What Is a Data Retention Policy for AI Training Data?

Footage of people is personal data, and a dataset can sit in a vendor’s storage long after it ships. Security reviewers and legal teams therefore ask what is kept, where, and for how long, and a vague answer can stall a pilot at contract review. A retention policy also tells you whether a participant’s deletion request can still reach a file once it has been delivered.

The site states rules for discarding and deleting data, not retention periods. Clips that cannot be reliably de-identified are discarded rather than shipped, and the participant release lets a participant request deletion of any recording not yet delivered. The /privacy and /dpa pages are drafts, placeholder language not yet reviewed by counsel, and the signed agreement controls.

A data retention policy answers three questions about a category of data: what is kept, for how long, and what happens when that time ends, whether the data is deleted, returned or archived. A retention period is the length of time. A deletion rule is the trigger, such as a request from a participant or the end of a contract. The two are related but not the same, and a policy can state one without the other.

For AI training data the categories differ in kind. Raw recordings, de-identified delivered datasets, participant and contact details, and website analytics carry different risks and are held for different reasons, so a single retention line rarely covers them all. This page is general information, not legal advice; your counsel decides what applies to your use.

Four groups matter. Raw footage is what the rig records before de-identification. Delivered data is the dataset the buyer receives, which the site says ships with faces and license plates blurred and audio scrubbed. Personal data is what identifies participants and prospects, such as contact details. Operational data includes analytics and the consent records the data card references per episode. The first table sets each group against what the site states.

Do not assume. A missing retention period is not a short one, and it is not a long one either. Wherever the site is silent, the table says Not stated.

Three rules appear on the site. First, the privacy page says clips that cannot be reliably de-identified are discarded rather than shipped, and the provenance page describes the consent chain as discarding footage at any broken checkpoint instead of downgrading it. Second, the participant release, shown verbatim on /provenance, says a participant may withdraw consent for future sessions and may request deletion of any recording not yet delivered to a licensee, and that recordings already delivered cannot be recalled from a licensee. Third, the anonymization service pages say the customer chooses, in writing, whether Firsthand returns only the anonymized output and securely deletes the source, or returns both and retains nothing.

The third rule describes footage a customer sends for anonymization, not participant footage collected to a buyer’s spec. The site does not say how it applies to custom collection. The second table lists each trigger and who acts.

Not stated: how long raw footage is held before or after de-identification, how long Firsthand keeps a delivered dataset or any backups, how quickly a deletion request is carried out, and whether deletion or return follows the end of an engagement. The privacy page also says analytics providers and retention periods will be listed after review, so cookie and analytics retention is not stated. The /dpa page says a sub-processor list for storage and delivery is available on request, and it states no deletion or return terms and no retention periods.

Ask for these in writing before the order form is signed. The signed agreement controls over any summary on the site, including this page.

The release says recordings already delivered cannot be recalled from a licensee. The licence terms point the same way: the site states a perpetual, irrevocable, worldwide grant, buyer-owned data and non-exclusive by default. Retention on Firsthand’s side and retention on the buyer’s side are therefore separate questions, and the second is governed by your own policy and the signed agreement.

Plan your own retention rules for delivered data. The data card lists consent artefact IDs and jurisdictions of capture per episode, which gives you a reference to key your own records to.

Put retention in the spec, not the aftermath. Where coverage is deep, a first validated batch lands in a median of 48 hours from signed spec; thinner environments typically take three to four weeks. Retention questions raised after that first batch compete with delivery for attention, so ask legal to read /privacy, /dpa and /provenance while the skill spec is written, and record the answers in the order form.

What is a data retention policy for AI training data?

A statement of what data a provider keeps, for how long, and when it is deleted or returned. For training data that shows people, it covers raw footage, delivered datasets and personal data. This is not legal advice.

Does Firsthand state a retention period for footage or datasets?

No. The site states discard and deletion-request rules but no retention periods. Raw footage, delivered dataset and backup retention are not stated, and the signed agreement controls.

Can a participant have a recording deleted?

The participant release lets a participant request deletion of any recording not yet delivered to a licensee. Recordings already delivered cannot be recalled from a licensee.

What happens to footage that cannot be de-identified?

The privacy page says clips that cannot be reliably de-identified are discarded rather than shipped, and the provenance page describes a consent chain that discards footage at any broken checkpoint.

Are the Firsthand privacy policy and DPA final?

No. The /privacy and /dpa pages are marked as drafts, placeholder language not yet reviewed by counsel. The signed agreement controls over anything on the website.

What should a buyer ask a data vendor about retention?

How long raw footage is kept before and after de-identification, whether deletion or return follows the end of the engagement, how fast deletion requests are carried out, and how backups are handled. Get the answers in writing before the order form is signed.