Short answer
The business holding an SOP often does not own all of it. Procedures written by employees may belong to the employer, but text copied from equipment manuals, consultant deliverables, vendor courseware and franchisor standards often belongs to someone else or arrives under license limits. Before licensing a procedure corpus for AI training, establish authorship per document, exclude or clear third-party content, and get express permission for each intended use.
This is not legal advice. The page sets out diligence questions and clause topics to work through with your counsel and the licensor.
Why a procedure library is a patchwork
A site's document-control system looks like one owner's collection. In practice it mixes material from many hands: SOPs drafted by the site's engineers, maintenance steps pasted from an equipment maker's manual, a quality manual written by a consultant, courses licensed from a training vendor, photos supplied by a parts distributor, and passages quoted from paid industry standards. Each source can carry a different owner and different limits.
Copyright is only one layer. Confidentiality clauses in OEM, supplier and customer agreements, franchise agreements and software-platform terms can restrict sharing even where the business wrote the text itself. A clean license has to clear all of them.
The legal debate about AI training on copyrighted material is unsettled. The US Copyright Office's report on generative AI training identifies the points in development where copying may implicate rights, analyzes fair use and examines licensing options (U.S. Copyright Office, 2025). A license that names each intended use, granted by whoever actually holds the rights, reduces how much a robotics program depends on how that debate is resolved.
Who may hold rights in procedure content
| Content source | Questions to resolve | Evidence to request |
|---|---|---|
| Written by the supplier's employees | Does any policy, agreement or collective arrangement reserve rights to authors? | Document-control author fields; authoring policy |
| Written by consultants or contractors | Did the contract assign rights to the supplier, or only grant a license? | The engagement contract's IP clause |
| Copied or adapted from equipment manuals | Does the equipment maker's manual license permit reuse, let alone AI training? | Manual terms; written permission; passage list |
| Vendor courseware in an LMS | Is the supplier a seat licensee with no right to sublicense? | Vendor contract |
| Franchisor operating manuals | Does the franchise agreement restrict disclosure or reuse? | Franchise agreement confidentiality and IP terms |
| Customer specifications and work instructions | Do supply or quality agreements make these the customer's property or confidential? | Customer agreement clauses |
| Quoted industry standards | Are paid standards reproduced at length inside procedures? | List of quoted standards and extent |
| Openly licensed material | Do NonCommercial or NoDerivatives terms apply? | License text per item |
| Embedded photos and diagrams | Who took or drew them: staff, vendors or photographers? | Image source list |
Open licenses need reading, not assuming. Under Creative Commons terms, NonCommercial is judged by the use rather than the user, NoDerivatives bars sharing adaptations publicly, and licenses with either term are considered non-free (Creative Commons). Labels can also be wrong: an audit of dataset licensing found that after re-annotation, 66% of analyzed Hugging Face licenses fell in a different use category, often labeled more permissive than the author's own license (Longpre et al., 2023). Ask for the underlying terms, not the tag.
Two neighbors cover specific holders in depth: OEM service documentation access for equipment-maker material, and LMS training courses for vendor courseware. Multi-site operating standards are covered on franchise operations manuals.
Approval trails are not ownership records
Regulated industries keep strong document control. Under US drug manufacturing rules, written production procedures are drafted, reviewed and approved by the appropriate organizational units and approved by the quality control unit (FDA / eCFR). That approval history is useful evidence of internal authorship, revision and currency. It does not show whether a step was pasted in from a supplier's manual, and approval by the quality unit does not transfer anyone else's rights.
Uses the license must name
Procedure data is used in more ways than training, and some uses reproduce the text.
- Training and fine-tuning of planners, instruction-following and procedural-reasoning models.
- Retrieval at runtime. A planner that retrieves procedure chunks stores the text in an index and may display or speak it to operators. This is reproduction at deployment, not just training; name it explicitly.
- Derived structure. Extracted step records, task graphs, plan exemplars, translations and embeddings: who owns them, and do they survive termination?
- Evaluation. Procedures held out as test cases, and whether benchmark results or example items may be published.
- Collection scripts. Using procedures to script commissioned demonstration recording shares them with demonstrators and collection providers.
- Recurring updates. Whether new revisions are delivered, and whether superseded text must be purged from indexes.
General robotics clause topics, including weights, field of use and deletion, are set out in licensing terms for robotics training data; this page covers what changes for documents.
Illustrative rights manifest for a procedure delivery
# Illustrative only — synthetic documents and parties; not a legal template
delivery: proc_corpus_supplier_S07_v1
documents:
- doc_id: SOP-PKG-0112 # case packer changeover, rev 9
authorship: employee (process engineering, site_03)
third_party_content: none identified
status: included
- doc_id: WI-MNT-0457 # filler lubrication
authorship: employee
third_party_content: steps 4-11 adapted from OEM manual OEM_M22, sec. 6
status: included_with_exclusion # steps 4-11 removed pending OEM permission
- doc_id: QM-001 # quality manual
authorship: consultant_C2 (2019 engagement)
third_party_content: assignment clause not located
status: excluded
- doc_id: LMS-FORK-03 # forklift course
authorship: training vendor V4
status: excluded # seat license only
redaction_log: redactions/S07_v1.csv # parameter, customer and supplier names
permitted_uses: [training, fine_tuning, runtime_retrieval_internal, evaluation]
derived_data: licensee owns extracted steps, task graphs and embeddings
The exclusion lines matter as much as the inclusions. A corpus that arrives with a documented exclusion list is easier to defend than one that claims to be clean.
Delivery mechanics
- Document inventory with ID, title, revision, effective date, owning department and author category for every file.
- Third-party content register listing passages, images and courses from outside sources, and their status.
- Exclusion list and confirmation that excluded files and passages are absent from the delivery.
- Redaction log showing what was removed for confidentiality, using the approach in redacting SOPs. Redaction protects confidential details; it does not change who owns the remaining text.
- Originals and extracts delivered together, so extracted steps can be traced back to a page and revision.
- Hashes and a manifest for acceptance and later deletion.
Limitations
The manifest is only as reliable as the supplier's records. Older documents often lack author fields, and pasted passages rarely carry attribution, so third-party content can survive a review. Similarity checks against known manuals help but will not catch everything. The supplier's warranty and indemnity terms decide who bears that residual risk; see warranty of title.
Questions to ask suppliers
- Who authors procedures at each site, and does document control record the author or department?
- Which documents were written or revised by consultants, and do their contracts assign rights to you?
- Which procedures copy or adapt equipment-maker manuals, and do you hold written permission to reuse them?
- Which training courses are licensed from vendors, and do those licenses allow sublicensing?
- Do any franchise, customer or supplier agreements restrict disclosure of these procedures?
- Are paid industry standards quoted at length inside the documents?
- Who supplied the photos and diagrams embedded in procedures?
- Can you deliver an exclusion list and warrant title for everything else?
- Will you permit runtime retrieval and display of procedure text, not only training?
These ownership checks sit alongside quality checks in the SOP corpus evaluation checklist.
Discuss your requirements
SourceX checks the data and the supplier's licensing permissions, agrees pricing and allowed uses in a license, and coordinates delivery; nothing is contracted until a supplier agrees. Tell SourceX which procedures you need and how you will use them, including runtime retrieval. It reviews the request and follows up about potential sources; a request does not guarantee a match. See also the procedures cluster, the humanoid training data guide, the page on licensing SOPs and playbooks and the guide to license terms.
Explore sourcing and licensing options with SourceX
Sources
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — where AI training may implicate rights, fair use and licensing options.
- About CC Licenses — NonCommercial and NoDerivatives terms.
- 21 CFR 211.100 Written procedures; deviations — drafting, review and approval of regulated procedures.
- The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI — license labels that differ from the underlying terms.