Stop Chasing Views: How to Reduce x-Risk as an AI Safety Content Creator

Many of the content creator fellows at plzdontkillus found my thoughts useful when I visited two weeks ago, so I’m now sharing a write-up here.

Many thanks to Maggie Munroe (FLI) and Chana Messinger (80,000 Hours) for their feedback on an earlier draft. Cross-posted to LessWrong.

As an x-risk content creator, your job is to increase the number of good actions that your viewers take and to increase the goodness of those actions.

Here's how to think about impact, your types of viewers, what calls to action to make, how to talk about x-risk, and what to do if you have an existing following.

Impact

As a content creator, your impact is indirect. Your impact lies in the impact that your viewers have that they wouldn't have had without you. Whatever your viewers do because of you that they wouldn't have otherwise done, that is your impact. Your job is to increase the number of good actions that your viewers take and to increase the goodness of each of those actions. You can model this as

Your Impact = Number of Views x Impact per View,

And the amount of impact that each view has can be seen as: Does it get the person to take action, and how good are those actions at reducing existential risk?

Actions can be pretty broad. For example, people might talk to others about AI Safety, change their voting behavior, donate money, or switch careers. In order for them to change such behaviors because of you, you will first need to change their mind. E.g. their beliefs and attitudes. This is necessary but not sufficient for behavior change. They can be as worried as you want about the future of humanity, but if they don’t change any behavior, then they haven’t reduced x-risk. And if they haven’t reduced x-risk, then you haven’t reduced x-risk either.

So while there are various ways for you to be impactful as a content creator, it will always have to end with people taking good actions because of you, and probably often involve you changing their views about AI in the process.

Two Types of Viewers

There are 2 types of people your videos might reach.

The first type includes those with the potential to switch careers and pursue a full-time career in AI safety. This applies to many people, but it depends on their skills and life circumstances, and it helps to have a technical background. However, as many of you are showing, a technical background is not needed to have a full-time impact on AI safety.

The second type includes people who, for whatever reason, are not in a position to pursue a full-time career in AI safety. These people might contribute in other ways. For example:

• Part-time or voluntary work in AI safety.
• Donating to the AI safety community.
• Content creation with a focus on AI safety.
• Personal outreach and conversations with friends (real life, whatsapp, texting, etc)
• Political engagement (voting, writing to politicians, signing public letters, attending protests).

By and large, these part-time contributions types are communication, advocacy, and donations. I might be missing stuff, though!

Three Types of Calls to Action

There are 3 types of things you can want from your viewers:

1. You could want them to just be informed with no specific call to action. The theory of change here is that at some point in the future, they will have some important thing that matters, and they might, for example, be more likely to vote for a political candidate that is favorable to AI safety, or they might be more likely to go to a protest about AI safety in the future. But you do not ask this of them, and potentially never will. Or even if you might in the future, for now you're just focused on building an AI safety-aware following. This style can be nice because people often react negatively to feeling like they're being sold something.
2. You could have direct-impact calls to action to your viewers, such as "Sign this public letter," "Donate to this organization," or "Go to this specific protest." You can find a list of calls to action for different audiences here: https://betterpath.ai/what-you-can-do/. There exist other lists from various other organizations too.
3. You could have a call to action towards the funnel of AI safety. You could send people off asking them to look at or sign up to Lens Academy or Blue Dot Impact. I find this model exciting because it means as a content creator, you only need to get your viewer interested enough to go to a website, and then it's that platform’s job to get them excited enough to engage more and more, and then get them to take a full 25-hour course, which is what really makes them ready to have a big contribution. I’d recommend this CTA for both people with full-time career potential and those with part-time contribution potential. Compared to the second category (direct-impact CTAs), I’d argue that those simple actions, like signing letters and donating, are more likely to consistently happen after people have deeply engaged with AI Safety for 25h, so even if those actions are the goal, an introductory course seems like a good starting point.

Three Approaches to Raising Concern About Existential Risk

I can see 3 ways to get people to worry about existential risk from advanced AI:

1. Go straight for the goal, talking about x-risk. This is what books like "If Anyone Builds It, Everyone Dies" and releases like AI2027 do. This is also what some AI safety videos from e.g. AI in Context and Drew Spartz's Species do. They will start and end the video talking about extinction risk.
2. Hook into people’s existing worries about AI, then move towards x-risk. Questionnaires show that people do not really start out caring about human extinction. They care about things like misinformation, deepfakes, job loss, scams, manipulation, and to some extent, power concentration or authoritarianism. What you can do is hook into these things and then try to get them to switch to more existential risks from that point on.
3. Talk about a different topic, attract a following with that, and occasionally mention AI safety. This seems potentially useful in that it can create wide societal support for AI safety, where it is associated not with a single political system or a single type of person, but is a worry that is broadly spread throughout the world. This might not work with many audiences, because usually it seems good to focus on one specific niche, but if your audience follows you for e.g. intellectual commentary about the world, I suspect this might work.

Moving from everyday AI worries to x-risk

If you go for option 2, where you meet people where they're at and focus on existing worries before you start talking about existential risk, it's important to think about how you're gonna move them from that starting point to where you want them to be. Some people think just spending attention on that starting point might be enough to have a positive impact and reduce existential risk. I'm not particularly convinced by that. I think you need to find actual ways to get people to care about existential risk and then get them to take actions about existential risk. Again, if people do not change their behaviors, they can’t reduce existential risk.

I can see a couple of ways to approach this, and I'm not sure which ones work well:

1. Within a video, you can start from something that people already care about, like job loss, and make that transition to existential risk within a few minutes. Some people seem good at doing this, and it seems worth experimenting with.
2. You can do some videos on easy-to-understand, everyday worries that are more introductory topics, and some videos on existential risk. Whether this works well, I have no idea, but it seems not harmful to experiment with. Mind you, you will likely get to see that your beginner-friendly videos get more views than the videos about existential risk. Do not take this as a bad sign. This is to be expected, and this does not mean that the beginner-friendly videos are doing a better job than your x-risk videos. They're simply doing a different job. This is a funnel. The top of the funnel of beginner-friendly videos will always get more views than the bottom-of-the-funnel advanced videos. What you're aiming for is behavior change with respect to existential risk, not views for the sake of views. That requires people coming out of the bottom of the funnel. The top of the funnel is just there to support that.
3. Have some channels by some creators focus on more tangible everyday worries and have other channels by other creators focus on existential risks. This might work, maybe, maybe, but it would need to be in a way that seeing the videos from a more tangible channel makes people more likely to watch the videos from an existential risk channel. It seems like it would probably be very helpful if the video channel that does content on everyday worries still mentions the existential worries, just kind of normalizes them and supports them with some kind of validity.
4. Hook people with everyday worries, like job loss or deepfakes, and then have a call to action around that – not around x-risk – sending them off to Lens Academy or Blue Dot with a special program on their website that is catered to that specific problem. For example, if somebody comes in worried about job loss, tell them, "Do you wanna learn what you can do to not lose your job to AI? Take this free 30-minute course." And then in that 30-minute course, let’s actually provide them with that value that is solving their actual problem. Their problem is, "I don't wanna lose my job." Our (Lens, Bluedot, I don’t care. Someone in the AI Safety community) solution is: "Cool, we'll teach you how AI works, we'll teach you how automation works, we'll teach you how AI can become embodied in the physical world. We'll teach you how feedback loops work and how competitive dynamics work. Based on that, you’ll be able to understand which jobs will be automated first and which ones will be automated last. Thus, you know which job you should choose if you don't wanna lose your job right away. We've now solved their problem, but in the process of doing so, we’re showing them they have another problem they weren’t aware of yet: that of x-risk. We have given them all the underlying nuggets for understanding existential risk. And we will then make the connection, make the bridge, and extrapolate from there and say, 'Hey, this is the first thing. See how all of these trends are continuing? We're actually gonna lose control and we're probably gonna die.' So now they're worried about existential risk. Now they're aware of this new problem that they already had but weren't aware of, and now they might be willing to take a full course about existential risk and then be routed to impactful actions after that."
   The model is: people have salient problems. Provide a solution to that problem. Whilst providing that solution, make them aware of another problem they're having that they weren't aware of. And then provide a solution to that second problem. It's really hard sometimes to go straight for that second solution because they're not yet aware of this second problem, and getting them to be aware of this second problem might take time.
   Another example. Same model, different problem: What we can do is tell them about deepfakes. People care about deepfakes. Have a call to action to a little quiz—a 2-minute quiz. Can you detect a deepfake? Can you distinguish a deepfake from a real picture? Can you detect a deepfake video from a real video? Cool. We have them do a little quiz like that on a website. Now we have them on a website. Now we can get their email address, capture them, and lure them further into the space.
   It's really hard to get people to engage deeply on a platform that is optimized for short-form content like TikTok, Instagram, or YouTube. But once you can route them to a different platform that is owned by the AI safety community—and I don't care if that's Lens Academy or 80,000 Hours or BlueDot—now we can reel them in and start digging deeper. Now we can give them long-form explainers. Now we can have them have conversations with other human beings. But we gotta get them out of short form, where we can only engage with them for 2 minutes at a time, and get them to a place where we can hook them long enough and provide them with enough mental calm and bandwidth to deeply engage with this topic for hours at a time in a way that moves from passive consumption to active engagement.

If You Have an Existing Following

If you have an existing following, you might want to take your time to slowly introduce your following to existential risk. You might want to take them on a journey with you. Depending on the circumstances, your journey might be faster than that of your following. So you might want to find your own journey and then slowly repeat that in a simulated fashion for your following.

If you have an existing following, you are going to lose part of your following. This is inevitable. You will also gain a new following. It could be that your new following is smaller than it was before. It could be that it's larger than before. This doesn't matter. If you want to reduce existential risk and your old videos with your old following were not reducing existential risk, then reducing the number of viewers you get but talking about existential risk means you’re increasing your impact. You are doing a good job. Just because you're losing viewers doesn't mean you are being less effective. You are doing something different that works in a different way, and it's a different game that you're playing. Please do not let your mind trick you into believing views are what matters.

Measuring Your Impact

There are different ways to try to be impactful. Those ways can be effective or ineffective at having impact, and their impact can be measurable or unmeasurable. It is possible to have a lot of impact in a way that is not measurable. However, this seems hard to achieve to me because you cannot measure whether you're having success. So how do you know if you're succeeding? And if you don’t have a measure of success, then how can you improve? This is sad, and as a community we should try and find better ways of measuring those things that cannot currently be measured.

There are also ways of having impact as a content creator that are measurable. In my eyes, the way to measure that is not through views, likes, or comments. Views, likes, or comments show you if people like what you're making and if people are engaging with what you're making. This is not your goal. Your goal is to have as many people as possible take as good actions as possible. If you do not get behavior change, then you haven’t reduced x-risk. Thus, you need to measure the behavior change if you want to know if you're being successful. Don't measure views; measure behavior.

One way to do this is to have call-to-actions and measure how many people act on them (e.g. how many people click on a relevant link). The goal should be something like maximizing the number of monthly clicks you get, rather than maximizing the number of clicks from each video. If you make sure that the place you're linking to is actually good and is likely to reel them into AI safety and make them do impactful things, then in my opinion it is enough for you to measure the number of people that you're sending that way and not worry about their later actions.

It is very easy to not care about call-to-action clicks and/or not track them. It’s easy to revert to using the feedback signals you're getting from the content creation platform, like views and likes, and start optimizing for the number of views you get. But the thing is, if all you want is views then there are other topics with which you can get more views. If your goal is to reduce the chance that we all die, then do not maximize the number of views you get. Maximize your impact. I think the easiest (though not the only) way to do that is by maximizing the number of people that click on a link that is impactful for them to click on. There too, though, don’t just maximize for CTA clicks, or you might be tempted to replace a hard, impactful CTA with an easier, more clickable, but less impactful CTA.

Getting more actions from your viewers involves getting more views, but it also involves getting more of those viewers to click on things. So what you get is:

Number of Clicks on a CTA = Number of Views x Percentage of Viewers that Click

Where should you put your call to action? You could put it at the end. This might mean you get a lot of views but not a lot of clicks per view. You can also put a call to action at the start of a video. This means your video will probably perform less well—you will not get as many views, but a bigger percentage of those people might click on the link. Which one is better? I don’t know. Just be aware: it doesn't matter how many views you get; it matters how many people click on that link. However, it's about the monthly number of clicks, not about the number of clicks on a specific video.

You will likely want to do some videos or quite a lot of videos without any call to action. But it seems very alien to me to have no call to actions in any video because then you have no measure of success. You probably want some videos without a call to action that are just optimized for virality while still doing something with AI safety. This builds your audience, and then you have other videos with a call to action that convert.

Conclusion

I’m very excited about people trying their hands at AI Safety content creation. You do not need to get it right, right away. You do not need to ever get it fully right. But I'm sure you all have an internal drive to get better over time, so please make sure that you aim for things that actually reduces existential risk. Have fun creating!

A lot of my thinking on this is informed by marketing books like Day Trading Attention and $100M Leads, and also my broader experiences working full-time in AI Safety field building as the founder of Lens Academy. Bear in mind that I’m neither a marketeer, nor a content creator, though. Also bear in mind that I’ve written up these thoughts as an addition to the thoughts many content creators already seem to have. If someone would start from scratch and only follow this guide, that’d probably not be very successful.