AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Source Summaries

Anticipatory AI Ethics (Lazar, Knight Columbia, May 1 2026)

high confidence · updated 2026-05-17

Defends anticipatory ethics applied to AI against moral and epistemic objections. Names the 'technological horizon' as the boundary of possible worlds reasonably understandable from current capabilities and institutional frameworks. Argues anticipatory AI ethics is required by four conditions: rapid technological progress, narrow research-to-deployment gap, high outcome variance, and available levers for ex-ante influence.

Contents

Author: Seth Lazar (ANU) Publication: Knight First Amendment Institute (Knight Columbia) / Knight Columbia Date: May 1, 2026 Type: Foundational academic essay introducing a named framework ("technological horizon") and a defense of anticipatory ethics within constrained scope.

Central thesis

Anticipatory ethics applied to AI faces two clusters of objections — moral (hype-amplification, attention-zero-sum, distraction from present harms) and epistemic (technological determinism, futurism-rarely-survives, complex-prediction-is-hard). Lazar defends a constrained, conditional-projection approach grounded in:

  • Epistemic humility — not unconditional all-things-considered predictions, but conditional hazard/opportunity identification
  • A clearly defined technological horizon — the range of possible worlds reasonably understood from current AI capabilities and institutional frameworks

The Technological Horizon

The technological horizon marks the boundary of possible worlds we can reasonably understand based on two constraints held relatively fixed:

  1. Current AI capabilities (or plausible extensions thereof)
  2. Current social, political, and economic structures into which AI will be deployed

Why both constraints matter:

  • Without fixing the deployment environment, uncertainty explodes (forecasting AI + conjuring an ecosystem)
  • Without fixing AI to plausible extensions of current systems, the space of possibilities explodes with no rules to constrain speculation — and we have no levers to intervene in purely hypothetical systems (short of shutting down all AI research)

Operating within the horizon allows constrained analysis: instead of "What will the impacts be?" we identify particular features of AI systems that, given a specific deployment environment, increase the probability of negative or positive outcomes. These discrete hazards and opportunities are what STS calls affordances — properties that make outcomes more or less likely without necessitating them.

Four conditions that make anticipatory ethics indicated

Anticipatory ethics is especially indicated when all four conditions are met:

  1. Rapid technological progress is underway
  2. The gap between fundamental research discoveries and society-wide deployment could be small
  3. Variance between possible outcomes is high
  4. Levers exist to influence outcomes ex ante

Lazar's assessment of current AI:

  • (1) Rapid progress: Clearly met. Cites the post-2022 transition from "narrow brittle superhuman competence" to "general-purpose models with discrete superhuman capabilities."
    • Sub-argument on reasoning models: pre-reasoning autoregressive LLMs had two in-principle upper-bound constraints — inability to differentially allocate compute conditioned on token importance, and path-dependent autoregressive nature. Reasoning models substantially mitigate both by allowing backtracking and proportional compute allocation. Does not imply unbounded returns but does mitigate two in-principle obstacles to fundamental progress.
  • (2) Narrow research-to-deployment gap: Contentious but Lazar argues yes. Engages directly with AI as Normal Technology's Narayanan-Kapoor methods-applications-diffusion distinction:
    • Acknowledges deployment lag exists
    • Counters: every sector now runs on platforms owned by companies financing frontier R&D; new AI models are structurally compatible with these systems; deployment to billions requires only an over-the-air update
    • Cites Microsoft+OpenAI in Copilot, Amazon's Alexa+ using Anthropic Claude, OpenAI's 400M+ MAU
    • "An army of start-ups has built LLM-based software predicated on the underlying models achieving a particular level of performance in the future. When sufficiently capable and efficient models are developed, there will be vessels waiting to carry them directly to market."
    • Likely a "jagged frontier" of deployment lag — some areas slow, some near-instant
  • (3) High variance: Clearly met. Even within the current technological horizon: order-of-magnitude global-growth-rate increase (Erdil-Besiroglu 2023); radical labor transformation (Susskind 2020); cyber attack-defense balance shifts (Guven 2024)
  • (4) Levers exist: Uniquely so for AI. Unlike nanotechnology, human cloning, or medicine, AI design does give ethicists explicit opportunities to shape societal impacts by shaping AI systems themselves. Anticipatory ethics can be both critique (cudgel) and roadmap (blueprint).

The Constrained-Analysis Method

"Given our current institutional context, what hazards and opportunities would arise from capabilities that plausible extensions of today's AI systems might realistically acquire?"

Lazar illustrates with language model agents (LMAs): three plausible futures (capability-reliability gap stall, widespread human-baseline competence, superhuman in every dimension). Anticipatory ethics doesn't bet on a single distribution — it identifies how each scenario would interact with the platform-economy context (large tech firms exercising power over users) to produce specific hazards (e.g., LMAs deployed in the platform economy create strong centralizing tendencies that threaten to further concentrate power). This is one feature, not a summary judgment.

Where the Technological Horizon Ends

Lazar's closing question: Are genuinely transformative AI systems within or beyond the horizon? This essay defends the horizon-based approach but flags the open question of whether AGI-level systems (or their proximate extensions) require methods beyond conditional hazard/opportunity identification.

Analytical context

  • Names the technological horizon as a citable analytical primitive that constrains both AI Existential Risk discourse (often technologically deterministic) and AI as Normal Technology discourse (sometimes accuses anticipatory ethics of distracting from current harms).
  • Pairs structurally with AI as Social Technology (Farrell-Shalizi, also Knight Columbia, May 11 2026) and Sociotechnical AI Risk Governance (Mulligan-Marda-Wang, Knight Columbia, March 16 2026) as part of a coherent Knight Columbia 2026 constrained-deployment-context analysis symposium.
  • Engages explicitly with AI as Normal Technology (Narayanan-Kapoor 2025) on condition 2 (research-to-deployment gap) — Lazar argues the gap is narrower than Narayanan-Kapoor allow.
  • Lazar himself is the load-bearing author of Agent Architecture Patterns-adjacent governance work (cited in the essay: Lazar et al. 2024; Chan et al. 2025; Kapoor et al. 2025).

Relationships

Sources

Categories: Source Summaries | global | opinion | framework | frontier-models | ethics | safety
What links here (8)