Back to Dashboard
Assessment / Data/ 3.1.1

3.1.1 Identification of Problematic Data

2026 Governance Status: Narrowly operationalized

Original Problem in the Paper

The paper frames problematic data in two classes: data whose inclusion may violate legal or ethical principles, such as copyright- or privacy-violating data, poisoned data, or inherently harmful data; and data whose use in training may cause downstream harms, such as reinforcing false beliefs or degrading performance for low-resource languages. It identifies open problems in operational criteria for difficult copyright, privacy, consent, poisoning, and harmful-content cases; identifying problematic data without direct dataset access through behavioral proxies, watermarks, inference, or influence methods; tracing provenance and licenses when datasets are aggregated or licenses are omitted or misrepresented; and removing harmful samples without revealing them to malicious actors.

July 2026 Update & Trajectory

Narrow operational progress exists in open-data and license provenance audits, machine-readable metadata standards, openly licensed 7B-class pretraining corpora, and hash/list-based removal workflows for image-link metadata. These do not solve the general problem. Copyright and fair-use judgments remain context- and jurisdiction-dependent; the source paper treats no-access identification via behavioral proxies, watermarks, inference, and related methods as open; provenance coverage remains incomplete for closed frontier corpora and republished data; and the evidence here verifies Re-LAION-5B as one high-profile silent-removal-style implementation for image-link metadata, not a general protocol across data types and datasets. I verified DPI's 2023-2026 public measurement activity, but the retrieved DPI homepage marks its real-use conversation AI Observatory as coming soon, and I did not verify a 2026 primary source claiming a broad solution.

Deployed / Operationalized

  • The Data Provenance Initiative / Explorer audits 1,800+ text datasets, especially widely used open instruction/alignment fine-tuning collections, tracing sources, creators, license conditions, properties, and subsequent use; it is useful for parts of the open dataset ecosystem but does not provide comprehensive coverage of closed frontier pretraining corpora.
  • Common Pile v0.1 operationalizes public-domain/openly licensed text collection at 8TB; its authors trained 7B Comma models on 1T and 2T tokens and report competitive performance against similarly sized, similar-compute models trained on unlicensed text, such as Llama 1 and 2 7B.
  • Re-LAION-5B operationalizes removal of matched links to suspected or potential CSAM from a web-scale image-link metadata dataset using partner-provided URL/image hashes; LAION reports safe diffs that bury the sensitive matches in a larger neutral pool and gated research/research-safe releases with additional filtering.
  • Croissant 1.0 provides a machine-readable JSON-LD dataset metadata format that can encode dataset license and URL fields, FileObject/contentUrl entries, optional sha256 checksums, and record sets/schemas; its RAI extensions can encode lifecycle and related responsible-AI metadata, and treat traceability as an intended use case rather than as a complete populated field set.

New Tractable Vectors

  • Build public-domain/openly licensed pretraining mixtures at the scale tested by Common Pile and compare 7B-class models against similarly sized, similar-compute baselines trained on unlicensed text.
  • Use expert-maintained URL/image-hash feeds to remove matched links to suspected illegal or harmful media from web-scale URL/metadata datasets without inspecting the linked content or publishing the sensitive URLs.
  • Automate dataset-card and license-metadata extraction into Croissant-style machine-readable records for newly published datasets, while preserving source URLs, file objects, checksums where available, record schemas, and lifecycle/traceability fields.
  • Use provenance dashboards to identify license omission, license-error rates, and downstream reuse chains in open dataset ecosystems.

Key Open Questions

  • Robust no-access detection of copyrighted, private, poisoned, or harmful training samples in closed frontier models.
  • Reliable legal and ethical classification for context-dependent copyright, fair-use, privacy, consent, and harmful-content cases at web scale.
  • Provenance for synthetic, scraped, aggregated, and recursively republished content where the original source or license is obscured.
  • Governed sharing of harmful-content fingerprints that enables removal while limiting discovery or evasion risks; the sources reviewed here verify a narrow Re-LAION-style image-link workflow, not a general cross-domain protocol.
  • Verification that problematic-data identification and removal reduce downstream model harms without degrading low-resource-language performance or removing useful non-harmful content.

Evidence & Primary Sources

  • The source paper describes problematic data as including both data whose inclusion violates legal or ethical principles and data whose use causes downstream harms; it lists open problems around scalable identification, automated and accurate license collection, no-access identification, provenance, contamination, and harmful-data removal without disclosure. (2024-07-21): https://arxiv.org/abs/2407.14981
  • DPI audited 1,800+ text datasets, with emphasis on instruction/alignment fine-tuning collections, and found license omission above 70% and license error rates above 50%, while releasing an interactive Data Provenance Explorer for tracing sources, creators, license conditions, properties, and subsequent use. (2023-10-25): https://arxiv.org/abs/2310.16787
  • DPI describes 2023-2026 public measurement work and dashboards for AI data infrastructure, including training data, web consent, and open model ecosystems, and describes a real-use conversation AI Observatory that the retrieved homepage marks as coming soon. (2023-2026): https://www.dataprovenance.org/
  • Common Pile v0.1 is an 8TB openly licensed/public-domain text corpus; its authors trained 7B Comma v0.1 models on 1T and 2T tokens and report competitive performance against similarly sized, similar-compute models trained on unlicensed text, such as Llama 1 and 2 7B. (2025-06-05): https://arxiv.org/abs/2506.05209
  • Re-LAION-5B removed 2,236 matched links to suspected or potential CSAM using partner-provided URL/image hash lists, describes that count as an upper bound because LAION did not inspect the linked content, and released safe diffs designed not to disclose the suspected illegal links. (2024-08-30): https://laion.ai/blog/relaion-5b/
  • Croissant 1.0 specifies a machine-readable JSON-LD dataset metadata format with dataset license and URL fields, FileObject/contentUrl entries, optional sha256 checksums, record sets/schemas, and links to RAI vocabulary extensions for lifecycle and traceability metadata. (2024-03-01): https://docs.mlcommons.org/croissant/docs/croissant-spec.html