AI provenance is the documented lineage of an AI system, covering the origin and composition of its training data, the sequence of training and fine-tuning stages it underwent, the versions of code and infrastructure used, and any post-training modifications such as alignment tuning or quantization. Maintaining provenance records makes it possible to trace unexpected model behavior back to its source—whether a contaminated dataset, a problematic annotation batch, or a misconfigured training run—and to reproduce or audit results independently.

Think of it like a chain of custody for evidence in a legal case. Each time the evidence changes hands, it is logged. If something looks wrong at trial, investigators can trace every step back to the crime scene. Without that chain of custody, the evidence's integrity cannot be verified.

In practice, provenance is often captured in a combination of model cards, dataset documentation (datasheets for datasets), and experiment-tracking systems such as MLflow or Weights & Biases. A minimal provenance record for a fine-tuned model might specify the base Foundation Model version, the fine-tuning dataset name and version hash, the RLHF reward model used, the compute cluster and software stack, and the date of each run.

AI provenance connects to AI Ethics (transparency depends on knowing what data was used), AI Governance (regulators increasingly require provenance documentation), Bias (tracing discriminatory outputs requires knowing which training splits contributed to them), Fine-tuning (each fine-tuning stage is a provenance event), and Pretraining (the base training run is the root node of the provenance graph).

You will encounter AI provenance requirements in the EU AI Act's technical documentation obligations, in enterprise procurement checklists for high-risk AI systems, and in open-source model release notes that list dataset sources and training recipe details.