Provable inference – Definition & Meaning

Definition of provable inference

: a proposed AI security technique that produces verifiable evidence that an output came from a particular approved set of model weights running under specified conditions
- Provable inference would let the auditor check that the response came from the approved model rather than a secretly modified copy.
- The lab is testing whether provable inference can detect a model that has been replaced after training.

Origin & history

first known use2026

Anthropic used the term in its 2026 Frontier Safety Roadmap for a research project on reliably signing model outputs so they can be attributed to a specific set of weights. It builds on ideas from cryptographic attestation, confidential computing, and verifiable computation.

Test yourself

Which of these is the meaning of provable inference?

Related terms

inference the act of running a trained model to produce an output, such as answering a prompt or… weights the learned numbers stored in a trained neural network that determine how strongly each part… model the trained program that results from machine learning: a set of learned numbers plus the… AI safety the field concerned with making AI systems behave as intended and avoid causing harm, from… open weights describing a model whose trained weights are published so anyone can download, run, and modify… black box a system whose inputs and outputs can be seen but whose internal reasoning cannot easily be…

Cite this entry

"Provable inference." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/provable-inference/. Accessed 4 Oct. 2026.