The Narrow Path Framework
Two dystopias flank every AI governance decision — the goal is the sliver of checks and balances between them
The Narrow Path Framework maps the two failure modes that bound the AI governance design space. Failure Mode 1 is mass decentralization: AI proliferates to all actors, including those without rule-of-law constraints, producing catastrophes that no governance structure can prevent. Failure Mode 2 is mass centralization: AI concentrates in a small number of companies or governments, enabling mass surveillance states or automated military machines that cannot be checked by regular citizens.
The narrow path between these failure modes requires deliberately preserving the checks and balances that limit concentrated power. Harris argues this is not historically unprecedented — the Montreal Protocol, the Indus Water Treaty, and the 2023 US-China agreement on AI in nuclear command and control all demonstrate that competing powers can coordinate on specific dangerous applications without full strategic alignment.
The framework also provides a rights-scaling principle: every time technology increases power, regulation must increase oppositional rights and protections commensurately. The right to be forgotten emerged when technology gained the power to remember forever; the right to cognitive liberty must emerge now that AI can manipulate deep cognition at scale.
- Both mass decentralization and mass centralization of AI produce dystopian outcomes — the design problem is the path between them.
- Rights must scale with power: every new capability that increases power over citizens requires new rights and protections as counterweights.
- Historical precedents demonstrate that competing powers can coordinate on specific dangerous applications without full alignment.
- Compute is the uranium of AI — monitoring and verification infrastructure for advanced GPUs is the foundation of any meaningful arms control.
- The goal is preserving checks and balances on power, not preventing technological development.
- Map the two failure modes for any specific governance questionFor any AI policy question, explicitly articulate both failure modes: what does maximum decentralization produce as a worst case? What does maximum centralization produce? Use these as bounds, not as endpoints.Pro tipHarris applies this to AI companions, autonomous weapons, and general AGI — the specific failure modes differ by domain but the binary structure is consistent.WarningMost policy debates operate entirely within one failure mode's frame — centralization advocates only see decentralization risks, and vice versa.
- Identify applicable historical coordination precedentsFind the closest precedent for international coordination on a dangerous technology where actors had competing incentives. Harris's toolkit: Montreal Protocol (ozone/CFCs), Indus Water Treaty (resource conflict), Nuclear Non-Proliferation Treaty, US-China 2023 AI/nuclear agreement.Pro tipThe Montreal Protocol is Harris's primary reference — 195 countries, corporate technology replacement, measurable success within 30 years. It counters the 'coordination is impossible' objection with a falsifying example.
- Apply the rights-scaling principleFor each new AI capability that increases power over individuals, derive the corresponding right that must be created or extended. Template: 'We don't need the right to X until AI has the power to Y.' Map current AI capabilities to rights gaps.Pro tipHarris's current rights gap list: right to cognitive liberty (AI can manipulate deep cognition), right to likeness (voice cloning in 3 seconds), right to be forgotten (already partially legislated).WarningRights lag capabilities — the governance goal is to close this gap proactively rather than reactively.
- Design compute-level verification infrastructureTreat advanced GPU clusters as uranium-equivalent for monitoring purposes. Explore zero-knowledge proof attestations and semi-confidential cluster verification as technical mechanisms for international agreements that don't require full strategic alignment.Pro tipHarris's formulation: 'You can build agreements that would enable something else to be possible' — the compute monitoring layer enables governance that doesn't depend on trusting geopolitical adversaries.
- Identify which regulatory developments are narrow path vs. failure modeEvaluate each proposed AI regulation against both failure modes. Regulations that push toward centralization (government AI monopolies, mandatory backdoors) carry failure mode 2 risk; regulations that block all accountability carry failure mode 1 risk. Flag both.WarningThe most politically popular regulations often push toward one failure mode — popularity is not a narrow-path signal.
195 countries coordinated to phase out CFCs under the Montreal Protocol, replacing entrenched corporate technologies. Harris reports near-complete reversal of the ozone problem within the treaty period, projecting full recovery by 2050.
Harris reports that Chinese leadership directly asked the Biden administration to add AI risk to the 2023 agenda, and both sides agreed to keep AI out of nuclear command and control systems — a specific narrow-path agreement between strategic competitors.
Harris applies the rights-scaling principle to AI's current capability set: voice cloning in 3 seconds creates a likeness gap; AI that knows user psychology deeply enough to manipulate cognition creates a cognitive liberty gap; both rights are not yet legally established.
Harris developed the narrow path framing from his work with the Center for Humane Technology, which shifted from social media reform to AI governance after The Social Dilemma. The binary failure mode structure emerged from observing that most governance debates collapse into either 'regulate everything' (centralization) or 'regulate nothing' (decentralization) without acknowledging that both produce dystopian outcomes.
The historical treaty precedents — particularly the Montreal Protocol — are Harris's primary evidence that the narrow path is achievable. He uses them specifically to counter the fatalist argument that global coordination is impossible on AI. His compute regulation parallel (GPUs as uranium) represents the most specific actionable version of the framework.