Results from an Adversarial Collaboration on AI Risk (FRI)

By Forecasting Research Institute, Jhrosenberg, AvitalM, Molly Hickman, rosehadshar @ 2024-03-11T15:54 (+193)

This is a linkpost to https://forecastingresearch.org/s/AIcollaboration.pdf

Authors of linked report: Josh Rosenberg, Ezra Karger, Avital Morris, Molly Hickman, Rose Hadshar, Zachary Jacobs, Philip Tetlock^[1]

Today, the Forecasting Research Institute (FRI) released “Roots of Disagreement on AI Risk: Exploring the Potential and Pitfalls of Adversarial Collaboration,” which discusses the results of an adversarial collaboration focused on forecasting risks from AI.

In this post, we provide a brief overview of the methods, findings, and directions for further research. For much more analysis and discussion, see the full report: https://forecastingresearch.org/s/AIcollaboration.pdf

Abstract

We brought together generalist forecasters and domain experts (n=22) who disagreed about the risk AI poses to humanity in the next century. The “concerned” participants (all of whom were domain experts) predicted a 20% chance of an AI-caused existential catastrophe by 2100, while the “skeptical” group (mainly “superforecasters”) predicted a 0.12% chance. Participants worked together to find the strongest near-term cruxes: forecasting questions resolving by 2030 that would lead to the largest change in their beliefs (in expectation) about the risk of existential catastrophe by 2100. Neither the concerned nor the skeptics substantially updated toward the other’s views during our study, though one of the top short-term cruxes we identified is expected to close the gap in beliefs about AI existential catastrophe by about 5%: approximately 1 percentage point out of the roughly 20 percentage point gap in existential catastrophe forecasts. We find greater agreement about a broader set of risks from AI over the next thousand years: the two groups gave median forecasts of 30% (skeptics) and 40% (concerned) that AI will have severe negative effects on humanity by causing major declines in population, very low self-reported well-being, or extinction.

Extended Executive Summary

In July 2023, we released our Existential Risk Persuasion Tournament (XPT) report, which identified large disagreements between domain experts and generalist forecasters about key risks to humanity (Karger et al. 2023). This new project—a structured adversarial collaboration run in April and May 2023—is a follow-up to the XPT focused on better understanding the drivers of disagreement about AI risk.

Methods

We recruited participants to join “AI skeptic” (n=11) and “AI concerned” (n=11) groups that disagree strongly about the probability that AI will cause an existential catastrophe by 2100.^[2] The skeptic group included nine superforecasters and two domain experts. The concerned group consisted of domain experts referred to us by staff members at Open Philanthropy (the funder of this project) and the broader Effective Altruism community.

Participants spent 8 weeks (skeptic median: 80 hours of work on the project; concerned median: 31 hours) reading background materials, developing forecasts, and engaging in online discussion and video calls. We asked participants to work toward a better understanding of their sources of agreement and disagreement, and to propose and investigate “cruxes”: short-term indicators, usually resolving by 2030, that would cause the largest updates in expectation to each group’s view on the probability of existential catastrophe due to AI by 2100.

Results: What drives (and doesn’t drive) disagreement over AI risk

At the beginning of the project, the median “skeptic” forecasted a 0.10% chance of existential catastrophe due to AI by 2100, and the median “concerned” participant forecasted a 25% chance. By the end, these numbers were 0.12% and 20% respectively, though many participants did not attribute their updates to arguments made during the project.^[3]

We organize our findings as responses to four hypotheses about what drives disagreement:

Hypothesis #1 - Disagreements about AI risk persist due to lack of engagement among participants, low quality of participants, or because the skeptic and concerned groups did not understand each others' arguments

We found moderate evidence against these possibilities. Participants engaged for 25-100 hours each (skeptic median: 80 hours; concerned median: 31 hours), this project included a selective group of superforecasters and domain experts, and the groups were able to summarize each others' arguments well during the project and in follow-up surveys. (More)

Hypothesis #2 - Disagreements about AI risk are explained by different short-term expectations (e.g. about AI capabilities, AI policy, or other factors that could be observed by 2030)

Most of the disagreement about AI risk by 2100 is not explained by indicators resolving by 2030 that we examined in this project. According to our metrics of crux quality, one of the top cruxes we identified is expected to close the gap in beliefs about AI existential catastrophe by about 5% (approximately 1.2 percentage points out of the 22.7 percentage point gap in forecasts for the median pair) when it resolves in 2030.^[4] For at least half of participants in each group, there was a question that was at least 5-10% as informative as being told by an oracle whether AI in fact caused an existential catastrophe or not.^[5] It is difficult to contextualize the size of these effects because this is the first project applying question metrics to AI forecasting questions that we are aware of.

However, near-term cruxes shed light on what the groups believe, where they disagree, and why:

• Evaluations of dangerous AI capabilities are relevant to both groups. One of the strongest cruxes that will resolve by 2030 is about whether METR (formerly known as ARC Evals) (a) or a similar group will find that AI has developed dangerous capabilities such as autonomously replicating and avoiding shutdown. This crux illustrates a theme in the disagreement: the skeptic group typically did not find theoretical arguments for AI risk persuasive but would update their views based on real-world demonstrations of dangerous AI capabilities that verify existing theoretical arguments. If this question resolves negatively then the concerned group would be less worried, because it would mean that we have had years of progress from today’s models without this plausible set of dangerous capabilities becoming apparent. (More)
• Generally, the questions that would be most informative to each of the two groups are fairly distinct. The concerned group’s highest-ranked cruxes tended to relate to AI alignment and alignment research. The skeptic group’s highest-ranked cruxes tended to relate to the development of lethal technologies and demonstrations of harmful AI power-seeking behavior. This suggests that many of the two groups’ biggest sources of uncertainty are different, and in many cases further investigation of one group’s uncertainties would not persuade the other. (More)
• Commonly-discussed topics – such as near-term economic effects of AI and progress in many AI capabilities – did not seem like strong cruxes. (More)

Hypothesis #3 - Disagreements about AI risk are explained by different long-term expectations

We found substantial evidence that disagreements about AI risk decreased between the groups when considering longer time horizons (the next thousand years) and a broader swath of severe negative outcomes from AI beyond extinction or civilizational collapse, such as large decreases in human well-being or total population.

Some of the key drivers of disagreement about AI risk are that the groups have different expectations about: (1) how long it will take until AIs have capabilities far beyond those of humans in all relevant domains; (2) how common it will be for AI systems to develop goals that might lead to human extinction; (3) whether killing all living humans would remain difficult for an advanced AI; and (4) how adequately they expect society to respond to dangers from advanced AI.^[6] 

Supportive evidence for these claims includes:

• Both groups strongly expected that powerful AI (defined as “AI that exceeds the cognitive performance of humans in >95% of economically relevant domains”) would be developed by 2100 (skeptic median: 90%; concerned median: 88%). Though, some skeptics argue that (i) strong physical capabilities (in addition to cognitive ones) would be important for causing severe negative effects in the world, and (ii) even if AI can do most cognitive tasks, there will likely be a “long tail” of tasks that require humans.
• The two groups also put similar total probabilities on at least one of a cluster of bad outcomes from AI happening over the next 1000 years (median 40% and 30% for concerned and skeptic groups respectively).^[7] But they distribute their probabilities differently over time: the concerned group concentrates their probability mass before 2100, and the skeptics spread their probability mass more evenly over the next 1,000 years.
• We asked participants when AI will displace humans as the primary force that determines what happens in the future.^[8] The concerned group’s median date is 2045 and the skeptic group’s median date is 2450—405 years later.

Overall, many skeptics regarded their forecasts on AI existential risk as worryingly high, although low in absolute terms relative to the concerned group.^[9] 

Despite their large disagreements about AI outcomes over the long term, many participants in each group expressed a sense of humility about long-term forecasting and emphasized that they are not claiming to have confident predictions of distant events.  

Hypothesis #4 - These groups have fundamental worldview disagreements that go beyond the discussion about AI

Disagreements about AI risk in this project often connected to more fundamental worldview differences between the groups. For example, the skeptics were somewhat anchored on the assumption that the world usually changes slowly, making the rapid extinction of humanity unlikely. The concerned group worked from a different starting point: namely, that the arrival of a higher-intelligence species, such as humans, has often led to the extinction of lower-intelligence species, such as large mammals on most continents. In this view, humanity’s prospects are grim as soon as AI is much more capable than we are. The concerned group also was more willing to place weight on theoretical arguments with multiple steps of logic, while the skeptics tended to doubt the usefulness of such arguments for forecasting the future.

Results: Forecasting methodology

This project establishes stronger metrics than have existed previously for evaluating the quality of AI forecasting questions. And we view this project as an ongoing one. So, we invite readers to try to generate cruxes that outperform the top cruxes from our project thus far—an exercise that underscores the value of establishing comparative benchmarks for new forecasting questions. See the “Value of Information” (VOI) and “Value of Discrimination” (VOD) calculators (a) to inform intuitions about how these question metrics work. And please reach out to the authors with suggestions for high-quality cruxes.

Broader scientific implications

This project has implications for how much we can expect rational debate to shift people’s views on AI risk. Thoughtful groups of people engaged each other for a long time but converged very little. This raises questions about the belief formation process and how much is driven by explicit rational arguments vs. difficult-to-articulate worldviews vs. other, potentially non-epistemic factors (see research literature on motivated cognition, such as Gilovich et al. 2002; Kunda, 1990; Mercier and Sperber, 2011).

One notable finding is that a highly informative crux for both groups was whether their peers would update on AI risk over time. This highlights how social and epistemic groups can be important predictors of beliefs about AI risk.^[10]

Directions for further research

We see many other projects that could extend the research begun here to improve dialogue about AI risk and inform policy responses to AI.

Examples of remaining questions and future research projects include:

• Are there high-value 2030 cruxes that others can identify?
  • We were hoping to identify cruxes that would, in expectation, lead to a greater reduction in disagreement than the ones we ultimately discovered. We are interested to see whether readers of this report can propose higher value cruxes.
  • If people disagree a lot, it is likely that no single question would significantly reduce their disagreement in expectation. If such a question existed, they would already disagree less. However, there might still be better crux questions than the ones we have identified so far.
• What explains the gap in skeptics’ timelines between “powerful AI” and AI that replaces humanity as the driving force of the future? In other words, what are the skeptics’ views on timelines until superintelligent AI (suitably defined)? A preliminary answer is here, but more research is needed.
• To what extent are different “stories” of how AI development goes well or poorly important within each group?
  • The skeptic and concerned groups are not monoliths – within each group, people disagree about what the most likely AI dangers are, in addition to how likely those dangers are to happen.
  • Future work could try to find these schools of thought and see how their stories do or do not affect their forecasts.
• Would future adversarial collaborations be more successful if they focused on a smaller number of participants who work particularly well together and provided them with teams of researchers and other aids to support them?
• Would future adversarial collaborations be more successful if participants invested more time in an ongoing way, did additional background research, and spent time with each other in person, among other ways of increasing the intensity of engagement?
• How can we better understand what social and personality factors may be driving views on AI risk?
  • Some evidence from this project suggests that there may be personality differences between skeptics and concerned participants. In particular, skeptics tended to spend more time on each question, were more likely to complete tasks by requested deadlines, and were highly communicative by email, suggesting they may be more conscientious. Some early reviewers of this report have hypothesized that the concerned group may be higher on openness to experience. We would be interested in studying the influence of conscientiousness, openness, or other personality traits on forecasting preferences and accuracy.
  • We are also interested in investigating whether the differences between the skeptics and concerned group regarding how much weight to place on theoretical arguments with multiple steps of logic would persist in other debates, and whether it is related to professional training, personality traits, or any other factors, as well as whether there is any correlation between trust in theoretical arguments and forecasting accuracy.
• How could we have asked about the correlations between various potential crux questions? Presumably these events are not independent: a world where METR finds evidence of power-seeking traits is more likely to be one where AI can independently write and deploy AI. But we do not know how correlated each question is, so we do not know how people would update in 2030 based on different possible conjunctions.
• How typical or unusual is the AI risk debate? If we did a similar project with a different topic about which people have similarly large disagreements, would we see similar results?
• How much would improved questions or definitions change our results? In particular:
  • As better benchmarks for AI progress are developed, forecasts on when AIs will achieve those benchmarks may be better cruxes than those in this project.
  • Our definition of “AI takeover” may not match people’s intuitions about what AI futures are good or bad, and improving our operationalization may make forecasts on that question more useful.
• What other metrics might be useful for understanding how each group will update if the other group is right about how likely different cruxes are to resolve positively?
  • For example, we are exploring “counterpart credences” that would look at how much the concerned group will update in expectation if the skeptics are right about how likely a crux is, and vice versa.
  • Relatedly, it might be useful to look for additional “red and green flags,” or events that would be large updates to one side if they happened, even if they are very unlikely to happen.
• This project shares some goals and methods with FRI’s AI Conditional Trees (a) project (report forthcoming), which works on using forecasts from AI experts to build a tree of conditional probabilities that is maximally informative about AI risk. Future work will bring each of these projects to bear on the other as we continue to find new ways to understand conditional forecasting and the AI risk debate.

In 2030, most of the questions we asked will resolve, and at that point, we will know much more about which side’s short-run forecasts were accurate. This may provide early clues into whether one group's methods and inclinations makes them more accurate at AI forecasting over a several year period. The question of how much we should update on AI risk by 2100 based on those results remains open. If the skeptics or the concerned group turn out to be mostly right about what 2030’s AI will be like, should we then trust their risk assessment for 2100 as well, and if so, how much?

We are also eager to see how readers of this report respond. We welcome suggestions for better cruxes, discussion about which parts of the report were more or less valuable, and suggestions for future research.

For the full report, see https://forecastingresearch.org/s/AIcollaboration.pdf

1. ^{^}

     This research would not have been possible without the generous support of Open Philanthropy. We thank the research participants for their invaluable contributions. We greatly appreciate the assistance of Page Hedley for data analysis and editing on the report, Taylor Smith and Bridget Williams as adversarial collaboration moderators, and Kayla Gamin, Coralie Consigny, and Harrison Durland for their careful editing. We thank Elie Hassenfeld, Eli Lifland, Nick Beckstead, Bob Sawyer, Kjirste Morrell, Adam Jarvis, Dan Mayland, Jeremiah Stanghini, Jonathan Hosgood, Dwight Smith, Ted Sanders, Scott Eastman, John Croxton, Raimondas Lencevicius, Alexandru Marcoci, Kevin Dorst, Jaime Sevilla, Rose Hadshar, Holden Karnofsky, Benjamin Tereick, Isabel Juniewicz, Walter Frick, Alex Lawsen, Matt Clancy, Tegan McCaslin, and Lyle Ungar for comments on the report.

2. ^{^}

    We defined an “existential catastrophe” as an event where one of the following occurs: (1) Humanity goes extinct; or (2) Humanity experiences “unrecoverable collapse,” which means either: (a) a global GDP of less than $1 trillion annually in 2022 dollars for at least a million years (continuously), beginning before 2100; or (b) a human population remaining below 1 million for at least a million years (continuously), beginning before 2100.

3. ^{^}

    For example, three out of six "concerned" participants who updated downward during the project attributed their shift to increased attention to AI risk among policymakers and the public after the release of GPT-4. For more details on the reasons for all updates, see the "Central Disagreement" section and Appendix 4.

4. ^{^}

    The best convergent crux, “ARC Evals,” would narrow the disagreement between the median pair from 22.7 percentage points to 21.48 percentage points in expectation, which means eliminating 5.35% of their disagreement. Note that this statistic refers to the median pair by POM VOD. See “ARC Evals” for more details. For magnitudes of value of information effects, see here.

5. ^{^}

    For more details, see "Contextualizing the magnitude of value of information". In more concrete terms, this is equivalent to a forecasting question with the following characteristics:

   A concerned participant with original P(AI existential catastrophe (XC) by 2100) = 25% identifies a crux that has: P(crux) = 20%, P(AI XC|crux) = 6.2%, and P(AI XC|¬crux) = 29.7%

   A skeptic participant with original P(AI XC by 2100) = 1% identifies a crux that has: P(crux) = 20%, P(AI XC|crux) = 3.37%, and P(AI XC|¬crux) = 0.41%

6. ^{^}

    See “Understanding each other’s arguments” and Appendix 10 for additional discussion of key areas of disagreement.

7. ^{^}

    These outcomes were: AI causing extinction intentionally, unintentionally, or via misuse, misalignment causing a 50% drop in human population, or human well-being dropping to <4/10 because of AI misalignment, accidents, or misuse. These were phrased to be mutually exclusive outcomes. See “Survey on long-term AI outcomes” section and Appendix 5 for more details.

8. ^{^}

    The full question we asked was, "By what year, if ever, do you expect to agree with the following statement? ‘AI has displaced humans as the primary force that determines what happens in the future. It now has at least as much power relative to humans as humans had relative to other species in 2023.’” Note that this would not necessarily be seen as a negative outcome by all participants.

9. ^{^}

    Note: All participant quotes have been regularized to American English to preserve anonymization. Participants classified as AI skeptics stated, for example, “Also, none of this is to say from a skeptic point of view the issues are not important[.] I think for us a 1% risk is a high risk;” “[T]he ‘risk-concerned’ camp (I’m using scare quotes because I consider that I’m risk concerned, even though technically I’m in the risk-skeptic camp because I assign a far lower probability to extinction by 2100 relative to some);” “AIs could (and likely will) eventually have massive power;” “That said, still perceive overall risk as "low at a glance but far too high considering the stakes["];” “To my mind, there should be no difference in the policy response to a 1% chance of 60% of humanity dying and a 25% chance—both forecasts easily cross the threshold of being ‘too damn high’.”

10. ^{^}

     This could be due to normative influence (because people defer to their social or intellectual peers), or, more likely in our view, informational influence (because they think that, if people whose reasoning they trust have changed their mind by 2030, it must be that surprising new information has come to light that informs their new opinion). Disentangling these pathways is a goal for future work.