Intellectual Property, Provenance and Output Ownership
Who owns what goes in, and what comes out?
Baseline: 3Enhanced: 2Advanced: 27 controls
Who it applies to
Enter for every AI system. Some of its controls apply only where you train or adapt the model (B1). Go as deep as your system letter, and implement in your mode.
Why it exists
Settle who owns what before the first valuable output exists — because AI creates two intellectual-property exposures that ordinary software never did, one on the way in and one on the way out. On the way in, models and the document collections they consult are built from material whose licence and origin are often unknown; whatever legal risk that material carries flows silently into every answer built on it, and courts in several jurisdictions are still testing whether training on copyrighted material is lawful at all. Do not let your organisation become the test case by accident. On the way out, the ownership of what AI generates for you — code, text, designs, improvements — is very often defined nowhere: not in your policy, not in the provider’s terms. Silence on ownership is the single worst position to occupy in a later dispute, and it is the default position of almost every adopter.
What to actually do
Track the origin and licence of everything the model ingests — including the documents it consults to answer. Licensing the model is not enough. The datasets used to train or adapt it, and the document collections your systems draw on when generating answers, each carry their own licence and copyright status — and a source you had no right to use contaminates every output built on it. Record, for each source: where it came from, what its licence permits, and who approved its use. Connect these records to the bill of materials in the Inventory dimension, so that when a dataset or source is later found to be unlawful or withdrawn, the question “which of our systems are affected?” has an answer measured in hours.
Define who owns the outputs before your people generate anything of value. State, in your own policy and in every provider contract, who owns the code, content, and improvements that AI generates for you, and under what terms you may use, modify, and redistribute them. Do not assume the answer is favourable: provider terms differ, and some jurisdictions do not grant ownership over machine-generated work at all. The moment to fix this is before the first valuable output exists — afterwards, you are negotiating over something that already matters.
Keep your own secrets out of other people’s models. An employee pasting proprietary code, unpublished figures, or a confidential contract into an external AI tool may be handing that material to another company — possibly permanently, if the provider trains on what it receives. Treat this as an intellectual-property leak, not a minor data-handling slip: forbid it in policy for unapproved tools, provide approved tools whose contracts prevent training on your data (the clause you secured in the Sourcing dimension), and back the rule with the technical controls that catch confidential material on its way out.
Check valuable outputs for copied material before they ship. AI systems sometimes reproduce, verbatim or nearly so, licensed or copyrighted material they encountered in training. Where the output feeds your products — code that ships, text that publishes, designs that sell — screen it for reproduced material rather than assuming originality. The screening is cheap; discovering the reproduction in a competitor’s infringement claim is not.
Turn the prohibition into a one-page matrix. Cross the classes of information you already hold — public, internal, confidential, personal data — with the AI tools you allow, and state for each pair whether the tool may be used, used with conditions, or not at all. Published with the acceptable-use policy, it answers the question staff actually ask: can I paste this here?
Controls
Baseline3 controls
Control 61
Record the origin, licence and lawful basis of every dataset you train or fine-tune on.
BaselinePreventivedeveloperper system
What it is
Before training, not after; a dataset without a recorded origin is not used.
How to implement it · Minimum content
Dataset name, version and description of its content
Origin: who created it, where and when it was obtained
Licence and what it permits (commercial use, modification, training)
Personal data present or not, and the lawful basis for using it
Approver and date of approval
Systems and model versions trained on it
Rule: a dataset without a completed record is not used for training
How to start
Before the next training or fine-tuning run, complete the record for each dataset and do not start the run until it is complete.
Evidence
The dataset records, dated before the training runs that used them.
Mode: light / standard / full
light
The developer keeps one record per dataset in a spreadsheet next to the project.
standard
The data or engineering lead approves each record and stores it with the bill of materials.
full
A data governance function holds a dataset catalogue and the training pipeline refuses unregistered datasets.
Control 62
Adopt an output-ownership and licensing policy for AI-generated work.
BaselinePreventiveall rolesper system
What it is
Covering what your people may generate, who owns it, and under what terms it may be used — before the first valuable output.
How to implement it · Minimum structure
Scope: code, text, images, designs and analyses generated with AI by staff and contractors
Ownership: work generated for the organisation belongs to it, as employment and contractor terms allow
What may be generated freely and what needs approval, e.g. published content or shipped code
Provider terms: what each approved tool's terms say about output ownership and use restrictions
Where ownership of machine-generated work may not be recognised, record the human contribution to valuable outputs
Owner of the policy; review when approved tools change
How to start
Read the output terms of each approved AI tool and write a one-page policy stating who owns what and any restrictions.
Evidence
The approved policy and the provider terms it relied on.
Mode: light / standard / full
light
The owner-manager writes it, with a one-off check by an external adviser if outputs are sold.
standard
Legal drafts it with the AI lead and the committee approves it.
full
Legal maintains it with intellectual-property specialists.
Control 63
Forbid placing confidential or proprietary material in unapproved AI tools, and provide approved alternatives.
BaselinePreventiveall rolesper system
What it is
The rule and the alternative arrive together, or the rule is ignored.
How to implement it · Minimum structure
Rule in the acceptable-use policy: no confidential, proprietary or personal material in unapproved AI tools
One-page matrix: information classes (public, internal, confidential, personal data) against allowed tools: may use, use with conditions, may not use
At least one approved alternative, whose contract excludes training on your data, available before the rule takes effect
Communicated with the secure-use guide and acknowledged by staff
Backed by the technical controls that catch confidential material on its way out, where in place
How to start
Publish the matrix together with access to at least one approved tool, then announce the rule.
Evidence
The rule, the matrix, the approved-tools list and staff acknowledgements.
Mode: light / standard / full
light
The owner-manager approves one business AI tool, sends the matrix by email and collects acknowledgements.
standard
IT provides the approved tools and the policy owner publishes the matrix on the intranet.
full
The matrix is tied to the data-classification scheme and enforced by data-loss prevention on every channel.
Enhanced2 controls
Control 64
Track origin and licence for every training source and consulted collection.
EnhancedPreventiveall rolesper system
What it is
Connected to the bill of materials, so contamination questions are answerable.
How to implement it · Minimum content
Source name, version and type: training dataset, adaptation set, or document collection consulted when answering
Origin and date acquired
Licence or copyright status and permitted uses
Approver and date
Systems using it, linked to their bill of materials
Status: active, withdrawn or disputed, with date
Test: from any source, list the affected systems within hours
How to start
Extend each system's bill of materials to list its sources, starting with the document collections your assistants consult.
Evidence
The source register linked to the bills of materials, and a test query listing the systems affected by one source.
Mode: light / standard / full
light
A named person keeps a sources tab in the inventory spreadsheet, linked by system name.
standard
The engineering lead keeps the register with the bill of materials and approves new sources.
full
A data catalogue tracks sources and lineage automatically and answers affected-system queries on demand.
Control 65
Fix output ownership in every provider contract.
EnhancedPreventiveall rolesper system
What it is
Your terms, not the provider’s defaults, decide who owns what the service produces for you.
How to implement it · Minimum structure
Ownership of outputs assigned to the organisation, or a clear licence covering all intended uses
Your rights to use, modify, redistribute and commercialise outputs
Provider rights: no reuse of your outputs beyond delivering the service
Support or indemnity if an output infringes third-party rights, where negotiable
Your terms take precedence over the provider's standard terms
How to start
Review current AI contracts for an ownership clause, list those relying on provider defaults, and fix them at renewal.
Evidence
The ownership clause in each AI provider contract, and the list of contracts still pending.
Mode: light / standard / full
light
Choose providers whose business terms already assign outputs to the customer, and keep that page on file.
standard
Procurement and legal insert a standard ownership clause in every AI contract.
full
Legal enforces the clause as a mandatory minimum, with deviations approved by risk.
Advanced2 controls
Control 66
Screen product-bound outputs for reproduced material.
AdvancedPreventiveall rolesper system
What it is
Applied wherever AI-generated work ships in your products or publications.
How to implement it · Minimum rules
Scope: AI-generated code that ships, text that is published, images and designs that are sold
Code: enable the coding assistant's filter for matches with public code; scan releases and check the licence of any match
Text: run a similarity check against published sources before publication
Images and designs: run a reverse image search for close matches
On a match: rewrite, license and attribute, or remove; record the decision
Record what was screened, how, and the result
How to start
Turn on public-code match filtering in coding assistants and add a similarity check to the publication checklist.
Evidence
Screening records for shipped outputs and the decisions taken on any match.
Mode: light / standard / full
light
The author runs a free similarity check before publishing and notes the result in the release checklist.
standard
Engineering and communications build screening into code review and editorial approval.
full
Automated scanning in the build pipeline and the publishing workflow, with matches routed to legal.
Control 67
Publish a training-content summary and a copyright-compliance policy where you provide a general-purpose model.
AdvancedPreventivedeveloperthe whole organisation
What it is
What those who build on your model, and those whose work it was trained on, need to know about what went into it.
How to implement it · Minimum structure
Applicability: only if you place a general-purpose model at others' disposal
Summary: main data sources by category (public web, licensed, user, synthetic), main datasets, period covered, languages
Collection practice: how content was gathered and how rights holders' machine-readable opt-outs were honoured
Copyright policy: how rights reservations are identified and respected, how licensed content is handled, contact for complaints
Published on your website and updated with each new model version
How to start
Confirm whether you provide a general-purpose model to others; if so, compile the source list from your dataset records and draft the summary.
Evidence
The published summary and policy, dated and tied to the model versions they cover.
Mode: light / standard / full
light
A small provider writes both documents from its dataset records, with a one-off review by an external adviser.
standard
The model team drafts them and legal approves before each release.
full
Legal and the data governance function maintain them as release artefacts, generated from the dataset catalogue.
Evidence
The provenance record for each training dataset; The ownership policy; the origin and licence register; provider contract clauses on ownership and training; screening records for shipped outputs; the confidentiality rule and its enforcement evidence. The published training-content summary and copyright policy, where applicable.