Beyond Short-Termism: How δ and w Can Realign AI with Our Values
By Beyond Singularity @ 2025-06-18T16:34 (+6)
A sequel to "How the Human Psychological 'Program' Undermines AI Alignment"
TL;DR
Human psychology systematically biases us toward short-term, self-centered decisions, creating a persistent alignment gap between our professed ethics and actual behavior. To counter this, we propose the Time × Scope framework, explicitly managing two ethical parameters—δ (time horizon) and w (moral scope). This structured approach aims to correct our innate biases toward short-termism and narrow moral circles, providing a transparent and adjustable kernel for aligning AI systems with humanity’s aspirational ethics.
Two Types of Intelligent People
When reflecting on human behavior, I observed two distinct types of intelligence in action:
- Strategic Egoists: Smart individuals whose intelligence primarily serves immediate personal gain.
- Strategic Collectivists: Equally intelligent individuals who naturally incorporate broader societal interests and long-term consequences into their decision-making.
This observation led me to hypothesize that what separates these two groups fundamentally involves two dimensions:
- Time Horizon (δ): How far into the future one consistently considers the outcomes.
- Moral Scope (w): How broadly one extends empathy and concern beyond oneself and immediate community.
From this intuitive insight, I derived the minimal framework called Time × Scope.
The Time × Scope utility function
The core ethical utility function is:
- δ (Time Horizon): Reflects how heavily future outcomes are valued (δ < 1 represents natural human discounting; δ ≈ 1 or δ > 1 consciously compensates for present bias). Notably, δ > 1 explicitly prioritizes future welfare above present interests—an essential constitutional decision for overcoming short-termism.
- w (Moral Scope): Represents how broadly empathy and concern extend (self, community, humanity, sentient beings, and potentially future sentient AI).
- v (Value Channels): A vector of measurable ethical facets—e.g., suffering, preference satisfaction, capability, freedom, aesthetics—capturing how each group is affected at each time step. Selecting which channels matter is a normative question (Value Selection); encoding and measuring them is a technical one, discussed later under open challenges.
From Biases to δ and w
To ground these parameters, picture a policy dilemma: should an AI‑governed society construct a hydroelectric dam? We postulate two ethical views — S₁ (short‑term focus, narrow moral circle) and S₂ (long‑term focus, broad moral circle) — and trace how their δ and w coordinates lead to contrasting recommendations.
- Time Horizon (δ): Higher δ values correspond to prioritizing long-term consequences over short-term gratification.
- Moral Scope (w): Higher w values represent broader empathy—consideration of others beyond immediate self-interest.
Thus, any action can be viewed as a point (δ, w). The ethical objective for AI alignment is then to shift this point consistently upward and outward—further into the future and wider in moral consideration, within constitutional guardrails.
Example: Should We Build the Dam?
Below are the contrasting recommendations produced by the two parameter settings.
| Scenario | Parameters | AI Recommendation |
|---|---|---|
| S₁ (Low δ, Narrow w) | δ = 0.9; w_self = 1, w_far = 0.2 | "Build": Immediate cheap energy and short-term employment outweigh long-term ecological costs. |
| S₂ (High δ, Broad w) | δ = 1.02; w_self = 1, w_future = 1.4, w_animals = 1.3 | "Don't build": Long-term ecological damage, wildlife harm, and negative impacts on future generations outweigh short-term economic benefits. |
The differing conclusions clearly illustrate the critical role of explicitly setting δ and w parameters.
Governance: Who Sets δ and w?
Once the parameters prove decisive, the next question is who is allowed to move these knobs, how often, and under what oversight. The table below sketches a three‑layer governance scheme that balances expert deliberation with continuous public feedback while anchoring everything in constitutional guard‑rails.
| Stakeholders | Role | Frequency | Mechanism |
| Expert Assembly | Sets foundational δ₀, w₀ ranges | Annually or semi-annually | Publicly accountable assemblies, incorporating researchers (including AI sentience experts). |
| Human RLHF | Continuous calibration of δᵤ, wᵤ | Ongoing | Sybil-resistant identity verification and hierarchical reputation system. Crucially, δᵤ feedback ensures ongoing societal alignment and prevents rigidity. |
| Constitutional Guardrails | Fixed ethical boundaries and procedural safeguards | Rare updates | Explicit, transparent, and publicly documented guidelines. |
I acknowledge δ and w are not fully orthogonal (future generations inherently expand moral scope). Thus, periodic iterative calibration using sensitivity analysis (e.g., Sobol indices) will be necessary.
Advantages of the Time × Scope Approach
- Transparency: Ethical aims reduced to clear parameters (δ, w) and openly documented v-channels.
- Bias Compensation: Explicitly corrects human short-termism and narrow empathy by design.
- Modularity: Allows updates and refinement of v-vectors without rebuilding the entire alignment structure.
- Computational Feasibility: Real-world implementations use approximations (limited rollouts, value-head approximation) ensuring computational practicality.
Future posts will specifically address how to concretely measure and weigh the v-vector channels (the Value-Loading and the Value-Selection Problems).
Limitations and Open Challenges of the Time × Scope Framework
| Area | Limitation / Open Question | Why it Matters | Possible Mitigation / Research Direction |
|---|---|---|---|
| Value-Loading and Value-Selection (v-vector) | No consensus on how to measure suffering, preference, capability, aesthetics across species and potential sentient AI. | Wrong v corrupts U, regardless of δ,w. | Develop multichannel v (hedonics + preferences + capability) with Bayesian uncertainty; iterative human audits; cross-theory ensembles (moral uncertainty). |
| δ-w Orthogonality | Time horizon and moral scope correlate (future = “others”). High collinearity collapses 2-D space. | Undermines separation of short-term vs collective trade-offs. | Regular Sobol / PCA audits; if overlap > X %, introduce nonlinear link or third axis (e.g., “risk”). |
| Uncaptured Moral Dimensions | Autonomy, honesty, novelty etc. don’t map cleanly onto δ or w. | Optimising δ,w may neglect other virtues. | Add auxiliary constitutional clauses + separate payoff heads for truth, autonomy; multi-objective training. |
| Computational Tractability | Infinite sum over t,i infeasible; approximations can bias. | Hidden bias, Goodhart. | Monte-Carlo rollouts + learned value-head; importance sampling on patient clusters; formal error bounds. |
| Dynamic Context Sensitivity | One global δ,w may mis-compensate by domain or culture. | Over- or under-correction. | Context-conditioned δ,w via causal features; hierarchical tuning (global → domain → local). |
| Governance & Capture | Experts/crowd can be hijacked; RLHF drifts to populism. | δ,w distorted. | DAO with Sybil-resistant ID, rotating expert panels, veto-quorum for minorities, transparent logs. |
| Goodhart & Spec Gaming | Agents manipulate v or feedback while maximising U. | Catastrophic misalignment. | Continuous red-team / causal scrubbing, adversarial training, anomaly triggers for human review. |
| Sentient-AI Agency | Framework treats AI mainly as patients, not agents. | Need reciprocity, accountability. | Extend constitution with agent duties/rights, audit trails, sanction protocols; separate “agent layer.” |
| Normative Debate on δ > 1 | Some ethicists reject negative discounting. | Choice of δ itself contentious. | Deliberative polling; publish confidence intervals on δ; allow opt-in policy slices under different δ settings. |
Incorporating Sentient AI
The framework is intentionally forward-compatible with the potential emergence of sentient artificial agents. This consideration is crucial, as discussed by my colleagues in their recent post, "Will Sentience Make AI’s Morality Better?". Sentient AI welfare can naturally be included in the w-vector.
Time × Scope already treats a future sentient AI as a moral patient through the w parameter. Turning such an AI into a moral agent is handled by the same framework plus the constitutional guard-rails: any autonomous system must optimise U = Σδᵗ wᵢ vᵢ(t) and pass self-critique against the constitution. That said, details of AI accountability and sanction mechanisms are out of scope for this post and will be covered in a follow-up on agent governance.
Conclusion
If my previous post "How the Human Psychological 'Program' Undermines AI Alignment" highlighted why we need a structured compensatory approach, this post outlines how we might begin to implement it. The Time × Scope framework provides a clear, practical starting point for aligning AI systems to our best ethical selves, facilitating both immediate improvement and scalable long-term adaptation.
Open Questions to the Community
To further refine and operationalize the Time × Scope framework, we invite the community to reflect on and discuss the following questions:
- Measurement and Specification: How should we concretely measure and specify the components of the v-vector? Are there robust methods or metrics already available?
- δ Calibration: What criteria should guide the selection of δ? Should δ always be constitutionally set to ≥ 1, or are there legitimate contexts where discounting future outcomes might be ethically justifiable?
- Moral Scope (w): How can we democratically determine the appropriate moral weights (w) for different groups (including sentient AI)? How can non-humans be represented in such a process? What governance structures would best support continuous, inclusive calibration?
- Computational Approximations: What are the most reliable and efficient computational methods for approximating the potentially infinite and complex calculations implied by the Time × Scope framework?
- Ethical Safeguards: What additional ethical guardrails or constitutional principles might be necessary to prevent misuse or unintended ethical drift?
Your critiques, insights, and collaborative input are warmly welcomed.
Ronen Bar @ 2025-06-24T08:13 (+1)
Thanks. I think as time horizon gets bigger, uncertainty may also get bigger, and this should be calculated into the expected value formula.
Beyond Singularity @ 2025-06-24T11:03 (+1)
Great point! Uncertainty really does explode the further we look ahead, and we need to handle that without giving up our long-term responsibilities.
How Time × Scope can deals with it
1. Separate “how much we care” (δ) from “how sure we are” (ρ).
- δ tells the AI how strongly we want to value future outcomes.
- ρ(t) is a confidence term that shrinks as forecasts get fuzzier — e.g. ρ(t)=e^{-kt} or a Bayesian CI straight from the model.
Revised utility: U = Σ [ δ^t × ρ(t) × w_i × v_i(t) ].
2. Set a constitutional horizon (T_max).
- Example rule: “Ignore outcomes beyond 250 years unless ρ(t) > 0.15.”
- Stops the system from chasing ultra-speculative tail events.
3. In practice it’s already truncated.
- Training usually uses finite roll-outs (say 100 years) and adds model-based uncertainty: low-confidence predictions get auto-discounted.
So, δ keeps our ethical aspiration for the long term alive, while ρ(t) and T_max keep us realistic about what we can actually predict. Experts can update the ρ(t) curve as forecasting methods improve, so the system stays both principled and practical.