'Future bias' and a possible solution to moral fanaticism | Christian Tarsney (2021)

Core Research Focus

• Christian Tarsney is a philosopher at the Global Priorities Institute (GPI) at Oxford University, researching how altruistically motivated agents should prioritize resources.
• His primary research interests include:
  • Epistemic challenges regarding predicting and influencing the far future (centuries/millennia).
  • Decision theory in contexts of extreme probabilities, specifically "fanaticism" and "Pascal's mugging."
  • Moral psychology, specifically "future bias" (temporal value asymmetry).

Future Bias and the Philosophy of Time

• Definition: Future bias is the intuition that people prefer negative experiences to occur in the past rather than the future, and positive experiences in the future rather than the past.
• Parfit's Thought Experiment: In "My Past and Future Operation," most people prefer a longer, more painful surgery in the past over a shorter, less painful surgery in the future, despite the total pain being greater in the past option.
• Experimental Findings: Tarsney and collaborators found that the "unaffectability" of the past explains part of future bias; when hypothetical scenarios allow retrocausal influence, people care more about past experiences, though a residual bias remains.
• Philosophical Explanations:
  • One theory posits that humans believe we "move" through time, making the future ethically distinct from the past.
  • Ontological theories of time relevant to this debate include:
    • Presentism: Only the present moment is real.
    • Growing Block Theory: The past and present are real; the future is not.
    • Eternalism: Past, present, and future are equally real (a "block universe").
• Rationality Debate:
  • Money Pump Argument: If future bias is merely a taste, it may lead to irrational cycles of choice where an agent is guaranteed to be worse off (e.g., accepting a trade to avoid future pain that they previously accepted to avoid past pain).
  • Time Travel Implications: If backward time travel were possible (in a self-consistent loop), the asymmetry of bias weakens, suggesting the bias is linked to the inability to affect the past.

The Problem of Fanaticism and Decision Theory

• Fanaticism Defined: A tendency of expected value maximization to prioritize options with tiny probabilities of astronomically good/bad outcomes over options with certain, high-value outcomes (e.g., preferring a 1 in $10^{100}$ chance of infinite happiness over 1 trillion certain lives).
• Objections to Fanaticism:
  • Standard Expected Utility Theory (EUT) does not force linearity; value functions can be concave (diminishing returns), preventing extreme outcomes from dominating decisions.
  • Arguments for fanaticism (e.g., Harsanyi's aggregation theorem) rely on controversial premises like ex ante Pareto principles.
• Tarsney's Proposed Solution (Stochastic Dominance with Background Uncertainty):
  • Stochastic Dominance: A rational agent should not choose an option that is strictly dominated by another across all probability distributions (i.e., Option A has a higher probability of being better than Option B for every possible outcome threshold).
  • Background Uncertainty: When accounting for uncertainty about the total value of the universe (background state), many moderate options become stochastically dominant over their riskier counterparts.
  • The Threshold Effect: This framework requires expected value maximization for outcomes smaller than the scale of background uncertainty (e.g., saving 100 lives with 99% probability vs. 1 life for sure).
  • The Escape Hatch: For outcomes vastly larger than background uncertainty (e.g., universe-scale consequences), the requirement to maximize expected value weakens, allowing agents to choose the "sure thing" without being irrational.
• Implications for Long-Termism:
  • This model supports prioritizing existential risk reduction where the probability of impact is not infinitesimally small (e.g., >1 in a billion), as these options become stochastically dominant.
  • It suggests that extreme "fanatical" choices driven by near-zero probabilities of infinite value are not rationally required.

Epistemic Challenges to Long-Termism

• The Core Tension: Long-termism relies on the vast scale of the future, but the ability to predictably influence the far future decays over time due to exogenous events (e.g., extinction, civilizational collapse).
• Modeling Decay: Tarsney models the persistence of interventions as an exponential decay rate (e.g., a 1% chance per century that an intervention's effect is washed out by a new catastrophe).
• Robustness of Existential Risk Reduction:
  • Even under pessimistic assumptions (e.g., a constant 1% annual extinction risk, no interstellar expansion), reducing existential risk retains significant expected value compared to short-term interventions.
  • Long-termism is robust if one assigns non-trivial credence (e.g., >1 in 1,000) to scenarios where human civilization becomes stable and expands, as these scenarios generate the bulk of the expected value.
  • Overconfidence in the "mediocrity" of the future (e.g., assuming we will never leave the solar system or achieve high welfare) is required to reject long-termism.

Moral Uncertainty and Rationality

• Internalism vs. Externalism:
  • Externalism: Moral principles have authority regardless of belief; one should do what the true moral theory dictates.
  • Internalism: One should act on one's best current normative beliefs; rationality depends on the agent's credences.
• The Regress Problem: Extreme internalism requires a higher-order principle to resolve moral uncertainty, which itself requires a meta-principle, leading to an infinite regress.
• Tarsney's Moderate Externalism:
  • Rationality is determined by an external criterion (e.g., stochastic dominance).
  • However, that criterion incorporates the agent's internal beliefs about empirical states and moral values.
  • Agents should hedge against moral uncertainty (like empirical uncertainty) using decision rules like variance normalization or expected choice worthiness.
• Practical Application: While theoretically developed (e.g., comparing value scales across theories), applying these tools in practice remains difficult due to the complexity of defining value functions for all conceivable outcomes.

Scope and Demandingness of Long-Termism

• Radical vs. Subtle Long-Termism:
  • Radical: Demands maximal resource allocation to the far future, potentially ignoring the present (e.g., "factories for factories").
  • Subtle: Focuses on "subtle" long-termism, where improving present institutions, social trust, and fairness is the best way to equip future generations to handle unknown challenges.
• Resource Allocation:
  • There is likely a point of diminishing returns where spending more on the far future yields less impact than addressing urgent near-term issues.
  • Many long-termist priorities (nuclear risk, pandemic prep, AI safety) are also high priorities for near-termist perspectives.
• Future Civilization Value:
  • Extrapolating current trends (e.g., violence reduction vs. factory farming suffering) is risky for predicting the far future.
  • A "outside view" suggests that intelligent agents may have an asymmetric tendency to pursue the good (via moral progress), though this is contested against evolutionary arguments.

Field Updates and Career Advice

• GPI Growth: The field of Global Priorities Research is expanding rapidly, with a strong pipeline of PhD students in philosophy and economics.
• Hiring Needs: The field currently prioritizes hiring for philosophy and economics roles, specifically for recent PhDs and pre-doctoral researchers.
• Future Priorities:
  • Tarsney would prioritize existential risk reduction, institutional improvement, and research into norms/values.
  • He is optimistic about future progress on moral progress, inductive reasoning, and the nature of consciousness.
• Competitive Debating:
  • Tarsney notes that competitive debate trains speed and argumentation but can encourage "game-playing" where bad arguments are deployed quickly to waste opponent time.
  • Skills transfer to academia only if the debater recognizes the distinction between "winning a debate" and "seeking the truth."

Fun Philosophy

• Barry's Paradox: A self-reference paradox involving the smallest natural number not nameable by an English expression of fewer than 100 characters. The phrase used to describe the number is itself fewer than 100 characters, creating a contradiction.