Case study · Data & governance · AI/ML

Making sensitive child audio usable, legally and ethically

Turned legal and ethical uncertainty into working processes: a 28-country data-protection handbook, risk assessments, an ethics board, and a licensing approach for releasing a model trained on children's audio.

My role
Led data-protection practice and model governance; co-authored a 28-country handbook
When
2021–2024
Read
3 min read
Diagram of the data lifecycle (collection, annotation, storage, model training, release and reuse) with the safeguard used at each step
The safeguard used at each step of working with children's recordings.

The problem

Our recordings capture young children and their families at home, including bystanders who never signed anything, in many countries. Each country has different rules, and most researchers don't have the legal support to read and interpret local legislation. In addition, using these data for AI modelling raised new questions: could releasing the model leak information about the families, and who should be allowed to use it? Research teams needed practical answers, not obscure legal advice.

Constraints

  • GDPR plus national legislation in every country where data was, or could be, collected
  • Community data sovereignty: some data came with requirements set by Aboriginal leadership
  • Multiple organisations had oversight but not the specific expertise (e.g., European Research Council ethics review, CNRS)
  • Safeguards had to protect families without making the data useless for science

My approach

I worked with specialists across law and ethics to design data-management plans and Data Protection Impact Assessments for research with children and caregivers, then turned them into day-to-day workflows across the data lifecycle: collection, annotation, storage, analysis and reuse. To give teams practical answers rather than legal texts, I co-authored a country-by-country handbook of the relevant legislation. During the ERC ethics review, I translated between ethics reviewers and technical staff and set up an ad hoc ethics board. For BabyHuBERT, I designed the governance, including a licensing approach that limits how a model trained on children's audio can be used.

Results

  • 28 countries, incl. 6 in continental Europe
    Legislation reviewed
  • Data from families in 25+ countries
    Programme scope

These were the first such practices for sensitive child-centred audio at CNRS, and the handbook is now public guidance for teams worldwide. I have since advised ethics boards and international scientific archives on responsible reuse of this kind of data.

Trade-offs

Release BabyHuBERT under a licence with conditions, rather than fully open or not at all

Why

Fully open weights would maximise reuse but give no recourse against misuse of a model trained on children's voices. Not releasing it would waste a public investment and slow down the field. A licence keeps the model useful while setting clear limits.

Alternatives considered
  • Fully open release
  • Keep the model internal
  • Share only through individual collaborations

Write country-by-country practical guidance instead of applying one global policy

Why

Applying the strictest regime everywhere blocks legitimate work, while leaving each team to work it out alone produces inconsistent and risky practice.

Alternatives considered
  • Apply GDPR-level rules everywhere
  • Leave compliance to each local team

Let community requirements shape access, not just the law

Why

Legal compliance is the floor. Trust from the communities who share their data is what makes the work possible in the long run.

What I learned

  • Governance moves faster when you sit between the lawyers/ethicists and the engineers and translate in both directions.
  • Writing the risk assessment early changes technical design choices; it should not be paperwork done at the end.

Skills used

  • GDPR & DPIAs
  • AI model governance
  • Licensing
  • Research ethics
  • Stakeholder management
  • Policy writing
All work