onTerms AI/ML Services Module v1.0 — Standard Terms (England & Wales)

onterms:ai:1.0.0:EW · Status: v0.9 DRAFT — Not legal advice.

For providers of AI/ML products and services (deeper than the SaaS module's AI knobs). Immutable adopt-verbatim body; typed Elections; composes with CORE, onDPA (onterms:dpa:1.0.0:EW) and the dispute module. Capitalised terms from the Dictionary.

0. B2B-only

Registry-verified RegistryAttestation at enrolment (spec/verifier.md §V2).

1. Training on customer data (default OFF)

1.1 The Provider will not use Customer Data, AI Input or AI Output to Train, fine-tune or improve any model, except a model dedicated to the Customer's own instance, unless the Customer opts in. Aggregated, de-identified Usage Data may be used per the Dictionary. Opt-in is an explicit, revocable Election and (per the order schema) requires this AI module to be incorporated.

• ai.training.onCustomerData · enum {off, opt_in} · off · ai.training.revocable · bool · true

2. Input/output ownership & licence

As between the parties, AI Output generated from the Customer's AI Input is the Customer's (treated as Customer Data); the Customer is responsible for its use; the Provider reserves that the same or similar output may be generated for others (no uniqueness/exclusivity). The Provider retains its models, weights and platform IP.

• ai.output.ownership · enum {customer, provider} · customer

3. Accuracy disclaimer (no warranty of correctness)

AI Output is probabilistic and may be inaccurate, incomplete or unsuitable; the Provider does not warrant that AI Output is accurate, error-free, non-infringing or fit for a particular purpose, and the Customer is responsible for reviewing Output before relying on it. This does not exclude liability that cannot lawfully be excluded.

4. IP indemnity for AI output (carved out by default)

4.1 AI Output is excluded from the IP infringement indemnity by default (carved_out). 4.2 A conditional shield is available by Election: the Provider defends third-party IP claims that AI Output infringes only where the Customer used unmodified, generally-available features, did not disable filters/guardrails, and used Output as documented — excluding claims from the Customer's inputs, modifications, or guardrail circumvention.

• ai.output.ipIndemnity · enum {carved_out, conditional_shield} · carved_out

5. Responsible & acceptable use

The Customer will not use the Services for prohibited uses (unlawful, harmful, deceptive, rights-infringing, or uses that produce illegal content), will keep meaningful human oversight where required, and will not present AI Output as human-generated where law requires disclosure. The Provider may suspend for serious or repeated breach.

• ai.use.humanOversightRequired · bool · true

6. EU AI Act role allocation (applies only where the Act applies) — [fix X-5]

6.1 Default: the parties do not assume EU AI Act applicability. The EU AI Act binds by territorial reach (output used in the EU); for an England & Wales-governed deal that is frequently not the case. The applicability flag therefore defaults to "assess", not "EU deployer = true". 6.2 Where the Act does apply, the Order allocates roles (provider/deployer), the Provider supplies the technical documentation, instructions for use and transparency information needed for the Customer's deployer duties, and the parties cooperate on logging, human oversight, and serious-incident handling; a change in how the Customer deploys that flips a party's role (Art. 25) is notified. 6.3 Art. 50 transparency (e.g. labelling AI interactions / synthetic media) is honoured where the Act applies.

• ai.euAiAct.applicability · enum {assess, applies_provider, applies_deployer, not_applicable} · assess (not defaulted ON)
• ai.euAiAct.systemRiskTier · enum {minimal, limited_transparency, high_risk, assess} · assess

7. Model & version changes; evaluation; transparency

7.1 The Provider may update models/versions provided it does not materially degrade the contracted capability or security during the term, with reasonable notice of material changes and a deprecation window for a named model where the Order pins one. 7.2 The Customer may evaluate/benchmark the Services for its own internal assessment. 7.3 The Provider makes available reasonable model documentation (e.g. a model card: intended use, known limitations, evaluation summary).

• ai.model.pinNamedModelEnabled · bool · false · ai.model.deprecationNoticeDays · int 30–180 · 90

8. Data protection & special-category data

Any Processing of personal data is governed by onDPA; the Customer will not submit special-category or other high-risk Personal Data to general features unless the Order/onDPA expressly provides for it.

9. Term, termination & liability (CORE)

Per CORE, with the AI-output carve-outs above. General cap and super-cap per CORE (1×-fees cap with a one-directional floor, 3× data/security super-cap, uncapped death/PI, fraud, IP indemnity, confidentiality, payment); breach of the AI-training restriction (clause 1) is treated as a confidentiality-grade matter. Disputes per the dispute module; governing law England & Wales default.

Elections (consolidated)

Hard-coded E&W overlays (not Elections)

Accuracy/no-correctness disclaimer cannot exclude non-excludable liability; EU AI Act duties are not defaulted ON for E&W deals (assess applicability); CRTPA 1999 excluded; death/PI + fraud never excluded; UCTA-reasonable mirrored caps with a one-directional floor; uncapped carve-outs; governing law E&W default; composes with CORE/onDPA/dispute.