Practitioner-written tasks
Who: people who do the job today or did it recently, recruited and paid by Datarena, each under a written consent and a copyright assignment signed before any work.
What is in a task: what was asked, the starting state, the steps, the end state and the check, written from memory under a made-up company name.
What is not: client names, client files, personal data of customers or colleagues, secrets, or an employer’s documents. Each practitioner confirms this for every task, and a second practitioner reviews it.
The practitioner’s own personal data, their name, contact and payment details, is kept only to pay them and to show the chain of rights, and is never in the task itself.
Speech and call recordings
We treat every voice recording as sensitive personal data, and handle it under the laws that apply, including Malaysia’s Personal Data Protection Act and the UK GDPR. This section describes our practice; the terms for each set are agreed in writing in its scope and licence.
- We record or license a call only when every speaker in it is an adult who agreed in writing that it may be used to train AI, or when the recording’s owner holds that basis and licenses the recording to us in writing.
- Names, identity and account numbers, addresses, payment details and other identifiers are removed from transcripts and silenced in the audio before anything reaches a buyer. The redaction spans are listed with each set. Where that work is done, by the owner or by us, is agreed in the scope.
- Each recording carries a rights note stating its source, the basis on which it may be used, and the licence.
- Anyone we have recorded can withdraw, as set out under Withdrawing consent.
What a buyer can license is described under contact-centre speech.
Earlier research: voice evaluation, October 2026
This note predates our current contact-centre speech work. In October 2026 we designed a voice evaluation that would record only the callers, typers and raters we recruit, each recruited, paid and consented individually by Datarena under a written consent form before any session. No session was recorded under it. It would have recorded:
- Our callers’ side of each test session with a platform’s staging agent or released model, by phone, typed, or through its API.
- The agent’s side of the same session, for rating only. We never publish agent-side audio without the platform’s written permission.
- Training demonstrations and reference clips, newly performed by our speakers for that purpose.
- Our raters’ labels and notes on each session.
Storage and access
Tasks, records, recordings and ratings are stored in access-controlled cloud storage, encrypted at rest. Access is limited to the founder and to the practitioners and raters assigned to a piece of work, for as long as they are working on it. Material is kept for the purpose and period set out in each consent or licence and then deleted, unless the person or company has agreed in writing to a longer period.
A practitioner or rater sees only the work assigned to them. Records are de-identified by their owner before they reach our storage, or reviewed with the owner inside their environment. For call recordings, where redaction is done is agreed in the scope of each set, as set out under Speech and call recordings.
Withdrawing consent
Anyone we have recorded can withdraw consent at any time by writing to support@datarena.ai. From that day we stop using their recordings: no new sessions, reports or datasets include them. Within 30 days we delete them from our systems and confirm in writing when it is done.
A practitioner who has written tasks can withdraw in the same way. Accepted tasks carry no personal data and were assigned to Datarena when they were accepted, so they remain licensed. Withdrawal stops the use of their recordings and of any draft not yet accepted, and we delete their name, contact and payment details, keeping only the assignment record that shows the chain of rights.
Withdrawal cannot recall copies delivered before it, inside a finished evaluation report or a licensed dataset, or undo a model already trained on them.
Changes and questions
When this page changes, the date at the top changes with it. If a change affects what a practitioner, recorded person or licensing company agreed to, we tell them in writing before it applies to their material.
Questions about any of this, or a request to see, correct or delete your details, go to support@datarena.ai.