Comment on “Forecasting is Way Overrated, and We Should Stop Funding It”

By /

Originally posted as a comment on this post. Reposting for visibility and since it is lengthy enough to be a standalone post. I plan to post a more comprehensive update in future describing FRI’s impact and theory of change in more detail.

Summary

[Relevant context/COI: I'm CEO at the Forecasting Research Institute (FRI), an organization which I co-founded with Phil Tetlock and others. Much of the below is my personal perspective, though it is informed by my work. I don't speak for others on my team. I’m sharing an initial reply now, and our team at FRI will share a larger post in future that offers a more comprehensive reflection on these topics.]

Thanks for the post — I think it's important to critically question the value of funds going to forecasting, and this post offers a good opportunity for reflection and discussion.

In brief, I share many of your concerns about forecasting and related research, but I'm also more positive on both its impact so far and its future expected impact.

A summary of some key points:

  1. Much of the impact of forecasting research on specific decision-makers is not public. For example, FRI has informed decisions on frontier AI companies' capability scaling policies, has advised senior US national security decision-makers, and has informed research at key US and UK government agencies. But, we are not able to share many details of this work publicly. However, there is also public evidence that forecasting research is widely cited and informs discourse and some decision-making (some examples below).
  2. AI timelines, adoption, and risk forecasts play a huge role in both individual career decisions and the broader AI discourse. Forecasting research still seems like one of the best tools available for getting specific and accountable beliefs on these topics. For example, comparing 'AI safety' community forecasts to more ‘typical’ experts’ forecasts seems especially important for understanding how much to trust each group’s views. These comparisons will become increasingly relevant for government policymakers over time, especially if there is extremely rapid AI capabilities progress that leads to major societal impacts in the short-run.
  3. When evaluating the impact of FRI-style forecasting research, I think the closest relevant comparison classes are more like broad public goods/measurement-oriented research (e.g., Our World in Data, Epoch) or think-tank research (e.g. GovAI, IAPS). By its nature, the impact of this kind of research tends to be more diffuse and difficult to measure. However, I'd be interested in more intensive comparative evaluation of this type of research and agree that funders should be responsive to evidence about relative impact in these fields.
  4. Forecasting research still has a ton of flaws, and its impact has been far from the dream I've long had for it. There are still big challenges around identifying accurate forecasters on questions related to AI, integrating conditional policy forecasts with actual decision-makers’ needs, and combining deep, individual qualitative research with high-quality, group-generated quantitative forecasts. 
    1. My extremely simplified narrative is: Tetlock et al. established the modern judgmental forecasting field and created a proof of concept for better forecasts on important topics (“superforecasting”)---this work was largely academic; some forecasting platforms were created to build on that work and apply it to a range of important issues; targeted efforts to make forecasting more directly useful to decision-makers are relatively nascent (i.e., have largely begun in the last few years), and are accumulating impact over time, but still have room for improvement.
    2. FRI’s research, in particular, aims to close many of the gaps left by prediction markets and historical forecasting approaches: it is particularly focused on conditional policy forecasts, medium-to-long-run forecasts that do not get much detailed engagement on prediction markets/platforms, and systematically eliciting forecasts from experts who would not typically participate in forecasting platforms but whom decision-makers want to rely on (while also eliciting forecasts from generalists with strong forecasting track records).
  6. However, some factors make the future potential impact of this work look more promising:
    1. AI-enhanced forecasting research is a huge factor that will unlock cheaper, faster, high-quality forecasts on any question of one's choosing. 
    2. The next few years of forecasting AI progress/adoption/impact seem critical, and like they'll deliver a lot of answers on whose forecasts we should trust. It seems good to be ready to support decision-makers during this time.
    3. Leaders in the AI space seem particularly interested in using forecasting in their decision-making; they tend to be both quantitative and open-minded. This creates more potential for forecasting to be useful. More minorly, prediction markets and forecasting are generally becoming more credible within governments. 

More detail on some select points below. This comment already got very long (!), so I’ll reserve more elaboration for a future, more comprehensive post.

Examples of impact

Forecasting research has informed some very important decisions. Unfortunately, many of the details of the relevant evidence here cannot be made public. However, there is evidence of substantial public citation of this research, and some public evidence of affecting particular decisions.

A few examples of relevant impact include:

  • Forecasting has been particularly relevant for decision-making around capability scaling policies. The near-term magnitude of AI-biorisk, how growing AI capabilities may increase it, and what safeguards need to be in place to respond to it, are highly uncertain. Frontier AI companies, the EU AI Code of Practice, and other governments are trying to track and respond to AI impacts on biorisk, cybersecurity, AI R&D, and other domains. We’ve had substantial engagement with the relevant actors, including some focused partnerships, and believe our work in this area has affected important decisions, though we unfortunately cannot share many of the details publicly.
  • Our work on ForecastBench, a benchmark of AI’s ability to do forecasting, showed that AI-produced forecasts could catch up to top human forecasters in roughly the next year if trends persist. This generated interest among senior decision-makers in U.S. national security. We cannot share details, but this is another example of important decision-makers paying attention to and using forecasts.
  • We have completed commissioned research to directly inform grantmaking at Coefficient Giving, and also have indirectly affected grantmaking. For an example of the latter, our work on the Existential Risk Persuasion Tournament (XPT) partially inspired Coefficient Giving (formerly Open Philanthropy) to launch an RFP on improved AI benchmarks. The XPT forecasts predicted that most existing benchmarks would likely saturate in the next few years, and showed that progress on these benchmarks was not crux-y for disagreements about AI impact. We were told that this played a role in the launch and conception of the RFP, and the XPT is cited in the public write-up. 

Some examples of more diffuse impacts — e.g., impact on public understanding of AI and research for policymakers or philanthropists, include:

For context: FRI has been operating for a little over 3 years, and we're accumulating substantially more momentum in terms of connections to top decision-makers as time goes on.

(To be clear: I am mostly discussing FRI here since it’s what I’m most familiar with.)

AI timelines, impact, and adoption forecasts drive a huge amount of career decision-making, attention, etc. 

Forecasts about AI timelines and risk have had major effects on people’s career decisions and the broader AI discourse. AI 2027 underlies popular YouTube videos, 80,000 Hours advises people on career decisions based on timelines forecasts, Dario Amodei’s “country of geniuses in a datacenter by 2027” forecast informs a lot of Anthropic’s work and policy outreach, the AI Impacts survey on AI researchers’ forecasts of existential risk is highly cited, etc.

A major reason I got into this field is that many people are making very intense claims about the effect that AI will have on the world soon, and I want to bring as much rigor and reflection as possible to those claims. So far, it looks like most forecasters are substantially underestimating AI capabilities progress (with some exceptions, e.g. on uplift studies); the evidence on forecasts about AI adoption, societal impacts, and risk is less clear, but I expect we will have more evidence soon, particularly from the Longitudinal Expert AI Panel (LEAP), especially as some forecasters are predicting transformative change in the next few years.

As the expected impact and timing of AI progress is sharpened and clarified, talent and money can be allocated more efficiently.

Case study: Economic impacts of AI

In some cases, it looks to me like forecasting research is picking relatively low-hanging fruit.

The economic impact of AI is a prominent topic of public discussion right now, and it is likely that governments will spend many billions of dollars to address it in the coming years.

Currently, economists hold major sway in public policy about the economic impacts of AI. Perhaps you think top economists, as a group, are badly mistaken about the likely near-term impacts of AI, as some Epoch researchers and others believe. Perhaps you think they are likely to be fairly accurate, as Tyler Cowen, Séb Krier, or typical economists believe. It seems like a valuable common sense intervention to at least document what various groups believe, so that when we are making economic policy going forward we can rely on that evidence to determine who is trustworthy. I believe that studies like this one (and its follow-ups) will be the clearest evidence on the topic.

Relevant comparison class for forecasting research

When thinking about the impact and cost-effectiveness of forecasting, I think it’s more appropriate to compare this work to public goods-oriented research organizations (e.g., Our World in Data, Epoch, etc.) and policy-oriented think-tank research (e.g. GovAI, IAPS, CSET, etc.).

I’ve been disappointed by most impact evaluation of think-tanks and public goods-oriented research that I’ve seen. I believe this is partly because it is very difficult to quantify the impact of this type of work because it has diffuse benefits. But, I still think it’s possible to do better and I would like FRI to do better on this front going forward. 

That said, I still believe there are reasonable heuristics for why this research area could be highly cost-effective. There are many billions of dollars of philanthropic and government capital being spent on AI policy topics. If there is a meaningful indication that forecasting is changing people’s views on these questions (as I believe there is; see discussion above), it seems reasonable to me to spend a very small fraction of that capital on getting more epistemic clarity.

My critiques of forecasting research

Forecasting research, and FRI’s research in particular, still has major areas for improvement.

Examples of a few key issues:

  • I've been underwhelmed by the accuracy of typical experts and superforecasters on questions about AI capabilities progress (as measured by benchmarks); they often underestimate AI progress (with exceptions). I think this underestimation is a useful fact to document, but it would be much more helpful if our research identified experts you should trust. We're in the process of identifying ‘Top AI forecasters' through LEAP and aim to share updates on this soon.
  • I think forecasting research is at its best when combined with in-depth research reports that provide more narratives and key arguments underlying forecasts. For example, Luca Righetti’s work on estimating (certain kinds of) AI-biorisk provides a lot of valuable analysis that usefully complements our expert panel study on the topic. [Note: Luca is an FRI senior advisor and a co-author of our forecasting study.] For decision-makers to build sufficiently detailed models, and for forecasters to test their arguments, we’d ideally have detailed research like Luca’s on most major topics where we collect forecasts — ideally from a few experts who disagree with each other. Unfortunately, this research often doesn’t readily exist, but we are investigating ways to generate it.
  • I have been somewhat surprised by how few experts in AI industry, AI policy, and other domains predict transformative impacts of AI similar to what are commonly discussed by AI lab leaders, people in the AI safety community, and others. This has made it harder to have a true horse-race between the ‘transformative AI’ school of thought that seems to drive a lot of discourse and decision-making vs. more gradual views of AI impacts. Though we have some transformative AI forecasters in our studies, in future work we aim to explicitly collect more forecasts from the ‘transformative AI’ school of thought in order to set up clearer comparisons between worldviews and to better anticipate what will happen if the ‘transformative AI’ school makes more accurate forecasts.

I will save other thoughts on how forecasting, and FRI’s research, could be made more useful to decision-makers for a future post.

But, to be clear: I have a lot of genuine uncertainty about whether forecasting research will be sufficiently impactful going forward. There are promising signs, and increasing momentum, but to more fully deliver on its promise, more improvements will be necessary.

Some notes on FRI-style forecasting research vs. other forecasting interventions

On the value of FRI-style forecasting research in particular:

  • Prediction markets do not have good ways to collect causal policy forecasts, but in our experience, conditional policy forecasts (e.g., how much would various safeguards reduce AI-cyber risk) are often the most helpful forecasts for decision-makers. 
  • Similarly, prediction markets do not create good incentives for longer run forecasts or low-probability forecasts, and incentivize against sharing the rationales behind forecasts. Directly paying and incentivizing relevant experts and forecasters to answer questions is often more useful.
  • Typical forecasting platforms do not get forecasts from the kinds of experts that policymakers typically rely on, and aren't the kind of evidence that can easily be cited in government reports. (This may be unfortunate, but it is the current state of the world.)

Reasons for optimism about future impact

Finally, there are a few factors that have the potential to dramatically change the field going forward:

  1. It looks like AI may soon make it >100x cheaper and faster to get high-quality forecasts on any topic of one’s choosing. Policy researchers will be able to ask the precise question they’re interested in, will be able to upload confidential documents to inform forecasts (something we’ve heard is especially important to decision-makers), and will be able to get detailed explanations for all forecasts. AI-produced forecasts will also be much easier to test for accuracy due to the volume of forecasts they can provide, and it will be easier to generate ‘crux’ questions since AI will not get bored of producing huge numbers of conditional forecasts (which are necessary for identifying cruxes). Building benchmarks and tooling to harness AI-produced forecasts will be a much larger part of our work going forward.
  2. The next few years seem very unusual in human history: very thoughtful researchers are predicting “Superhuman Coders” by 2029, with attendant large impacts. There is a spectrum of views, but the scope for disagreement among reasonable people about what the world will look like in 2030 is huge. This is a particularly important time to make predictions testable, update on what we observe, and make better policy and personal decisions on the basis of this information.
  3. People working in the AI space seem particularly interested in using forecasting, perhaps due to a mix of being quantitatively oriented and because they’re facing unusual degrees of uncertainty. This bodes well for forecasting being useful in the coming years. More minorly, it appears that there is a broader cultural change around forecasting-related topics. Prediction markets are increasingly being cited by government officials, and the public is paying more attention to them than ever before. Much of the impact for prediction markets specifically seems negative (e.g. via incentivizing gambling on low-value topics), but the broader cultural shift suggests there may be an opportunity for better uses of forecasting to enter public consciousness as well.


Discuss

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top