Language-data coalitions need a provenance ledger, not just more speech

Associated Press reported that the Gates Foundation has convened 60 organisations, including AI developers, companies and philanthropies, to coordinate work on languages that are underrepresented in AI systems. The coalition says it wants to reach more than 3 billion people over five years. Governance details are still being defined, and a secretariat is expected to track commitments.

The number is a reach ambition, not evidence that a model works for 3 billion people. A language-data programme has at least four distinct stages: collecting speech or text, establishing lawful and community-supported rights, documenting representation and quality, and testing whether deployed systems perform safely in a particular task. Progress at one stage does not establish progress at the next.

Track the data lineage

Project Vaani illustrates both the opportunity and the measurement burden. Its published site describes a goal of more than 150,000 hours of audio from about 1 million people across all 773 districts in India. It currently reports roughly 31,000 hours and describes intended diversity across language, region, education, urban-rural setting, age and gender. Its research paper documents multi-stage automated and manual quality checks for the released subset.

For each dataset, record who collected it, the consent basis, licence, permitted uses, compensation or community benefit, collection geography, speaker attributes, transcription method, quality checks and withdrawal route. Keep those records linked to every model and evaluation that uses the data. A pooled coalition without this lineage can increase volume while making accountability harder.

Separate representation from performance

Coverage should be reported by language and dialect, not only total hours. Then test the deployed task: medical triage is different from classroom tutoring, agricultural advice or customer support. Measure recognition error, meaning preservation, unsafe advice, abstention, appeal and performance for small subgroups. Include locally defined failure cases and independent reviewers who speak the relevant varieties.

Publish a coalition scorecard

The secretariat should report a small set of denominators for every commitment: languages proposed, communities consulted, datasets accepted, hours released under usable terms, model evaluations completed and deployments monitored. It should also show attrition between stages. A dataset may be collected but withheld because consent, transcription quality or licensing fails; hiding that loss would turn an operational problem into an inflated coverage claim.

Governance needs decision rights as well as reporting. Name who can approve reuse, challenge a label, restrict a sensitive application, request correction and withdraw future access. Record whether a community representative, dataset custodian, model developer or deployer owns each decision. Funding totals and partner counts cannot substitute for those assignments.

The counterargument is that detailed documentation can slow urgently needed inclusion. That trade-off is real, especially for small organisations. The response is a shared minimum record and reusable templates, not no record. Proportionate stewardship makes a coalition more scalable because partners can compare assets without renegotiating basic facts each time.

More representative data is a necessary input, not a completed outcome. The useful operating artifact is a provenance ledger that connects community terms to dataset releases, model versions and task-level evaluations. Use the Skills Intelligence methodology to keep coalition commitments, measured coverage and deployment decisions separate.