An Epistemic Audit for Existential Risks from AI
This article presents a framework for auditing one's epistemic uncertainty about existential risks from AI, featuring a structured list of questions and domains. The author emphasizes that the framework itself is more valuable than the specific questions, and encourages dynamic updates and community contributions.
x
An Epistemic Audit for Existential Risks from AI — LessWrong
Linkpost for alexandermullerakm.substack.com
3
^
The best audience is likely people with slightly more time available. It’s probably best suited for when you’re entering the field, though I think there is value for anyone, no matter the experience.
^
By epistemic uncertainty I mean the uncertainty about a belief that can, at least in principle, be reduced or falsified with more evidence and study—as opposed to aleatoric uncertainty, the randomness that remains even with perfect knowledge (think of a fair coin flip). Everything in this post concerns the former, so I’ll simply write “uncertainty” from here on.
^
I know this mapping can never be complete. As one example, each of the lethalities in Yudkowsky's list could be phrased as a question here. I appreciate comments discussing domains and subdomains that I missed that meaningfully matter when it comes to existential risks from AI. I will be updating this list dynamically, or at least every three months (when I rerun my own audit of this). If you’re interested in an incredibly long list of questions (more than a 1,000), see this Google Doc which Fable 5 created based on the (paraphrased) prompt where I gave it this post as input and asked to expand on it extremely thoroughly. (I have only skimmed the Doc and haven’t changed a single word, but some new questions seemed interesting).
^
Ordering the domains along a causal chain is expository, not probabilistic. I am not claiming that existential risk is the conjunction of these steps, e.g., domains 7 and 8 are explicitly routes that bypass the chain. Nothing here should be multiplied through. See the multiple-stage fallacy discussion for why that would bias estimates downward.
^
If you have the time to do a deep dive, skip step 2 because it might anchor you unnecessarily. It’s always better to first go through each question, write down what you think, your uncertainty, and in the end aggregate into a summary then do this in reverse. But, knowing that many people will likely not have the time to do this, first doing a summary might already give a quick helpful overview.
^
I’m uncertain about these anchors, and feel that making them more granular and specific would be better. However, I’m unsure what these anchors and specifications would look like, and think the current setup is better than nothing. Curious to receive better alternatives; I’ll make edits.
^
For illustration purposes, I have filled in the domain summary and one question. Importantly, the domain summary doesn’t actually say anything about your object-level beliefs, but purely about your (epistemic) uncertainty with regards to these questions. The other tabs in the Google Sheets allow you to write down object-level beliefs per question and what evidence would significantly change your beliefs.
^
The point of this post is not necessarily the questions (while I do think some of them are quite useful), but more so the framework itself. A lot of these questions have underpinning assumptions that should be the actual questions, and would be better if they are more narrow and falsifiable. However, at least for people still somewhat new to the field, I think the current question should be valuable. One can always swap in other questions and still use this framework if preferred.
^
I know this mapping can never be complete. I appreciate comments discussing domains and subdomains that I missed that meaningfully matter when it comes to existential risks from AI.
3
Click to highlight new comments since: Today at 11:06 AM
Moderation Log
More from Alexander Müller
View more
Curated and popular this week