Maintaining Strategic Human Capacities

There are two dominant perspectives on the problems of AI deskilling and skill supersession: it's a labor problem (individuals won't get jobs) or it's a flourishing problem (individuals won't write poems, talk to friends, etc.).

While both of these are worth considering, they're missing a bigger problem: AI deskilling and skill supersession create strategic vulnerabilities. The inability of humans to participate in key functions in society has ramifications for our survival across a range of risks, major and/or existential: AI failure, AI misalignment, and in AI-human negotiations. For instance, pulling a kill switch on a misaligned AI is politically untenable if society crumbles without AI-assistance.

Our project is on strategic resilience: the ability of a human society to maintain power over critical systems relative to advanced AI. The first step investigates what I call a Strategic Human Capacity Reserve (SHCR). We've articulated a research agenda, which we presented[1] at the TAIGR workshop at ICML. Mentees would engage with us in taking the next steps: determining what threats would require what key human capacities, and thinking about how to design policy interventions to ensure those capacities survive the AI transition.

Within this larger project, we have two subprojects which would benefit from SPAR mentees.

1. Threat modeling. This could include writing scenarios for a given threat (for example, manipulation of a political decision maker's AI guidance/information environment, or emergency which degrades access to autonomous cybersecurity defense). It could also include higher-level comparisons of potential threats, with the goal of classifying and prioritizing them. Mentees might research historical analogies, contemporary forecasting, and related AI incident reporting.

2. Research one potential policy or legal option for preventing human deskilling in a given area. Co-write a paper draft. For example, a philosophical paper in favor of a norm of "AI sabbaths" or a white paper on government investment in firm-level subsidies for redundant cybersecurity talent.

Both of these projects would contribute to a publication which would eventually be submitted to either a journal like AI & Society or a conference like IASEAI.

[1] https://joanobryan.wordpress.com/wp-content/uploads/2026/07/human_capacities_final-.pdf

There are three reasons we may need to rely on strategic human capacities:

1. We can't use AI because it stops doing the thing we want, or because it's destroyed by an adversary.
2. Human expertise is needed to detect certain AI failures.
3. Human expertise is used to generate option value for cases where we could use AI, but for whatever reason choose not to.

Depending on the situation, different human capabilities will matter. Some will be fully general: e.g., can humans get through a truncated period of time without any AI assistance? Others will be specific to places and people: e.g., how does a politician make decisions if we're worried his or her AI assistance has been manipulated?

Given the above, maintaining strategic human capacities could lower risks from transformative AIs in the following ways:

1. It makes it more politically/socially feasible to trigger an AI "kill switch" if necessary.
2. Protected capacities allows for human oversight of critical industries/systems, allowing us to detect AI failures early.
3. It lowers the stakes of an attack on domestic AI systems if society is robust and individuals can go about their lives without AI assistance in an emergency, lessening the chances of great power war.
4. It makes decision-makers more robust to potential manipulation from AI guides, lowering the chance that we will be gradually disempowered in that way.
5. A policy like human-only interpretability research could determine when misaligned AIs are lying to us (low likelihood, but high value).

For exceptional mentees, I am willing to give the level of autonomy I would give to an early PhD student: they can choose a project within the topic, I will give feedback on each stage of the research and paper drafting, they will be first author upon finishing the publication.

In general, mentees can expect to meet with me once a week for 1 hour to discuss a project which we outline together. I will give them feedback on the work that they have done so far, and direct them for that week's work. For instance, I will tell them which areas to read about, how to organize their writing, who to reach out to for interviews, etc.

If the mentee is newer to policy research, I will give them the scaffolding I give one of my undergraduate research assistants: a detailed outline with a list of questions for them to pursue line by line.

Undergraduate level writing experience: papers written without LLM assistance which receive A-level grades in senior-level classes.

EST is best