led to actual action (eg. emails to representatives), like Siliconversations.

To begin with, some similar analysis could be done for many of the creators that Marcus and Austin originally listed:

• Robert Miles recently did a video on AI Safety Career Advice. He links to 80k career advising in the description. 80k could tell us in the comments if they actually experienced an increase in traffic after this video came out.
• Similarly, 80k's AI In Context has a full section at the end on "What to do" saying to look at the links in description, and there are many links to eg. 80k's website, bluedot or controlai's website. Would be useful to study if there was any increase in traffic after the video got its first ~1M views.
• ControlAI could give information on how much increased traffic they observed on their website / to their tool for each of their collaborations.

Once all this data is gathered (or estimated), the one formula for measuring the impact of content based on how much it leads to "taking action" could be: 

Quality-adjusted views = Views × Qa × Qf × Qm x successful CTAs per view

However, I believe doing this might actually be a bit short-sighted, because although Siliconversations has been quite successful at getting people to email their representatives in one video, as Marcus also notes, some other kind of content might just be the first step in a career change that might take years.

Moving Talent Through The Video Pipeline

As Austin puts it, there are actually at least three ways an AI Safety video could have an impact:

1. Introducing people to a topic
2. Convincing people to take an action (change careers, donate)
3. Providing information to people working in the field

I think podcasts, especially technical ones like Dwarkesh / Cognitive Revolution or The Inside View are helpful with 3., but can also do 2. (eg. this interview with Yampolskiy led to a lot of people signing up for PauseAI). Targeted videos like Siliconversations definitely help with 2. And shortform content is helpful for 1^[2]. 

More generally, in another post I argued that in order to fully understand AI Safety arguments, you actually need a great deal of repeated exposure:

The first time I wanted to learn about AI 2027, I listened to Dwarkesh's interview of Daniel Kokotajlo and Scott Alexander to get a first intuition for it. I then read the full post while listening to the audio version, and was able to grasp many more details and nuances. A bit later, I watched Drew's AI 2027 video which made me feel the scenario through the animated timeline of events and visceral music. Finally, a month ago I watched 80k's video which made things even more concrete through the board game elements. And when I started cutting out clips from multiple Daniel Kokotajlo's interviews, I internalized the core elements of the story even more (though I still miss a lot of the background research).

What this means for me is that, instead of trying to compare many different interventions in the AI Safety video space that are quite different in nature, maybe it would make more sense to consider them as multiple steps of a funnel, and you can then compare each different step.

Categorizing The Video Pipeline

To give an analogy, if we were to think about AI Safety comms in a similar way to what is described in Prospecting for gold, I'd split things into:

• The people hunting for gold: those are the people who come up with new AI Safety ideas worth spreading. Think AI 2027, Scott Alexander, Tom Davidson's Takeoff speed model or Yudkowsky's and Soares' new book.
• The messengers going around telling the location of the gold once the gold has been found: think of these people as "amplifiers" of the original idea. Specific categories include:
  • Podcasts: Cognitive Revolution, AIRN, FLI podcast, Doom Debates or The Inside View, interviewing gold hunters Tom Davidson, Eliezer or Daniel Kokotajlo.
    • Short-form clips based on podcasts / interviews can be seen as amplifying the message from this messenger (eg. what I do on TikTok).

• Scenario visualisers: which includes Drew's AI 2027 or Joshua Clymer videos, Rational Animations' Alien Message or my video on "Scale was all we needed at first".
• Technical explainers: Robert Miles falls into this, but also some of the content by Rational Animations (eg. this recent one).

My claim is that it doesn't quite make sense to compare all of these apples to apples, because they're all useful steps in a diverse ecosystem, so it would make most sense to compare them by category if you really wanted to discuss cost-effectiveness, thinking about where the current bottleneck is in the entire pipeline (aka, what is something neglected that could increase the overall conversion of views -> AI Safety talent).

So an adjusted formula could be:

Quality-adjusted views = Views × Qa × Qf × Qm x successful CTAs per view x Qn

Where Qn could quantify whether something is neglected / a bottleneck in the AI Safety video space or not.

But one thing that I haven't mentioned, which makes things a little bit more complex, and at the same time even more exciting, is that it's possible for video content to actually be hunting gold, not simply amplifying.

Original Video Content

Alongside Siliconversations, one content creator who I think not everyone has heard about is Suzy Shepherd (known by EAs for having made that GWWC ad).

Last year she won some FLI contest with her short-film Writing Doom, which I believe she produced on a tight budget (< $20k, if not <$10k), and has now reached 500,000 views, with potentially many viewer minutes since the video is 27 minutes long.

Even without additional arguments, Suzy would already be a great contender for Marcus' and Austin's AI Safety Givewell recommendations.

But I'll make an even stronger recommendation, by emphasizing that Suzy did write this entire fiction herself. That's an original screenplay.

I think there's a slight difference between that and say AI 2027, since it's a different way of framing the original AI Safety arguments from Bostrom and Yudkowsky than original research.

But I still think that if we're trying to measure "how impactful is that view / $", it would make sense to think about whether the content is truly original, or if it's amplifying an idea that has already been amplified elsewhere (say a podcast or video about AI 2027).

To be more concrete, say someone on YouTube watches 5 videos on AI 2027 (like I did) and watches Writing Doom once. I think because Writing Doom adds diversity to the mix (by being about something different), and is on its own an original idea (not eg. amplifying a blogpost), the Writing Doom will change this person's mind much more than the average of these 5 AI 2027 videos ^[3].

So a final formula could be:

Quality-adjusted views = Views × Qa × Qf × Qm x successful CTAs per view x Qn x Qo

Where Qo is a factor indicating how original the content is.

Conclusion

1. I think Austin and Marcus' framework has a lot of good elements, especially the three factors Quality, Fidelity and Alignment of message.
2. Viewer Minutes is the wrong proxy if what we actually care about is people taking action, and instead we should probably try to multiply views by successful Call-To-Actions / views.
3. Instead of comparing creators that do quite different things, it might make sense to instead compare creators in the same category, and see which are bottlenecks in maximizing throughput through the AI Safety video -> AI Safety talent funnel
4. All else being equal, original content should get a higher score, because it would, on the margin, change people's minds more and add diversity.

1. ^{^}

   Note that, as I mention in a comment, some the data that was estimated / given was wrong by an order of magnitude, including the TikTok viewer minutes data and the amount of views of the FLI podcast.

2. ^{^}

   Especially if you're introducing a new audience (people on TikTok) to a topic, instead of other platforms where a lot of creators already exist (eg. YouTube).

3. ^{^}

   In particular, my intuition from talking to people who do watch AI Safety content on YouTube, is that they tend to be subscribed to all the different channels. So you could realistically expect that if all creators post about the same idea, it doesn't quite make sense to consider all the views from the different content as separate.