Training data transparency gap

What it is

What a model was trained on determines much of what matters about it: whose data is inside it, what it may reproduce, what biases it carries, what rights were needed. More and more parties ask. Regulators ask about lawfulness and bias. Customers ask in due diligence. Rights holders ask whether their works were used. The EU AI Act requires providers of general-purpose models to publish a summary of the content used for training. The risk is that the organisation has no answer, because it kept no record of its own training data or because its supplier discloses nothing.

How it happens

• Datasets were assembled over time with no record of their sources.
• A pre-trained model was adopted whose provider describes its data in general terms only.
• Fine-tuning data was drawn from many internal systems with no inventory.
• Records exist and are held by people who have left.

Why it matters

An organisation that cannot describe its training data cannot show a lawful basis, answer a rights request, respond to a copyright claim or complete a customer's questionnaire. Silence is read as having something to hide.

Warning signs

• No dataset documentation exists for models in production.
• Supplier responses on training data are vague.
• Due diligence questions about training data take weeks to answer.
• The organisation has no position on what it would disclose publicly.