Anticipatory AI ethics is a framework defended by Seth Lazar in Anticipatory AI Ethics (Knight Columbia, May 2026). It holds that anticipatory ethics applied to AI is necessary under four conditions but must be constrained — to within what Lazar calls the technological horizon — in order to avoid the moral and epistemic objections that often attach to it. Rather than unconditional all-things-considered predictions about AI's future, the framework calls for conditional identification of discrete hazards and opportunities.
Conditions that make anticipatory ethics indicated
Lazar argues that anticipatory ethics is especially indicated when four conditions hold, all of which he contends are met for current AI:
- Rapid technological progress is underway. Lazar grounds this in a sub-argument about the post-2022 transition from "narrow brittle superhuman competence" to "general-purpose models with discrete superhuman capabilities," and on reasoning models substantially mitigating two in-principle obstacles to LLM progress.
- The gap between fundamental research discoveries and society-wide deployment could be small. Lazar argues this is the case, engaging directly with AI as Normal Technology (Narayanan-Kapoor 2025); he cites Microsoft+OpenAI Copilot, Amazon Alexa+ with Anthropic Claude, OpenAI's 400M+ monthly active users, and the "vessels waiting to carry capable models to market" pattern.
- Variance between possible outcomes is high. Lazar holds this is met, pointing to an order-of-magnitude growth-rate increase, radical labor transformation, and shifts in the cyber attack-defense balance, all within the current horizon.
- Levers exist to influence outcomes ex ante. Lazar argues this holds uniquely for AI compared with nanotechnology, human cloning, or medicine, because AI design itself gives ethicists explicit opportunities to shape societal impacts by shaping AI systems.
The technological horizon
The technological horizon is the framework's central construct. It marks the boundary of possible worlds that can be reasonably understood based on two anchors:
- Current AI capabilities, or plausible extensions thereof.
- Current social, political, and economic structures into which AI will be deployed.
Lazar argues both constraints are necessary. Without fixing the deployment environment, uncertainty explodes, because one must forecast AI capabilities and also conjure up a whole ecosystem. Without fixing AI to plausible extensions of current systems, the space of possibilities explodes with few rules to constrain speculation, and there are no levers to intervene in purely hypothetical systems short of shutting down all AI research.
Operating within the horizon allows what Lazar calls constrained analysis: instead of asking "What will the impacts be?", the method identifies 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 science and technology studies (STS) calls affordances — properties that make outcomes more or less likely without necessitating them (Davis 2020).
The constrained-analysis method
Lazar's recommended question is:
"Given our current institutional context, what hazards and opportunities would arise from capabilities that plausible extensions of today's AI systems might realistically acquire?"
The method is conditional. It does not commit to a probability distribution over future capabilities; it identifies how each plausible scenario would interact with the platform-economy context to produce specific hazards. Lazar's illustrative example is that language model agents (LMAs) deployed in a platform economy create strong centralizing tendencies — a feature of the interaction between LMAs and the platform economy — without implying that the world is inevitably headed toward platform agents.
Lazar emphasizes that understanding the technology alone is insufficient. Anticipatory ethics must be paired with anticipatory social science (Nelson & Banks 2018), a reasonable model of the world into which AI is being deployed. Existing regulatory frameworks, political alignments, economic incentives, cultural norms, and practices all contribute to that model.
Objections and Lazar's responses
Lazar groups the standard objections to anticipatory ethics into moral and epistemic categories and responds to each.
Moral objections
- "Amplifies AI hype." Lazar responds that conditional projections of consequences cannot plausibly contribute to AI hype, and that yelling "hype" is not a counterargument to demonstrated progress.
- "Distracts attention from current harms" (the zero-sum-attention objection). Lazar responds that attention is not zero-sum in any structural sense, that allocation depends on political aptitude, and that the claim is unlikely to be resolved by a priori judgment.
- "Making common cause with industry boosters by accepting their capability claims." Lazar responds that accurate assessment of technological capabilities sometimes requires language that would otherwise be hyperbolic, and that reality should not be described inaccurately to avoid appearing hyperbolic.
Epistemic objections
- "Slips into technological determinism." Lazar responds that the horizon-constrained approach is explicitly designed to identify hazards and opportunities (affordances), not all-things-considered predictions, and that anticipatory ethics is precisely aimed at foregrounding human agency by identifying key points for intervention (Johnson 2011).
- "Futurism rarely survives." Lazar responds that conditional, horizon-constrained analysis is not futurism, but the identification of structural pressures and design levers operative now under realistic extensions.
- "Prediction of complex social phenomena is generally bad." Lazar concedes the point and argues it is the reason anticipatory ethics should not engage in unconditional all-things-considered predictions; the horizon-constrained, hazards-and-opportunities approach explicitly avoids them.
Relation to other frameworks and to policy
The pairing with anticipatory social science connects the framework to two other essays from the same Knight Columbia 2026 symposium. Sociotechnical AI Risk Governance (Mulligan-Marda-Wang) and AI as Social Technology (Farrell-Shalizi) both hold that harms emerge from sociotechnical assemblages rather than from model capabilities alone, and both reject the framing that AI capabilities determine outcomes.
The framework sits at the boundary with AI Existential Risk discourse. Lazar's closing question is whether genuinely transformative AI systems fall within or beyond this horizon: the essay defends the horizon-based approach but flags as open the question of whether AGI-level systems, or their proximate extensions, require methods beyond conditional hazard/opportunity identification. Lazar's stated stance is that even granting the existence of AGI-level systems, the horizon-based approach is the right starting point, because it forces engagement with the deployment context and because the levers for intervention are operative now. On this basis the framework critiques x-risk discourse for technological determinism while not rejecting the underlying inquiry, and it defends anticipatory ethics against the "distraction from current harms" critique associated with the "AI as normal technology" position by arguing that anticipatory work is not zero-sum with present-harm work. As a method for treating speculative claims about future AI capabilities, the framework holds that such claims should be conditional and horizon-constrained rather than unconditional.
Relationships
- supports: Anticipatory AI Ethics (Lazar, Knight Columbia, May 1 2026) (canonical source)
- engages-with: AI as Normal Technology (Narayanan-Kapoor 2025) — Lazar argues the methods-applications-diffusion gap is narrower than they allow; not a hard contradiction, but a contested empirical claim
- engages-with: AI Existential Risk — Lazar's framework critiques the determinism without rejecting the inquiry
- complements: Sociotechnical AI Risk Governance (Mulligan-Marda-Wang), AI as Social Technology (Farrell-Shalizi) — same Knight Columbia 2026 symposium; all three share the deployment-environment-matters framing
- related: Affordances (STS concept), AI Power Concentration, Agentic AI, Seth Lazar, Australian National University (ANU), Knight First Amendment Institute (Knight Columbia)
Sources
- Anticipatory AI Ethics (Lazar, Knight Columbia, May 1 2026) — Lazar, Anticipatory AI Ethics, Knight Columbia, May 1 2026 (canonical anchor)
- Johnson (2011) — foundation paper on anticipatory ethics
- Lazar & Nelson (2023) — companion essay on bioethics anticipatory framing
- Lazar et al. (2024); Chan et al. (2025); Kapoor et al. (2025) — Lazar's prior work on platform-power and agentic AI