AI Safety Communication Tips

My name is Brett Bricker. My expertise is in communication. I earned a PhD in rhetoric, and my primary academic emphasis is public consumption of scientific discourse (climate change, autism-MMR vaccine conspiracies, COVID skepticism, cyber-hacking, etc.). I’ve done some fieldwork in those areas, including communication coaching. I’m pretty new to the AI safety field, so thank you for welcoming me, and hearing me out.

Overall, I’m impressed with the public-facing components of the AI safety field. A lot of the technical scientists who are “going public” are authentic, play well with diverse audiences, and do a pretty good job of avoiding excessive jargon. If you are already doing that work and want to brush up; or, if you are considering developing a more public-facing image, here are some tips.

1. Imagine your audience.

Before you start preparing, spend some time thinking about who will be consuming the message. Actually imagine them consuming your content. Will it be on a screen? Will the audience be captive, like in a classroom? Or will they be a voluntary audience, meaning they can turn you off whenever they want? What percentage of your audience do you imagine will be multitasking?

Once you’ve done that, consider: what do you know, and what are you assuming about your audience? What do they already know about AI? What concerns brought them to this conversation? What might make them skeptical of your argument?

Sometimes the forum itself will tell you a lot about the audience. The 80k podcast listeners might have a specific worldview and background knowledge. Joe Rogan listeners will have much more heterogeneous (and lay) priors. An interview on CNN, or a BBC one-line quote will reach much broader, and thus much different, audiences.

It’s easy to lose track of how much background knowledge goes into the conversations that scientists have with colleagues. A term a technical scientist uses every day might be unfamiliar to someone who is smart and interested, just inexperienced in the field. This is true for all sorts of jargon (TedAI, RSI, AGI, RSI, etc.) but applies to even the simplest messages. When a public speaker says, “Hugging Face,” some folks will have zero background, others will think of the emoji, and others will be able to fill in a lot more of what those two words fully represent. For those that have no background, give them a way into the conversation. Explain the concept, offer an example, use an extended metaphor, and make the connection to something they care about.

The goal is to make your argument understandable without making it misleading. That requires a serious consideration of who constitutes the receiving audience.

2. Set a goal.

What do you want your audience to take away from the conversation? Try to answer that question in one sentence. After consuming lots of AI safety content, a recurring theme is that the (potentially, implicit) goals of AI safety communicators cited most in the media are to provide an accurate and frightening message about AI safety.

That makes sense. There are a lot of cases where this is exactly what the goal should be. Distinct goals are worth considering:

· Give people a set of responses to the “China problem,”

· Teach audiences how to think about probability and uncertainty,

· Introduce people to policy solutions,

· Help people understand why a new “warning shot” is different from the last one, etc.

You probably won’t be able to communicate everything you know. That’s fine. Pick a useful, realistic goal, and give your audience what they need to get there.

3. Tell a story.

It’s possible to retain a commitment to accuracy, while also developing a narrative structure to the message. A good story has a plot, characters, and points to a specific lesson. Give your audience something concrete to follow. If you’re explaining your research, walk them through what drew you to the dilemma, the specific question you were trying to answer, what you expected to find, and what happened. What surprised you? Why did it change how you think about the problem? In the AI safety environment, it is increasingly necessary to distinguish the story you’re telling from other stories that are already being told.

Every audience might need a different story, and every question gives you an opportunity to tell a new story. If someone asks you, “what brought you to this field,” there are a multitude of answers that can be drawn from. If you say, “I came to this field as a rationalist,” that is a choice, and that choice has costs/benefits. Saying, “I came to this field because I loved math,” is also a choice, with different costs and benefits. If both are true for you, both are options, and persuasion/credibility-building should drive the choice. If someone asks you, “what frightens you,” there are lots of options here, as well. Choose an example that helps your audience understand your answer. There isn’t a lot of evidence that existential risk is the most persuasive frame, but that research is limited and likely needs an update to account for recent warning shots.

4. Embrace evidence-based messaging.

Be willing to ask whether our communication is actually working. Effective messaging requires systematic testing. There’s already-existing work on this, I encourage folks to draw from. An explanation can feel clear to you, yet still leaves listeners confused. An argument can be persuasive to your colleagues and do very little for the people we’re trying to reach.

This is where feedback, surveys, and focus groups can be useful. With a little funding and effort, it’s possible to vet phrasing and pre-test messages. This can be done via survey and/or focus groups. What did listeners understand? What did they remember? Did they draw a conclusion you didn’t intend?

My anecdotal take based on AI-safety content consumption has led me to believe that most AI safety communicators have cached takes to common questions. It’s worth taking the extra step to consider if those are landing. Be open to learning that your cached take isn’t making your case stronger and be willing to adapt.

5. Practice!

Say it out loud. Record yourself and listen. Give yourself a time limit. Ask someone to interrupt you with a question. You’ll learn a lot about your explanation when you have to deliver it to another person.

I’d especially encourage you to practice with someone who doesn’t share your background. Ask what they understood, where they wanted more explanation, and what felt unconvincing. Then make a few changes and try again. Leave time to practice the questions, too, including the ones you hope nobody asks.

If a practice partner would be helpful, I’d be happy to work with you. Through a BlueDot rapid grant, I’m offering communication training, mock interviews, message testing through focus groups and surveys, and public speaking review sessions to AI safety researchers at no cost. Bring an upcoming presentation, an interview, or an idea you’re having trouble explaining. We can work on it together. I want to help. Contact me via this website or my personal email: [last name]312 at gmail.

I’m not encouraging speakers to hide their connection to EA/rationalism. When asked directly, never lie. But, open-ended questions offer a range of options, and I’m asking for resonance and persuasiveness to guide the choice of the communicator.

Stealing this term from Saheb Gulati. Think of these like heuristic fallbacks that represent default answers to common questions.