In February, the Swift Centre for Applied Forecasting launched a competition designed to bridge the gap between abstract AI safety research and the realities of government decision-making. See the original post here.
Most AI policy work today functions as a literature review of technical risks. While valuable, this rarely moves the dial for a policy official who has 15 minutes to read a brief and 48 hours to make a recommendation. We wanted to test a different model: Forecasting-led, decision-ready policy advice.
The competition was simple:
- Swift Centre provided forecasts across 5 AI related scenarios, from agentic capabilities, to workforce impact and autonomous weapons.
- I provided a 4 page policy template similar to what is often submitted to political leaders and advisors.
- Participants submitted their own policy advice, based on the forecasts, outlining the key options and their recommendation for what a stated political leader should do in response.
The 5 forecasts and 29 policy submissions can be found on an open dashboard.
A team of judges with experience in energy, national security, military, international affairs and AI policy across the UK and US graded the entries. We'll be working with the authors of the highest rated pieces to get their advice directly to relevant decision makers and any potential organisations who may be looking for policy advisors.
A few submissions of note:
If this cause interests you and you'd be keen to see this sort of project be refined and expanded, I'd be happy to speak further.
Reasons we did this
I launched this project to address several systematic gaps (and personal frustrations) I’ve observed in the AI safety policy space.
1. Ineffective Policy
A significant amount of AI safety policy today is essentially a literature review of technical capabilities paired with high-level suggestions that conveniently avoid the trickiest implementation challenges. Though interesting to read for many in the AI safety community, from my experience this type of work has a very low EV if the objective is to inform or persuade policy or decision makers (because they won't read it, or if they do, you aren't helping them solve their practical barriers (i.e. political, institutional, procedural).
2. Not Enough Work Targeting the Highest Value Parts of the Decision-Making Chain
Policy impact is often determined by where you intervene in the decision-making chain. In central governments, the process typically looks like this:
- (a) Identification: An issue is identified in the world.
- (b) Oxygen: The issue is amplified by civil society, research, or industry through long reports, press releases, round tables etc.
- (c) Monitoring: Policy officials in government start paying attention and discussing if to escalate.
- (d) Escalation: Officials write to senior leaders to highlight the issue.
- (e) Steer: Decision-makers give a response (e.g. ignore, monitor, or bring options).
- (f) Advise: Officials write specific options and a recommendation.
- (g) Decision: The leader makes a choice.
- (h) Implementation: Officials draft law, build business cases, or initiate consultations.
Most AI safety work currently focuses on stages (a) and (b). However, I have found it incredibly successful to just go directly to (f) or even (h).
Naturally you need some (a) and (b) to create the conditions for engagement. But by providing fully drafted legislation or policy advice aligned to what they'd get from their officials, you bypass the miscommunication and mis-prioritisation risks that happen when an idea is handed off between stages. This is especially vital today given government departments have increasingly limited capacity and are crisis-driven. If you provide a tangible, finished product, and can get it in front of a decision maker, you remove many of the systematic delays that the institutions create and push the discussion to focus on the tangible choices/actions that you recommend they take.
3. Copying Think Tanks
New AI safety organisations often copy the methods of legacy think tanks whose actual impact is questionable. Frequently, a think tank's influence isn't based on the quality of their reports, but on their proximity to power (e.g. in the UK, the Centre for Policy Studies has written the same type of reports for years, yet surprisingly they are far more influential when their friends in the Conservative party are running the Government).
The good thing is, a lot of these think tanks write bad policy. They are overly complicated, try to do too much, and spread across dozens of pages. So there is a great opportunity to do it better, which the AI safety field would be wise to invest in.
The truth is, policy at its core is a simple, two-element exercise:
- Predicting what the future is likely to be.
- Predicting how your intervention will change the likelihood of that future.
These elements are rarely made explicit in policy reports, leading to massive levels of misinterpretation between the writer and the decision-maker. Make these explicit and you can create a more tangible, action orientated outcome.
4. The "What Next?" Question for Talent
There is a lot of money being spent on training courses and fellowships, but very few outlets for people to get their hands dirty.
Numerous people say they want to get into AI policy and governance but do not have a realistic mechanism to test their fit for such work or before applying to courses or fellowships.
I facilitate a number of BlueDot Impact cohorts, and though I think the courses are great and provide a lot of value, the number one challenge participants have is the "what next?" question. These can include incredibly experienced professionals who's expertise (such as on international relations, diplomacy, defence policy) could be instrumental in solving fundamental challenges when it comes to governing AI.
I am loathed to tell them to take another course or do a fellowship unless I feel it would be especially useful - I want them to take action and start doing something. I often end up suggesting they at least start a substack and produce content on their interest areas: because it's tangible and beneficial to their knowledge/brand and maybe it'll be picked up.
However, I think there could be much better off/on ramps for these people to tangibly deploy their expertise and ideas so that decision makers can benefit. The above competition was one way to test this.
