#Responsible AI Use Policy

These are the rules. Everything after this clause is detail.

1.1 A person decides. No output of an artificial-intelligence capability is, or becomes, a decision. Every consequential action (clinical, financial, employment-related or otherwise affecting a person) is taken by a named human being who is accountable for it and who can see what the capability suggested and what it drew on.

1.2 No clinical decision, ever. No capability diagnoses, triages, prescribes, doses, or determines treatment. DIS-GL-028 is the boundary and it is not moved by this Policy, by an Order Form, by a configuration, or by a request from a hospital. A hospital cannot opt into having the Platform make a clinical decision.

1.3 It is always visible that a machine produced it. An output produced or materially shaped by a model is labelled as such at the point of use, with the capability, the model version and the time. Pensieve does not present machine-generated content as though a person wrote it, and no configuration removes the label.

1.4 The hospital's data trains nothing. Customer data is not used to train, fine-tune, evaluate or improve any model, for that hospital or for anybody else. The commitment and its reasoning are in DIS-GL-027 Section 4 and it is contractual in DPA-GL-001. This Policy adds one thing: no Pensieve employee may seek a hospital's permission to relax it as part of a commercial negotiation.

1.5 Everything is auditable. Every invocation of a capability that touches personal data is logged under DIS-GL-013: who, when, on what record, which capability, which model version. A suggestion that influenced a decision can be reconstructed afterwards.

"Human in the loop" is meaningless unless it says who, and when.

2.1 No capability may be configured to act without a person, and Pensieve does not build an "auto-approve" control for any of the above. Where a hospital asks for one, the answer is no, and the reason is this clause.

2.2 Named accountability. For each artificial-intelligence capability it enables, the hospital records the role accountable for its outputs. For clinical-adjacent capability, that is the hospital's clinical governance lead. The record lives with the enablement record under POL-GL-058 5.4.

Binding on every user of the Platform, in addition to ADD-GL-004 and ADD-GL-005.

3.1 Do not rely on an output as a clinical decision, or record one as though a clinician made it.

3.2 Do not disable, hide, crop or paraphrase away the machine-generated label at 1.3 when copying an output into another system or a printed document.

3.3 Do not copy patient data out of the Platform into an external artificial-intelligence tool: a public chatbot, a transcription service, a coding assistant. That is a disclosure of patient data to a third party under the hospital's own controllership, and it is outside everything Pensieve has assessed. Where a hospital wants such a capability, it should ask for it inside the Platform, where it is governed.

3.4 Do not use an output to make a decision about a person's employment, credit, insurance or access to care.

3.5 Do not attempt to make a capability produce content outside its purpose, or to extract training data, model weights or system instructions from it.

3.6 Do not present Platform-generated content to a patient as clinical advice.

4.1 Build a capability that takes a clinical decision, however configured, however requested, however priced.

4.2 Train on customer data (see 1.4).

4.3 Allow a model available to one hospital to see or learn from another hospital's data (DIS-GL-027 Section 4).

4.4 Send personal data to a model provider that reserves a right to train on it. Where a third-party provider is used, the position for each provider is in DIS-GL-027 Section 5 and the provider is a sub-processor listed in DIS-GL-009, notified under POL-GL-055.

4.5 Publish an accuracy, sensitivity or specificity figure Pensieve has not measured, or imply clinical validation that has not happened (DIS-GL-027 Section 9).

4.6 Enable an artificial-intelligence capability in a hospital's tenancy without the hospital enabling it. Every such capability is opt-in and reversible (DIS-GL-027 Section 7).

4.7 Use an artificial-intelligence capability to make a decision about a hospital, a credit decision, a support prioritisation that disadvantages it, a renewal price, without a person deciding. POL-GL-065 3.8 is the related commitment on pricing.

4.8 Use a hospital's data, or a model derived from it, in Pensieve's own research or product development without the ethics review that DIS-GL-027 Section 8 describes and, in most framings, participant consent. Pensieve does not do this today at all.

5.1 A person about whom a decision is made is entitled to know that an artificial-intelligence capability was involved. Because of 1.1, the answer is always that a person decided and a capability assisted; the assistance is nonetheless disclosed.

5.2 Data-protection rights (access, correction, erasure, grievance) are exercised under POL-GL-053 and POL-GL-066. Where the individual is a patient, the hospital is the Data Fiduciary and the request is made to the hospital; Pensieve assists it under DPA-GL-001.

5.3 Where a jurisdiction confers a specific right in relation to automated decision-making, that right is honoured in that jurisdiction. Pensieve's design position, that there is no solely automated decision producing a legal or similarly significant effect, means the right rarely bites, and Pensieve says so rather than claiming an exemption.

5.4 A person may complain about an artificial-intelligence capability directly to info@pensievelabs.org, whether or not they are a Pensieve customer.

Each of these is a gate, and each produces a record. The internal mechanism is POL-GL-132; the gates are published here so a hospital can ask for the evidence.

6.1 The inventory is published before enablement, not after. A hospital should never find a capability in the Platform that is not in DIS-GL-027.

8.1 Failure of T-1, T-3 or T-4 is a serious internal incident, logged in REG-GL-203 and reported to the affected hospitals.

9.1 Reviewed semi-annually under POL-GL-502, in step with POL-GL-132, because this area moves faster than the annual cycle.

9.2 A change that would weaken 1.1, 1.2 or 1.4 is not a policy change; it would be a change to what Pensieve is, and would be notified to every Customer sixty (60) days in advance with the reasoning published.