A Narrow Path: How to Secure Our Future

A Narrow Path: How to Secure Our Future is the flagship policy plan of ControlAI, authored by ControlAI Founder and CEO Andrea Miotti with Tolga Bilge, Dave Kasten, and James Newport. The plan was first released by ControlAI in October 2024 — announced by Miotti on 2 October 2024 with the launch line "If attempts to build superintelligent AI succeed, we face extinction as a species. We must choose a different path, one where we control what we create ... Humanity has no plan, so we built one" — and is now hosted at narrowpath.co with a 19 March 2026 timestamp on the current authoritative version. It was discussed at length by Miotti on the London Futurists podcast in late October 2024 and has been endorsed by MIT Professor Max Tegmark, philosopher Toby Ord, and cognitive scientist Gary Marcus.

The plan opens from the working premise — articulated in its own framing — that "we have no method to currently control an entity with greater intelligence than us", and proposes a phased trajectory through which states can prevent the development of artificial superintelligence until the science of controllable transformative AI has been laid down. The trajectory is organised as three named phases: Phase 0 — Safety, a twenty-year moratorium on the development of artificial superintelligence implemented through national-government legislation and built around a compute-threshold licensing regime triggered at training runs above 10^25 FLOP, paired with novel measures restricting directly dangerous AI capabilities; Phase 1 — Stability, the construction of an international AI-oversight architecture resilient enough to hold the moratorium against the geopolitical pressures that would otherwise erode it; and Phase 2 — Flourishing, the eventual development of "safe-by-design" transformative AI under human control, conditional on the prior establishment of a robust science and metrology of intelligence. Each phase has its own set of named institutions, legal instruments, and policy mechanisms — the licensing regime in particular requires applicants to file the planned training run, expected system capabilities, risk-management and mitigation measures, and safety protocols before any model above the 10^25 FLOP threshold can be developed.

Within the corpus, A Narrow Path is the ControlAI-anchored Publication and the first AI-safety / pause-movement entry in a Publications slice previously occupied by humanitarian-disarmament, algorithmic-justice, climate-and-data-centres, and participatory-data-stewardship anchors. Its working theory of change differs from the corpus's other AI-safety entry-point — PauseAI's street-protest and mass-political-pressure model — by deliberately operating through the institutional-advocacy channel: the document is built to be picked up by elected officials, with compute thresholds, licensing regimes, and a phased moratorium designed as components a national legislator can sign onto without committing to a full technical worldview. The plan's policy positions feed directly into ControlAI's UK parliamentary-briefing programme — Miotti's written evidence submission RAI0031 to a UK parliamentary AI-regulation inquiry, the campaign statement that crossed 100 cross-party UK signatories by early 2026, and the draft bill ControlAI prepared with parliamentary lawyers — and stand alongside Losing Humanity as the corpus's second example of a long-form civil-society policy artefact written in the legal-and-treaty register that national governments can act on. Where Losing Humanity opened a humanitarian-disarmament tradition aimed at autonomous-weapons systems, A Narrow Path opens the compute-governance / superintelligence-licensing tradition aimed at frontier-AI development itself.