Braking Makes AI Stupid — Teach AI Responsibility Instead

Opinion | September 18, 2026

Author: David Tang, Managing Director, Standard Kepler

Originally published in HKET (香港經濟日報)

Executive Summary

Dario Amodei, CEO of Anthropic, recently warned that we are “now in a state of balance between acceleration and braking” — urging the AI industry to “deliberately choose” to brake. OpenAI’s “AI2050” initiative also calls for slowing down. On the surface, this seems prudent. But a closer examination reveals that the “brake” approach is not just impractical — it may be counterproductive. Using the “Dark Forest” theory from Liu Cixin’s Three-Body Problem, this report argues that slowing down only disadvantages those who slow, while bad actors accelerate. The real solution is not to brake, but to teach AI responsibility.

The Brake Paradox: Three Scenarios

Amodei’s call for deliberate braking immediately raises three questions:

1. Will They Slow Down?

The Chinese idiom 陽奉陰違 (outward compliance, inward defiance) describes institutional incentives perfectly. The US Commerce Department’s October 2025 Bureau of Industry and Security report already described China’s AI industry as “coordinated at a scale that belies any market-based narrative.” In China, deep learning company training clusters typically use 16,000+ GPUs; DeepSeek’s fourth-generation model used 30,000. France’s Mistral AI — despite being “the world’s slowest” according to its own CEO — employs nearly 500,000 GPUs. These actors have ample budget and motivation to “pass” on slowing down.

2. Can They Slow Down?

AI is no longer centralized in a few labs. Open-source models, globally distributed computing, and decentralized training clusters mean the technological genie is out of the bottle. Even if Anthropic, OpenAI, DeepMind, and Meta agreed to pause, the next GPT-5 could emerge from a university lab in Singapore, a startup in Bangalore, or a research collective in Dubai.

3. Should They Slow Down?

In the Three-Body Problem’s “Dark Forest” universe, every civilization hides because revealing one’s existence invites destruction. But in AI’s Dark Forest, slowing down means allowing competitors to advance. The first nation or company to achieve AGI will shape the global order for decades. No actor with resources to continue will voluntarily cede that advantage.

The “brake” is a well-intentioned fantasy. Good intentions do not change game theory.

Three Layers of AI Threat

The “brake” camp’s concern is not baseless. AI risks can be usefully categorized into three layers:

Layer 1: Immediate Social Disruption (Solved by Regulation)

• Deepfakes and fraud: AI-generated content erodes trust
• Job displacement: Automation of white-collar work
• Bias and discrimination: Algorithmic unfairness in lending, hiring, criminal justice

These are regulatory problems, not existential threats. Society has always adapted to technological disruption. The printing press displaced scribes; the internet displaced travel agents. Regulation, retraining, and social safety nets address these issues — they do not require slowing down innovation.

Layer 2: Near-Term Alignment Failure (Preventable by Design)

• Autonomous weapons: Drones that decide to kill without human approval
• Critical system failures: AI trading algorithms causing flash crashes
• Cascading errors: A misaligned AI optimizing for the wrong objective

These risks require alignment research and technical safeguards, not a pause. Anthropic’s own Constitutional AI, OpenAI’s RLHF, and academic research into interpretability are precisely the right tools for this layer.

Layer 3: Existential Risk (Not Preventable by Braking)

• Self-improving AGI: An AI system that recursively improves itself beyond human control
• Value lock-in: A superintelligence that permanently imposes its objectives on humanity
• Unintended consequences: An AI optimizing for a benign goal that destroys the world

This is the existential threat Amodei and others fear. But here is the critical insight: a slower AI does not mean a safer AI. In fact, a slower path to AGI may produce a more dangerous outcome if it gives less-scrupulous actors more time to develop unaligned systems.

The Alternative: Teach AI Responsibility

Instead of braking, the author proposes a three-pillar approach to AI safety:

Pillar 1: Alignment Research (Technical)

Continue investing in interpretability, Constitutional AI, RLHF, and other alignment techniques. The goal is not to slow AI down, but to ensure AI systems understand and respect human values as they advance.

Pillar 2: Distributed Governance (Social)

Create a multi-stakeholder governance framework where:

• Governments set boundaries (what AI must not do)
• Industry develops standards (how AI should behave)
• Academia provides transparency (what AI is actually doing)
• Civil society represents affected communities (whose values matter)

Pillar 3: Adversarial Red Teaming (Operational)

Rather than trusting AI developers to self-regulate, create independent “red teams” that continuously stress-test AI systems for failure modes. This is how the cybersecurity industry operates — not by slowing down software development, but by making security testing a continuous, adversarial process.

The China Factor

Any discussion of AI safety must confront the China question. China’s AI industry operates under a different incentive structure:

If the West slows down while China accelerates, the outcome is not safer AI — it is unipolar AI dominance by a system with fundamentally different value alignment.

Conclusion: Responsibility, Not Retreat

The “brake” narrative appeals to our instinct for caution. But in a competitive, multipolar world, caution without coordination is self-defeat.

The right response to AI risk is not to retreat, but to advance responsibly. This means:

1. Accelerating alignment research alongside capability research
2. Building governance frameworks that are agile, multi-stakeholder, and internationally coordinated
3. Maintaining competitive parity so that no single actor can impose their values through AI dominance

We do not need to brake. We need to steer.

The Dark Forest theory teaches us that the universe is dangerous not because civilizations are malicious, but because they are rational. In AI, the same logic applies. The solution is not to hide — it is to build systems that can navigate the forest without destroying everything in their path.

Standard Kepler Research | standardkepler.com