Derived Records in Training Data: Tracing Rights in Transcripts, Translations and Summaries to the Source

This page covers inheritance from human-created source records. Model-generated corpora with no human source are covered in provenance records for synthetic training data. For the wider framework of source, rights and permitted use, start at the data provenance hub.

Why derived records inherit the source's rights

Rights in the source record usually set the ceiling for every record derived from it. The Copyright Act defines a derivative work as one based on preexisting works. Its examples include translations, abridgments and sound recordings, and it covers works made of "editorial revisions, annotations, elaborations, or other modifications" when the result is original [1]. Copyright in a derivative reaches only the new material and never cleans up the underlying work, so a licensed translation of an unlicensed contract is still an unlicensed contract.

Non-copyright restrictions travel too, and in practice they bind more often. Privacy obligations, confidentiality clauses in customer agreements, recording-consent rules and sector rules attach to the information itself. Those obligations survive transcription, translation and summarization. A summary of a support call that keeps the customer's account problem and city is still derived from that customer's personal data.

Derived corpora also mix parents with different terms. Research on dataset documentation notes that one dataset may be built from multiple sources, each under a different license [6]. The Data Provenance Initiative audited more than 1,800 text datasets and reported license omission rates above 70% and error rates above 50% on popular hosting sites [7]. Derivation is where those errors compound. For multi-source cases, see tracing upstream licenses in aggregated and derived datasets.

Who owns transcripts of recorded calls

Ownership of a call transcript usually depends on three things: who spoke, who recorded, and who transcribed. The recording party typically holds the audio file. Unscripted business conversation is often thin as copyrightable expression, so the binding constraints tend to be contract and privacy, not copyright. For scripted content such as sales scripts, training material or recorded presentations, the script author's rights carry through to the transcript (see copyright in operational business records).

Consent to record is not consent to license. California Penal Code section 632 requires all-party consent to record a confidential communication [8]. A compliant "this call may be recorded for quality and training purposes" notice answers that question for the recording. It does not by itself establish that the transcript may be licensed to a third party for model training. Check what the notice and the customer contract actually said; consent and notice records lists what to request.

Channel layout matters for attribution. Telephony platforms such as Twilio record either mono or dual channel, and dual channel keeps each call leg on its own channel [9]. With dual-channel audio, a transcript can attribute turns to agent and customer reliably. That is what lets you apply different rules to employee speech and customer speech. With mono audio, diarization errors can put customer personal data into turns labeled "agent."

Machine-generated transcript and summary ownership

A machine-generated transcript or summary probably adds no new copyright of its own, but it can add contract terms. The US Copyright Office's January 2025 copyrightability report concluded that AI outputs are protectable only where a human determines sufficient expressive elements, and that prompts alone are not enough [4]. An ASR transcript or LLM summary with no meaningful human editing therefore sits on the source's rights rather than creating a new rights holder. A human-edited summary may carry new authorship that needs its own chain of title.

The tool's terms are the second layer. Speech-to-text, machine translation and LLM APIs each have service terms about output ownership, use restrictions and data retention. Some enterprise agreements also limit using outputs to build competing models. Record the tool, version and the agreement in force when each record was produced. A later change to those terms does not show you which terms applied to outputs produced in 2024.

Training-use questions are separate from ownership. As of October 2026, the Copyright Office's Part 3 report on generative AI training remains a pre-publication version [5]. Treat derivation as a documentation problem rather than a fair-use argument you can rely on.

Translation rights in training data

A translation has two potential rights holders: the author of the source text and the translator. Translations are a named category of derivative work [1]. Copyright vests in the author unless the work is made for hire, in which case the employer or commissioning party is treated as the author [2]. A commissioned translation is one of the enumerated categories that can be a work made for hire, but only if the parties signed a written agreement saying so [1]. Otherwise ownership moves only through a signed transfer [3].

For a translation memory (TMX) or bilingual corpus, check three items. First, whether the client's source documents could be licensed at all. Second, whether the language-service provider's contract assigned the translations to the client or treated them as made for hire. Third, whether segments came from machine translation post-editing, which brings in the MT engine's terms. Agencies often reuse memories across clients, so a single TM can hold segments with different owners. See licensing translation memories for AI training for what a buyer request should cover.

Are annotations derivative works

Most labels are not derivative works, but they are still bound by the source's terms and by the annotation contract. Section 101 counts annotations and elaborations as derivative works only when the result is an original work of authorship [1]. A class label, bounding box or "resolved / not resolved" flag is unlikely to meet that bar. Free-text rationales, rubric explanations and preference justifications written by experts may.

The practical risks sit elsewhere. Annotation vendors' master agreements define who owns deliverables, whether the vendor may reuse your source data, and whether annotators' work product is assigned. Employee-written labels, such as a support lead's QA scores, fall under employment terms and internal policy; see employee-authored records in training data. Labels that encode outcomes from the source system also inherit its personal-data status.

Derived-record rights decision table

Use this table to decide which rights to check for each type of derived record before it enters a post-training mix.

Illustrative example: invented to show structure; it does not describe an available dataset.

A lineage schema for derived records

Every derived record needs a pointer to its parent and a description of the transformation, stored at record level and not only in a dataset card. Datasheets for Datasets frames collection, preprocessing and labeling as questions the dataset creator should answer [11]. Croissant gives a machine-readable format, with a Responsible AI extension, that can carry dataset-level provenance [12]. Neither holds per-record parent IDs on its own, so add them to your records. The rights metadata schema shows how to encode permitted uses alongside.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "record_id": "sum-000184",
  "record_type": "summary",
  "parent_ids": ["ticket-55120", "call-rec-55120-a"],
  "parent_license_ids": ["lic-supplier-07-v2"],
  "transformation": {
    "kind": "llm_summarization",
    "producer": "machine",
    "tool": "summarizer-api",
    "tool_version": "2026-06-01",
    "tool_terms_ref": "terms-snapshot-2026-06",
    "prompt_template_id": "tmpl-ticket-sum-v4",
    "human_edited": false,
    "editor_role": null,
    "produced_at": "2026-07-14T10:02:11Z"
  },
  "deidentification": {
    "method": "ner_replace_v3",
    "applied_before_transform": true,
    "sample_checked": true
  },
  "permitted_uses": ["sft", "eval"],
  "inherits_restrictions_from": ["lic-supplier-07-v2"],
  "status": "active"
}

Four fields carry most of the weight. parent_ids lets you propagate a takedown or license expiry to every child. producer and human_edited decide whether a new author exists. tool_terms_ref freezes the output terms that applied, and applied_before_transform shows whether personal data was removed before an outside service saw it. When a parent is withdrawn, a query on parent_ids should return every transcript, translation, summary and label to delete or quarantine, including records in held-out eval sets. The data lineage glossary entry defines the broader concept.

Privacy rules that survive transformation

Personal-data obligations follow the information, not the file format. A transcript, translation or summary that still identifies a person stays personal data. For health records, HIPAA de-identification requires Safe Harbor removal of 18 identifiers or an Expert Determination [13]. A summary generated from a de-identified note is only as safe as the input. A summary generated from an identified note and then scrubbed needs its own check, because LLMs paraphrase identifiers into forms that pattern-based redaction misses. See reviewing a HIPAA Expert Determination report if health data is in scope.

Order the pipeline deliberately. De-identify before sending text to a third-party transcription, translation or summarization service where you can. Record the order in the lineage record, because reviewers will ask. Summarization datasets built from customer-service dialogs, such as TWEETSUMM, show how closely summaries track the source conversation [10].

Request checklist for derived datasets

Ask these questions of any supplier offering transcripts, translations, summaries or labels:

• What is the parent record for each derived record, and is the parent ID delivered?
• What license or authorization covered the parent, and does it extend to derivatives and to third-party AI training?
• Who produced each derivative (employee, contractor, vendor or machine), and is there a signed assignment or work-made-for-hire agreement where one is needed?
• Which tool and version produced machine outputs, and which output terms applied at that time?
• Was de-identification applied before or after transformation, and was a sample checked?
• For recordings: what notice was given, were all parties' consents obtained where required, and is the audio mono or dual channel?
• Can a parent withdrawal be propagated to every child record?

A supplier that cannot answer the first two questions is not offering a licensable derived dataset. If you would rather describe the records you need and have the parent rights assessed per dataset, see how SourceX works with AI data buyers. The chain-of-title guide lists the documents that back each answer.

Sourcing derived records with source lineage

SourceX sources operational datasets, including support and sales histories and engineering records, from US companies on request, and every dataset is rights-reviewed for ownership and consents before delivery. Each is delivered under a license that defines records, uses, term and delivery, with diligence materials on source, rights, preparation and allowed use prepared per dataset. If you need transcripts, translations or summaries with their source documented, such as meeting transcripts licensed for AI training, describe the data you need at SourceX for AI data buyers.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.