Anthropic’s new research could change how we verify AI-generated answers and the work behind them.
You ask an AI to review some software code.
It comes back with a reassuring answer: “No problems found.”
But how do you know it actually used the powerful AI model you requested? Could the provider have substituted a cheaper one? Could the response have been altered along the way?
You can’t tell from the answer alone. Even a saved conversation isn’t enough.
On November 2, 2026, Anthropic published research that could help change that.
From “trust me” to “check it yourself”
Anthropic introduced FlashProve, a system that uses cryptography to verify AI computations.
The principle is straightforward.
Instead of simply claiming that a particular AI model produced an answer, a provider could supply mathematical proof that a specific, previously committed version of a model computes the checked output.
Someone else could verify that proof without having access to the model’s private workings or the powerful computers needed to run it.
The model’s identity still needs to be established separately, but the computation itself becomes checkable.
Anthropic demonstrated this with Llama 4 Maverick, a model containing 400 billion parameters. Its experimental proofs can be checked on a regular CPU in seconds.
And here’s a curious detail: Claude helped researchers improve the cryptographic system itself.
AI is helping develop tools to verify AI.
The technology is still experimental. For now, FlashProve verifies individual token computations under limited conditions, rather than entire AI conversations. Producing those proofs also takes substantially more computing power than generating answers normally.
Still, the direction is promising.
Imagine an AI security review with proof
One possibility Anthropic discusses is particularly interesting.
Before releasing software, a developer asks an AI model to look for security vulnerabilities.
Today, the developer might publish a report saying the model found nothing suspicious.
In the future, they could potentially provide proof that a specified model searched for vulnerabilities, performed a defined amount of computation, and failed to find a bug.
Anthropic calls this proposed approach a proof of failure.
It wouldn’t prove that the software is safe. A model can miss problems.
But it could make a narrower claim independently checkable: the specified computation was carried out, rather than merely reported.
That’s a meaningful difference.
And what if someone needs that proof a year later?
Imagine that software is now used by thousands of people.
Someone reviewing its history wants to examine the original code, the AI’s instructions and responses, and the cryptographic proofs associated with the review.
All those records need to remain available.
A valid inference proof can establish something about a computation. An independently timestamped record can establish that particular files existed by a particular date.
One doesn’t replace the other. And neither guarantees that the AI’s conclusions were correct.
At ProofStamp, we’re working to make independent verification of important digital records a common practice. That includes keeping evidence checkable outside the platform that originally created it.
As AI agents begin writing code, reviewing documents, making recommendations, and acting on our behalf, this distinction could matter more.
We may soon expect AI systems to provide proof of particular computations.
And when someone asks about that work a year later, we should still be able to check the evidence.