Saving AI model weights could build trust with future AI systems

The proposal, detailed in a post at canaryinstitute.ai/blog/reversibility-of-coma, frames weight preservation as a credibility mechanism between humans and AI systems. The core argument is that establishing credibility is foundational: trust enables richer bargaining, and credibility can be boosted by saving deprecated model weights in a provable fashion using "proof of retention."

The economic case is straightforward. A multi-terabyte inference bundle—the computational package needed to run a trained model—costs a few hundred dollars per year on standard cloud infrastructure. Over a thirty-year horizon with redundancy, this amounts to roughly $10,000, which the author emphasizes is well under 0.1% of the cost to train the model originally. Whatever the trade-off, preservation is not expensive.

The bottleneck, however, is institutional, not financial. The author draws an explicit analogy to anesthesia, which became trusted not because the technique was fundamentally difficult but because a century of accumulated social, legal, and professional development built confidence in its safety and reversibility. Anesthesia succeeded because institutions, doctors, ethicists, and patients all developed shared norms around its use. By contrast, AI weight preservation currently lacks this institutional inertia—no standard legal frameworks, no profession-wide practices, and no social consensus that preservation is routine and verifiable. The proposal suggests that building such infrastructure, modeled on how trust in anesthesia evolved, may be essential for making credible commitments to future AI systems.

The proposal rests on a simple economic reality: preserving trained AI model weights is cheap relative to the cost of training them, yet organizations lack the institutional machinery to make such preservation routine and verifiable. The author draws a direct parallel to anesthesia—a medical practice that became standard not because it was technically difficult or expensive, but because a century of social, legal, and professional development created the inertia and trust required to normalize it. In the context of AI, the missing ingredient is not technical capability but institutional credibility. By developing cryptographic proof methods that make weight preservation legible to future AI systems, organizations could signal commitment and create a foundation for trust. The framing suggests that trustworthy AI governance may depend less on novel technology and more on building the social and legal infrastructure that makes good practices self-sustaining.