SynWeb

1st Workshop on Trustworthy Synthetic Data for the Web: Generation, Verification, and Governance

Half-day workshop at TheWebConf 2027, Dublin, Ireland. Date to be announced.

  1. GenerateCorpora, profiles, records, and benchmarks from language and multimodal models
  2. ConstrainRetrieval, knowledge, schema, and rules
  3. VerifyFactuality, consistency, uncertainty, provenance
  4. CurateFilter, document, and combine with observed data
  5. EvaluateMeasured on search, recommendation, and other web tasks
  6. ReleaseLicensing, privacy, governance, and reporting

About the workshop

Large language and multimodal models are now used to create training corpora, benchmarks, user and item profiles, structured records, and evaluation data for web-facing systems. At the same time, machine-generated content is becoming part of the web itself. This raises open questions about provenance, contamination, distribution shift, privacy, and whether synthetic data actually improves downstream web tasks.

SynWeb brings together researchers and practitioners in web data management, information retrieval, NLP, recommender systems, and responsible AI to study this problem end to end: how synthetic web data should be generated, verified, documented, stress-tested, and evaluated before it is reused. The program combines technical contributions with a benchmark design session and a community roadmap.

Objectives

  1. Verification-first design principles for synthetic data used in web-facing systems, with explicit attention to provenance, grounding, uncertainty, and reproducibility.
  2. Web-specific failure modes, including fabricated entities or relations, synthetic-content contamination, distribution shift, privacy leakage, and feedback effects when generated content is repeatedly reused.
  3. Shared evaluation protocols that connect intrinsic data quality to measurable downstream utility on search, recommendation, fact verification, knowledge extraction, and content moderation.
  4. A community roadmap and benchmark specification identifying priority tasks, reporting practices, and reusable resources for trustworthy synthetic web data.

Why now

Access to platform and user-generated data is increasingly constrained by licensing, privacy, cost, and platform policy, while generative models make large synthetic datasets cheap to produce. Web systems also operate in environments that mix human- and machine-generated content. Methods for grounded generation, uncertainty estimation, provenance tracking, and model-based verification have advanced quickly, but they are usually evaluated in isolation. The community still lacks common protocols for deciding when synthetic data is fit for a particular web task and how its limitations should be reported.

Focus: the data layer

Recent TheWebConf workshops address trustworthiness at the level of models, grounding, harmful content, and agent architectures. SynWeb focuses on the data layer: how synthetic web data is generated, verified, documented, combined with observed data, and evaluated before downstream web systems reuse it. Synthetic-content detection is studied where generated content enters, contaminates, or augments the pipelines used to train, evaluate, or operate web systems.

What the workshop will produce

Participants will jointly identify evaluation requirements, benchmark tasks, and reporting practices. The organizers will publish the outcome as the SynWeb roadmap, with a preliminary benchmark specification and a minimum reporting checklist for trustworthy synthetic web data.

See topics, submission types, and formatting rules in the call for papers.