When Is a Summary of Data Used to Train GPAI Sufficiently Detailed?
summary
EU law soon forces GPAI providers to publish a “sufficiently detailed” training-data summary and the Commission’s AI Office is drafting a template that lists concrete fields: overall dataset size, data provenance (sources and % from scraped vs licensed vs user data), diversity metrics, preprocessing and filtering steps, use of synthetic or distilled-model data, and model-to-model dependencies. Companies push back citing trade secret and security; open-source hubs already publish much of this. Template plus guidelines aim to balance enforceability and commercial sensitivity; obligation kicks in when a model is placed on the EU market (from 2 August).
AI-written summary · about
chapters (7)
- 0:00AI training data transparency
- 10:36EU template for training data
- 24:24Open source transparency needs
- 36:07Trade secrets vs transparency
- 51:40Wikipedia in AI datasets
- 1:00:05Q&A on model distillation
- 1:12:31Closing remarks on transparency and AI development
related talks
01h 17mBias, Discrimination and Data Protection in the Age of GPAI
39mGPAI Code of Practice & Systemic Risk
35mThe European AI Act: a data scientist’s perspective
50mAI Development and Copyright: Public ≠ Free Use
01h 14mLatest Developments in AI and Data Protection Legal Uncertainty Despite
53m10 Key Challenges for AI within the EU Data Protection Framework
28mAI Act - What is important to know from a developers perspective?
01h 11mThe Intersection of AI and Regulation How Organisations and Regulators Should...