Research map / Large language models / AI and society
Position: Data Authenticity, Consent, & Provenance for AI are all broken: what will it take to fix them?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that no complete data provenance system exists for foundation model training data and calling for a unified, structured standard covering authenticity, consent, and legal metadata.
Problem
Foundation model training data is massive, widely sourced, and under-documented, which causes problems in tracing authenticity, verifying consent, privacy, bias, and copyright compliance.
Approach
Surveys and compares existing provenance tools (C2PA, robots.txt opt-outs, consent registries, datasheets/data cards, Data Nutrition Labels, Data & Trust Alliance standard, Common Crawl, Hugging Face, Data Provenance Initiative) and proposes five properties and a set of metadata fields for a unified data provenance standard, with stakeholder-specific recommendations.
Key results
- Comparison table of seven provenance interventions across coverage, verifiability, searchability, extensibility, and attribution, finding no single solution is complete.
- Proposes five requirements for a unified data provenance framework: modality and source agnostic, verifiable, structured, extensible, and symbolically attributable.
- Identifies trade-offs among Common Crawl (accurate, wide coverage, little metadata), Hugging Face (broad but inaccurate/uneven), and Data Provenance Initiative (accurate, detailed, limited scope).
Surprising findings
- Perverse legal incentives discourage companies from disclosing dataset information because disclosure may increase the probability of legal action.
- Popular AI systems such as ChatGPT, Bard, and Llama 2 disclose little or no basic information about their training data.
Limitations
- Position paper with no experiments; claims rest on qualitative survey and citations.
- Proposed standard is not instantiated or validated; adoption challenges and maintenance costs are acknowledged but not resolved.
- Legal analysis focuses on U.S. fair use and EU AI Act, with other jurisdictions only briefly noted.
Experimental setup
Baselines: C2PA robots.txt Data Nutrition Label Data & Trust Alliance Data Provenance Standard Common Crawl Hugging Face Datasets
Builds on: datasheets for datasets data cards data statements Data Provenance Initiative
Closest papers by meaning
- The Foundation Model Transparency Index (NeurIPS 2026 · similarity 0.86)
- A Systematic Review of NeurIPS Dataset Management Practices (NeurIPS 2024 · similarity 0.84)
- Consent in Crisis: The Rapid Decline of the AI Data Commons (NeurIPS 2024 · similarity 0.83)
- Bridging the Data Provenance Gap Across Text, Speech, and Video (ICLR 2025 · similarity 0.83)
- Differentially Private Synthetic Data via Foundation Model APIs 1: Images (ICLR 2024 · similarity 0.81)
- CDI: Copyrighted Data Identification in Diffusion Models (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.