3.1.1 Identification of Problematic Data
Original Problem in the Paper
Paper frames two problematic-data classes: samples whose inclusion itself violates legal/ethical principles (copyright/privacy, poisoning, inherently harmful data) and samples whose use causes downstream harms (false beliefs, low-resource-language degradation). Dedicated open problems: operational criteria for difficult cases such as copyright/harm; identifying problematic data without dataset access via behavioral proxies, watermarks, inference/influence methods that remain non-robust; tracing provenance and licenses despite aggregated/misrepresented licenses; silent removal of harmful samples without revealing them to malicious actors.
July 2026 Update & Trajectory
Narrow operational progress exists: open-data/license provenance audits, machine-readable metadata standards, openly licensed pretraining corpora, and hash/list-based removal workflows. Not solved in general: copyright infringement still requires legal/context judgments; behavioral detection without data remains non-robust; provenance coverage is incomplete; silent removal has one high-profile image-link implementation but no universal protocol. I verified 2026 activity for DPI’s 2023-2026 dashboards, but did not verify a 2026 primary source claiming broad solution.
Deployed / Operationalized
- Data Provenance Initiative / Explorer audits 1,800+ text datasets, tracing source, creators, licenses, and uses; useful for open/fine-tuning datasets but not comprehensive frontier pretraining corpora.
- Common Pile v0.1 operationalizes licensed-data collection at 8TB and trains 7B models on 1-2T tokens, showing licensed/open data can support competitive open LMs at modest scale.
- Re-LAION-5B operationalizes harmful-image-link removal via partner-provided URL/image hashes and a diff pool that avoids disclosing suspected CSAM URLs; gated research releases include research-safe filtering.
- Croissant metadata can encode license, content URLs, checksums, record sets, and lifecycle/RAI metadata, making responsible dataset metadata machine-readable where adopted.
New Tractable Vectors
- Build licensed/public-domain pretraining mixtures large enough for 7B-class models and compare against unlicensed baselines.
- Use expert-maintained hash feeds to remove known illegal media from web-scale URL datasets without inspecting or publishing the material.
- Automate dataset-card/license metadata extraction into Croissant-style machine-readable records for newly published datasets.
- Use provenance dashboards to identify license omission/error rates and downstream reuse chains in open dataset ecosystems.
Key Open Questions
- Robust no-access detection of copyrighted/private/harmful training samples in closed frontier models.
- Reliable legal/ethical classification for context-dependent copyright/fair-use, privacy, consent, and harmful-content cases at web scale.
- Provenance for synthetic, scraped, and recursively republished content where original source/license is obscured.
- Governed sharing of harmful-content fingerprints that enables removal without enabling discovery or evasion.
- Verification that problematic-data removal actually reduces downstream model harms without over-filtering marginalized or low-resource content.
Evidence & Primary Sources
- DPI audited 1,800+ text datasets and found license omission above 70% and license error rates above 50%, releasing an interactive Data Provenance Explorer. (2023-10-25): https://arxiv.org/abs/2310.16787
- DPI describes 2023-2026 public dashboards covering training data, web consent, open model ecosystems, and real-use conversations. (2023-2026): https://www.dataprovenance.org/
- Common Pile v0.1 is an 8TB openly licensed/public-domain pretraining corpus and trained 7B Comma models on 1T/2T tokens competitively with similarly sized models trained on unlicensed text. (2025-06-05): https://arxiv.org/abs/2506.05209
- Re-LAION-5B removed 2,236 suspected CSAM links using partner hash lists and released safe diffs that do not disclose the suspected illegal links. (2024-08-30): https://laion.ai/blog/relaion-5b/
- Croissant 1.0 provides machine-readable dataset metadata including license, URL, file objects, sha256 checksums, record schemas, and RAI lifecycle/traceability extensions. (2024-03-01): https://docs.mlcommons.org/croissant/docs/croissant-spec.html