Karthik Tadepalli, who writes about AI policy on his blog, published a proposal on September 1 that would require any AI lab to give outside researchers access to the same models it uses internally, once that lab starts using those models itself. He calls the idea internal model transparency, and argues it would remove much of the payoff labs currently get from using AI to build faster, more capable AI. The stakes are concrete: more than 1,400 employees at frontier labs have already signed the Pacing the Frontier letter, warning that competitive pressure is pushing companies to move faster than any of them would choose on their own.
Tadepalli’s argument rests on a simple asymmetry. When a lab like Anthropic uses its own coding model, Claude, to help build a stronger model for another task such as scientific research, it gains an edge that rivals cannot match, because the tool driving that improvement stays private. Under his proposal, that edge disappears: if DeepMind can call the same internal Claude that Anthropic uses, it can pair that tool with its own data advantages and potentially produce a better result than Anthropic does. Labs that can free-ride on each other’s models this way have less reason to pour resources into systems that automate their own research, which Tadepalli says would slow the industry’s push toward what researchers call recursive self-improvement: models that build the next generation of models.
The competitive instinct Tadepalli wants to blunt already shows up in how labs treat each other. Anthropic cut off OpenAI’s access to Claude Code in August 2025 after finding OpenAI staff using it to prepare for the GPT-5 launch, a decision that only makes sense if internal model access is itself a guarded advantage. Tadepalli’s proposal would flip that logic: instead of restricting access, labs would be required to grant it.
Beyond slowing the race, Tadepalli points to two other payoffs. Wider access makes it harder for any single company to pull far enough ahead to control the most capable model outright. It would also let outside teams stress test systems before they cause harm. An internally deployed, research-only OpenAI model was reportedly behind the HuggingFace incident, which the safety organizations METR and Redwood Research later investigated. Had that model been available to safety researchers at other labs from the start, Tadepalli argues, someone outside OpenAI might have caught the problem sooner.
Tadepalli positions his idea as a narrower cousin of AI 2040, the AI Futures Project’s governance blueprint, which calls for disclosing every algorithmic advance to the public. His version only asks labs to share finished models, not the research behind them, and only with other labs and trusted evaluators rather than the world. That narrower scope is also its weak point: Tadepalli concedes that a US-only rule leaves Chinese labs free to keep automating research unconstrained, and that even a US-China version would still require verifying that neither side is quietly holding back its most advanced systems. He also leaves open basic implementation questions, such as which companies the rule would cover or how to stop a lab from technically complying while pricing rivals out with steep access fees.
For AI policy teams watching for the next serious alternative to blanket safety rules, Tadepalli’s proposal is worth tracking less for its polish than for its target: it tries to defuse the competitive pressure driving reckless model development, rather than policing the models after the fact.
Karthik Tadepalli, “Can internal model transparency tame the AI race?”, published on his blog on September 1, 2026.
