📈 economics

Technology Speed Limits

This paper demonstrates that an adaptive speed limit, which caps the rate of technological increase, is the unique time-consistent mechanism capable of delivering optimal worst-case guarantees for technology regulation under private learning through both scaling and waiting.

Original authors: Andrew Koh, Sivakorn Sanguanmoo

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew Koh, Sivakorn Sanguanmoo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the mayor of a city, and a brilliant but reckless tech company wants to build a new kind of flying car. You know this invention could change the world for the better, but you also know it could crash and kill everyone if it's not ready.

The problem is that you don't know when it will be safe. You only learn two things:

By Doing: As they build more cars and fly them, you learn how they handle. But if you let them build too many too fast, and one crashes, the damage is done. You can't "un-fly" the car.
By Waiting: As time passes, even if they build nothing new, you might learn from accidents elsewhere or from new scientific discoveries that the cars are dangerous.

The paper by Andrew Koh and Sivan Sanguanmoo asks: What is the best rule for the mayor to give the company to ensure safety without stifling progress, especially when the company might care more about profit than public safety?

The "Speed Limit" Solution

The authors argue that the best rule is an Adaptive Speed Limit.

Think of it like a construction site with a "Maximum Daily Progress" sign.

The Rule: The company is allowed to increase the number of flying cars by a certain amount each day (e.g., "No more than 10 new cars per day").
The Adaptation: This limit isn't fixed forever. If the company builds 10 cars and they fly perfectly, the mayor might say, "Okay, tomorrow you can build 15." But if a car crashes or a new danger is discovered, the mayor can instantly lower the limit to 0.

Why is this the "Perfect" Rule?

The paper proves two main things using complex math (which we'll translate into metaphors):

1. It's the "Safety Net" for the Worst-Case Scenario (Robustness)
Imagine you are playing a game where you don't know the rules of the opponent. Maybe the company is super smart and learns fast; maybe they are greedy and ignore risks. Maybe they learn by testing; maybe they learn by reading reports.

The authors show that the Adaptive Speed Limit is the only rule that guarantees the best possible outcome even in the worst-case scenario.

The Metaphor: Imagine you are walking through a minefield. You don't know where the mines are. The "Speed Limit" is like a rule that says, "Take one step every 10 seconds."
- If you walk too fast (no limit), you might step on a mine immediately.
- If you stand still forever (total ban), you never reach the treasure.
- The speed limit ensures that no matter where the mines are hidden or how fast you could learn, you will never step on two mines at once. It gives you the "option value" to stop if you hear a click.

2. It's the Only Rule the Mayor Won't Cheat On (Time-Consistency)
This is the trickiest part. Imagine the mayor promises the company: "I will never let you build more than 10 cars a day."

The Problem: Tomorrow, the company builds 10 cars and they are amazing! The mayor gets excited and thinks, "I should let them build 20 tomorrow to get the benefits faster!" But if the company knows the mayor might change their mind, they might take huge risks today, betting the mayor will loosen the rules later.
The Solution: The authors prove that the Adaptive Speed Limit is Time-Consistent. This means the mayor has no incentive to change the rule later.
- The Metaphor: It's like a parent telling a child, "You can eat one cookie an hour." If the child eats one, the parent still wants to wait an hour before the next one, because they know that if they let the child eat two now, the child might get a stomach ache later. The rule remains the best rule at every single moment, so the parent never needs to lie or break their promise.

What About the "Bad Guys"?

The paper assumes the tech company (the "Agent") might be misaligned with society (the "Principal").

They might not care about the crashes (negative externalities).
They might want to get rich super fast (disproportionate winners).

The Speed Limit works even if the company is trying to game the system. Because the limit is a "hard cap" (you literally cannot go faster than the limit), the company can't trick the system. They can choose to go slower than the limit if they want to be extra careful, but they can never go faster.

The "Buying Time" Concept

The paper also mentions that slowing down "buys time."

The Metaphor: Think of the flying car as a fire. If you pour water on it too fast, you might make it spread. If you pour it slowly, you give yourself time to build a fireproof wall (mitigation) or figure out a better way to put it out.
The Speed Limit ensures that the technology doesn't grow so fast that we run out of time to build the safety measures (like new laws, vaccines, or AI safety research) needed to handle it.

Summary in One Sentence

The paper proves that the smartest, most reliable way to regulate dangerous new technology is to set a moving speed limit that can be tightened or loosened based on new information, because it is the only rule that protects society from the worst risks while still allowing progress, and it's the only rule the regulator will actually stick to.

Technical Summary: Technology Speed Limits

Problem Statement
The paper addresses the optimal regulation of new technologies where learning occurs through two distinct, interconnected margins: learning-by-doing (gaining information by scaling up the technology) and learning-by-waiting (gaining information simply as time passes). The central tension arises because learning-by-doing is risky due to the irreversibility of technology deployment, while learning-by-waiting is costly due to the delay of potential gains.

The authors model a scenario where a principal (representing society) seeks to regulate an agent (e.g., a firm) whose incentives are misaligned with social welfare. This misalignment may stem from negative externalities (where the firm underestimates harm in a "bad" state) or disproportionate gains (where the firm captures excessive benefits in a "good" state). The core challenge is designing a policy mechanism that remains optimal regardless of the specific learning process the agent employs or their specific risk preferences, and which remains credible over time without requiring the principal to commit to future actions.

Methodology
The authors develop a stochastic control model on a discrete time and technology level space ( $T \times L$ ).

Learning Framework: Unlike standard multi-armed bandit models, the filtration of information is modeled as a multiparameter field. The authors explicitly avoid the "commutation" condition common in literature, allowing for dependence between learning-by-doing and learning-by-waiting. This captures the reality that new information might retroactively alter the interpretation of signals received at lower technology levels.
Mechanism Design: The principal offers a path-adapted mechanism consisting of flow transfers (taxes or subsidies) based on the realized technology path. The agent chooses a non-decreasing path of technology levels to maximize their discounted expected utility net of transfers.
Robustness Criteria: The authors define two robustness problems:
1. Learning-robustness: Optimality against any admissible learning filtration ( $F$ ).
2. Dually-robustness: Optimality against any admissible learning filtration ( $F$ ) and any admissible agent preference structure ( $U$ ).
Time-Consistency: A mechanism is time-consistent if, after any decision history, its continuation is conditionally undominated. This rules out mechanisms where the principal would have an incentive to deviate from the policy once new information arrives or the technology level changes.

Key Results
The paper establishes two primary theorems regarding Adaptive Speed Limits, defined as a cap on the rate at which the technology level can increase per unit time, where the cap itself evolves based on the information revealed by the principal's filtration.

Theorem 1 (Robust Optimality): The adaptive speed limit, derived from the solution to the principal's "direct control problem" (where the principal chooses the path directly assuming full information), is both learning-robust and dually-robust.
- Intuition: The proof relies on comparing the option value of slowing down. If an agent chooses a path slower than the speed limit, it is often a precautionary move to preserve the option to halt if adverse information arrives later. The authors show that for any preference structure, the principal's payoff under the agent's optimal path (constrained by the speed limit) is weakly higher than the payoff under the principal's direct control. Since the direct control value serves as an upper bound for any mechanism (even with commitment), the speed limit achieves this bound robustly.
Theorem 2 (Uniqueness and Time-Consistency): Among all dually-robust mechanisms, the adaptive speed limit is the unique time-consistent mechanism (assuming a finite technology space).
- Intuition: The proof utilizes a backward induction argument over the time-level grid. It demonstrates that any time-consistent, robust mechanism must impose the same hard limit as the adaptive speed limit and must provide the agent with "flat" continuation transfers below that limit. If transfers were to slope downward (making the agent pay more for scaling), the principal could "iron" these transfers to induce the agent to stop prematurely, strictly improving the principal's payoff. This deviation would violate time-consistency. Therefore, the only mechanism that cannot be improved upon ex-post is the one with zero transfers below the limit and an infinite penalty above it.

Significance and Claims
The paper claims to provide the first theoretical foundation for adaptive speed limits as a robust regulatory tool.

Unification of Learning: It unifies learning-by-doing and learning-by-waiting within a single framework using multiparameter filtrations, moving beyond the restrictive commutation assumptions of previous literature.
Simplicity vs. Complexity: The authors demonstrate that despite the complexity of the agent's stochastic control problem (involving general learning processes and preferences), the optimal regulatory mechanism is simple and interpretable: a dynamic cap on the rate of scaling.
Commitment-Free Implementation: A critical contribution is showing that this optimal mechanism does not require the principal to commit to future policies. It is the only mechanism that remains optimal even when the principal cannot bind themselves to future actions, addressing the time-inconsistency problems often associated with technology regulation.
Mitigation of Risks: While the model abstracts from specific mitigation technologies, the authors note that the framework can be extended to include endogenous transitions where waiting allows for the development of safety measures, in which case the adaptive speed limit remains the unique time-consistent robust solution.

The authors conclude that speed limits offer a principled way to regulate the externalities posed by risk-seeking firms, balancing the need for learning against the risks of irreversible deployment, without relying on complex, non-robust, or commitment-dependent policies.