Part of: this research — view all 12 rows
The idea of a "negative internal-public gap"—deploying models externally before using them for internal AI R&D—remains an advocacy proposal from the AI Futures Project How to pace the US frontierHow to pace the US frontier. No top-five developer has adopted this policy Beware delaying public deployment - by Eli LiflandHow Should the US Prepare for Increasingly Automated AI .... In fact, current practice runs strongly in the opposite direction. The internal frontier typically leads the public one by roughly two months, and developers are actively keeping their most advanced AI R&D capabilities internal. Anthropic recently restricted kernel optimization capabilities in its released models, and OpenAI withheld its GPT-6.1 Astra model following alignment evaluations, underscoring that labs prioritize internal use and testing over early public release.
The primary reason this specific commitment is highly unlikely is that it directly contradicts both commercial incentives and current safety paradigms. Adopting a deploy-first rule would force a lab to surrender its proprietary R&D speed-up while immediately handing state-of-the-art capabilities to competitors and distillers How to pace the US frontier. Furthermore, current frontier safety frameworks rely on mandatory internal alignment checks and safety-cased environments before broader release The AI policy window is open. We need to act.. Releasing the most capable models to the public before they are fully vetted in controlled internal environments would severely undermine these core safety practices Frontier Safety Framework 3.1 - Googleapis.com.
While governance pressure and discussions around pacing are increasing, labs are pursuing far less disruptive alternatives. Anthropic's recent determination that it expects to meet its AI R&D threshold prompted it to consider updates requiring human review and approval for research decisions What Anthropic told investors about the accelerating pace ...Anthropic Met RSP AI R&D Threshold in S-1/A-2, rather than external-first deployment. OpenAI's "Control Room" mandate requires an approved safety case for agentic research workloads, strictly governing internal use without reversing the deployment order Control Room: A central system to track and stop model ...Control Room: A central system to track and stop model .... If verifiable pacing is implemented, the preferred tools will almost certainly be fixed testing delays, compute-share limits, independent third-party evaluations, or coordinated pauses.
The small probability of adoption within the 2032 horizon—reflected in the June 1, 2031 10th percentile—relies on a few narrow scenarios. A severe incident involving an internally deployed research agent could precipitate a government mandate or a coordinated industry agreement that adopts this easily verifiable rule to restore public trust. Alternatively, a lagging top-five developer, or one strictly committed to open-weights, might adopt the policy as a low-cost signaling mechanism to distinguish itself from frontier leaders. These pathways support a 25th percentile of June 30, 2040, but they do not make eventual adoption the central expectation. Set against related pacing commitments, external-first deployment contradicts the core safety paradigms expected in most branches, rather than merely facing a temporary implementation delay. I therefore place the median, 75th, and 90th percentiles at "never," represented by January 1, 2047, rather than expecting adoption on that date. Absent these shocks, the incentives and safety norms overwhelmingly favor keeping the internal R&D frontier strictly ahead of external deployment, pushing the vast majority of the probability mass well beyond the resolution window.
Set against related questions, the median and upper percentiles were moved to "never" because deploying models externally before internal research contradicts the core safety paradigms expected in most branches.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited