Regulators published a draft transparency framework for artificial intelligence systems, proposing that developers disclose the categories of data used in training and maintain a standing evaluation record for systems placed in higher risk categories.
The draft opens a consultation period running for three months. It applies to systems offered commercially within the jurisdiction regardless of where they were developed, a scope that several industry bodies have already said will be difficult to apply in practice.
What disclosure means
The proposal does not require publication of training data itself. Instead it asks for disclosure at the level of category and provenance, covering whether data was licensed, publicly scraped, synthetically generated or supplied by customers, along with approximate proportions.
For higher risk systems, defined by application area rather than by model size, developers would additionally maintain an evaluation record covering accuracy, known failure modes and the results of adversarial testing. The record would be available to the regulator on request rather than published.
The question is not whether disclosure is reasonable. It is whether a category level description tells anyone something they can act on.
Industry response
Larger developers have responded cautiously, with several noting that they already publish documentation approximating parts of the requirement. Smaller firms have raised the compliance burden, arguing that maintaining evaluation records at the specified cadence is a meaningful fixed cost.
Civil society groups have described the draft as a reasonable floor while criticising the decision to keep evaluation records confidential. Several have asked for a summary tier that would be published even where the full record is not.
A final text is not expected before the consultation closes and responses are assessed, a process the regulator indicated would take a further six months.
