A thread gaining momentum on Lobsters right now, tied to a blog post from July 14, raises a question that security people have been quietly frustrated about for a while: How do you actually know that the AI service you're using is running the model they claim? And that the output hasn't been tampered with along the way?

This isn't about hallucinations or bad answers — it's about trust at the infrastructure level. Could a cloud provider swap out your model for something cheaper? Could a compromised server send you manipulated responses? Right now: yes, and you wouldn't necessarily know.

Verifiable inference is the missing link between "we promise we're running GPT-X" and "here is the cryptographic proof."

What the community is debating essentially comes down to three approaches:

TEEs (Trusted Execution Environments) are the hottest option right now. Intel TDX, AMD SEV-SNP, and ARM CCA create isolated "vaults" inside the processor where code cannot be tampered with — not even by the operating system or cloud admins. NVIDIA began rolling out GPU support for TEEs in 2024, making this relevant for actual AI workloads. The point is that you can obtain a cryptographic attestation confirming that this exact model ran this exact prompt in an unmodified environment.

Can you prove the AI isn't lying? Now people want answers - Bilde 1

Cryptographic signing is the simpler sibling — NVIDIA has already been signing models in the NGC Catalog using the OpenSSF standard since March 2025. This solves the provenance problem (is this the genuine model?), but not runtime manipulation.

Zero-knowledge proofs are the elegant but heavyweight solution. Mathematical proof without revealing the input. The problem is that ZKP for large language models is computationally demanding enough to make it impractical today — hence the interest in the TEE alternative.

Why does anyone care now? Because AI is being used in increasingly critical contexts — legal documents, medical advice, financial decisions. The question of who actually made the decision and which model was behind it is becoming regulatorily relevant. EU AI Act implementation is underway, and auditability is no longer just nerd talk.

These are early signals from a technical niche, and none of the solutions are production-ready at scale yet. But the fact that the Lobsters crowd — typically more skeptical and technically rigorous than Reddit — is picking this up is worth noting. Watch for whether this surfaces at security conferences like DEF CON or Black Hat this fall.