Transparency in AI

Transparency in AI is openness about how an artificial intelligence (AI) system is built, trained, evaluated, and used, together with telling people when they are dealing with one. It answers questions about the whole system: what data trained the model, who built it and for what purpose, how well it performs and for whom, what it should not be used for, and whether a chat partner, image, or decision came from AI. Transparency lets users calibrate their trust, lets auditors and regulators check a system's claims, and lets researchers find flaws such as algorithmic bias, which is why the OECD AI Principles and most other AI ethics frameworks list it as a core principle.

Forms of Transparency#

• Disclosure: People are told when they interact with an AI system or see AI-generated content. Article 50 of the EU AI Act, applicable since August 2, 2026, requires chatbots to make clear that users are talking to a machine, requires providers of generative systems to mark their output in a machine-readable format, and requires deployers to label deepfakes. Standards such as Content Credentials (C2PA) and AI watermarking implement the marking.
• Documentation: A model card reports a model's intended use, training data, evaluation results broken down by group, and known limitations, and Datasheets for Datasets does the same for the data, recording its motivation, composition, collection process, and recommended uses. Under the EU AI Act, providers of general-purpose AI models must also publish a summary of their training content using a template the European Commission released in July 2025.
• Access: Releasing weights, code, or data lets outsiders inspect, reproduce, and test a model directly instead of relying on its developer's reports.
• Explanations of outputs: Individual decisions can be explained with explainable AI (XAI) methods, which the GDPR and the EU AI Act require in some high-stakes settings.

Access comes in degrees. A closed model can be used only through its developer's API, an open-weight model publishes its parameters, and the Open Source Initiative's Open Source AI Definition 1.0, released in October 2024, also requires the complete training and inference code and enough detail about the training data, including its provenance and processing, for a skilled person to build a substantially equivalent system:

Transparency vs. Explainability#

The terms overlap and are sometimes used together, as in "transparency and explainability". Transparency concerns the system and the process that produced it, and it holds even for a model nobody can interpret, such as when a developer documents training data and test results. Explainability concerns why a model produced a particular output, such as which pixels drove a classification. Interpretable models, such as small decision trees, are understandable by design, and mechanistic interpretability reverse-engineers the internal circuits of large networks. Data provenance, the record of where data came from and how it changed, supplies much of what transparency reports contain.

Documentation That Travels with the Model#

Documentation is most useful when it stays attached to the model after deployment. Ultralytics YOLO26 writes a description of the model and its training dataset, the library version, the export date, the task, the input size, and the class names into each exported model, such as the metadata properties of an ONNX file, and Ultralytics reads those fields back to run the model correctly:

import onnx
from ultralytics import YOLO

# Export YOLO26 to ONNX; the exporter embeds provenance metadata in the file
model = YOLO("yolo26n.pt")
onnx_path = model.export(format="onnx")

# Anyone holding only the exported file can read where it came from
metadata = {prop.key: prop.value for prop in onnx.load(onnx_path).metadata_props}
for key in ("description", "author", "version", "date", "task", "imgsz"):
    print(f"{key}: {metadata[key]}")

The printed fields identify the model, its training data, the software that exported it, and when, with no access to the original training run.

Measuring Transparency#

Stanford's Foundation Model Transparency Index scores major AI developers on 100 indicators covering the data and compute behind their flagship models, the models themselves, and how they are used after release. In the 2025 edition, which scored 13 developers, the average fell from 58 in 2024 to 40 out of 100, with developers most opaque about training data and training compute, while IBM set the index's highest score at 95.