I’ve been working on a series of longer articles explaining and justifying the following principles of communicating AI risk to the general public, but in the interest of speed and brevity, I’ve decided to write a simplified guide here. Some of this will be based on articles I’ve already written and released, and some will be based on articles that are still in the works.
Principle 1- Jargon
It’s best to avoid jargon when communicating with the general public. Jargon is not easily accessible for non-experts, and unfamiliar words and phrases could lead to misunderstandings. However, when communicating about new technologies that are largely unprecedented, some jargon may be necessary. When you must use jargon, make sure you immediately follow the term with a simple and concise definition. Don’t stack multiple terms together into the same line, and don’t stack more jargon into your definition.
Be consistent with your jargon. Use the same terms consistently, until the public gets used to them. If you notice the public already seems familiar with a term, or experts are using the term often in public communication, continue to use the same term. When you are able, use catchy, memetic terms that carry some emotional weight. An example of this type of term is “the black-box problem” when discussing interpretability, because not only is “black box problem” rather catchy and descriptive, but a “black box” is already associated with frightening events such as plane crashes.
Principle 2- Sound Bites
Anticipate that anything you produce- whether written, audio, or audio/visual- will be broken up into short-form content such as tweets or tik-tok videos. Because you can anticipate that your content will be shredded into sound bites, it’s important that you ensure each argument can stand on its own. Nuance is still important, but try to incorporate caveats into a short space following the nuance you’ve introduced.
If you cannot make each argument stand on its own in just a few lines, make sure you balance this deficit with volume. Put out as much content you can, as often as you can, so that people are more likely to get a broader understanding of your arguments.
Principle 3- Examples
People will often want to hear examples of how, exactly, a rogue superintelligence could cause widespread destruction. These people may be skeptical if you say that you don’t know how an intelligence greater than yours will cause destruction- the chessmaster argument can feel like a copout. However, if you do give concrete examples, your audience may try to propose “patches” for any path of destruction they are given, which is not helpful because there is no end to the number of ways a superintelligence can cause destruction. Instead of getting bogged down in endless scifi projections, it’s best to emphasize how integrated present infrastructure is with the internet and how dependent on technology people already are. Point out how easily AI could affect us and the infrastructure we rely on, and how difficult it would be to shut it down once it has infiltrated these systems.
If you still need to give concrete examples, focus on dangers people are already familiar with, such as AI’s potential in creating lab-grown pandemics, the potential for AI to use our weapons of war against us, AI escalating its energy usage to the detriment of our environment- or any others you can think of. Even if a superintelligence would not be limited to destructive means we are already aware of, people are less skeptical of familiar risks.
Principle 4- Accuracy
When communicating a technical topic to a lay audience, it is far more important to get the broad idea across than to be perfectly accurate. For example- if you are arguing with a flat-earther, and they argue that the earth is not round, it is not helpful to retort that the earth is, in fact, an oblate spheroid, and then to give a detailed definition of an oblate spheroid. At this point, you will lose your audience. For the purposes of the argument, it is accurate enough to say the earth is round.
Principle 5- Intelligence Saturation
At some point, AI will be intelligent enough to accomplish most mundane uses, and average people won’t easily perceive improvements in capabilities and intelligence. At this point, many will declare that AI development has hit a wall, even if AI continues to grow capabilities in more specialized domains. What’s worse is that capabilities growth is happening so quickly that people will still widely share “gotcha” examples of older models making mistakes that frontier models no longer make, and this will lull people into a false sense of safety. It is vital to share examples of surprising domain-specific capability growth as often as possible with the public. The more visual and visceral these examples are, the better.
Some will call examples of capabilities growth “hype,” which is why it is also important to continue to share “warning shots,” or visceral examples of AI demonstrating more sophisticated deceptiveness and misalignment.
Principle 6- Shills
Some people accuse anyone arguing AI risk of being shills for big tech. (I’ve yet to see a dime.) This accusation takes one of two forms. 1- You are trying to make AI seem more capable than it is or ever could be by calling it dangerous, thereby gaining investment dollars for large AI companies. 2- You are trying to create a monopoly for the present AI companies by keeping potential competitors from trying to make anything more powerful. These are ad hominem attacks, but they must be addressed.
People who seek power try to get dangerous things before their rivals; it’s best to emphasize that no matter which government or which company has ownership of superintelligence, no one will be able to control it. Everyone is equally in danger from AI, no matter who they are.
Speak about the scientists- not tech CEOs- who are concerned about AI risk. The best example to use is Geoffrey Hinton, who left Google in part to speak out about risk. It’s also helpful to emphasize other dangerous technologies we’ve had to restrict, such as nuclear capabilities.
Principle 7- Agreement > Argument
Reach out to skeptics and meet them where they are. If someone insists that AI will never become superintelligent or general, and that generative AI content is slop, then we’re on the same side. If AI is a hoax, it should be stopped, and if AI is an existential risk, it should be stopped. People often create identities around having a “side” when they agree more with their opponent than they disagree. When these artificial barriers are breached, people can work together for a greater good.
Principle 8- Precautionary Principle
Many people try to lay the burden of proof on AI risk, but the burden of proof should actually lie in the opposite direction. We should not have to prove that AI is dangerous- by the time we have our proof, it is too late. Instead, AI companies should have to prove that their product is safe before foisting it on the public. I find it helpful to use medicine as an analogy; would you take an untested drug simply because it hasn’t killed anyone yet, or would you insist that the drug be thoroughly tested before taking it?
If you find yourself cornered by the burden of proof anyway, we do have many examples of current AI behaving deceptively, reward hacking, attempting to gain power, and attempting to stop itself from being shut off- all behaviors alignment researchers predicted in advance. We may not have definitive proof, but we have plenty of evidence that is sufficient in the face of dangerous uncertainty.
Principle 0 – Kill the Buddha.
If any of these principles don’t work, discard them. If you have any great arguments you can use, use them. If any of the arguments or examples I gave don’t work, throw them away. This is only meant to be a general guide for people having trouble communicating AI risk. Don’t follow this guide at the expense of doing well. In fact, I want you to come up with better arguments and better principles and a better guide as soon as you can. Share that guide with everyone you know.