#Responsible AI Principles

STM-GL-034 | Version 1.0.0 | Last Modified On 03 August 2026

---

#Purpose

This is Pensieve Labs's statement of how it thinks about artificial intelligence in a hospital. It is the principles document that produced the two operational documents beside it, and it does not repeat what they say.

A principle that cannot be tested against a control is a slogan. Each principle below is stated with the mechanism that makes it real and the document that records it. Where Pensieve Labs has a commitment rather than a control, Section 10 says so plainly rather than dressing an intention as a fact.

The whole of this document reduces to one sentence: in Pensieve, artificial intelligence assists a person and never replaces one, and it never crosses the clinical boundary in DIS-GL-028. No principle in it is moved by a configuration, an Order Form, a price, or a request from a hospital.

---

#Applicability

These principles are a property of the software and of how Pensieve Labs builds it, not of where the software runs. They apply identically in all four deployment models.

Where a capability requires an external model provider, its availability by deployment model is stated in DIS-GL-027 Section 6, not here.

---

#1. The clinical boundary comes before every principle

Pensieve is a hospital operating system: an administrative and record-management platform. It is not a medical device, and no capability that uses a model changes that. No capability diagnoses, triages, calculates a patient-specific dose, computes a clinical risk or early-warning score, interprets an image or a signal, or recommends treatment. The machine-learning functions in the Platform are administrative and operational: coding suggestions for billing, denial prediction, demand and stock forecasting, roster optimisation, document classification, transcription without clinical interpretation, and search over records a user is already permitted to see.

The boundary, jurisdiction by jurisdiction and with the provision relied on in each, is DIS-GL-028. The feature-level inventory that stays inside it is DIS-GL-027. This statement adds one thing to both: the boundary is the first principle, not a caveat to the others. A hospital cannot opt into having Pensieve make a clinical decision, because that capability is refused at design and is not built (POL-GL-132 Section 3.1; ADD-GL-005).

---

#2. Human accountability for every output that touches a person

Every consequential action, whether 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. The capability assists; the person decides.

This is enforced, not asserted. For each artificial-intelligence capability a hospital enables, the role accountable for its outputs is recorded. No auto-approve control and no bulk-accept control is built for any suggestion class: a coder accepts, edits or rejects each coding suggestion; an unsigned transcription is a draft, not a record; a ranked queue is shown alongside the unranked list. The operational detail is POL-GL-060 clause 2. The design rule behind it is simple: a capability that cannot be given a human decision point is not built.

---

#3. The principles, and the mechanism behind each

Sections 4 to 7 expand the four principles a hospital's data-protection and clinical-governance functions test hardest.

---

#4. Provenance and traceability

This is P4 in full. Every invocation of a model-assisted capability that touches personal data is logged: who invoked it, when, on which record, which capability, and which model version produced the output. A suggestion that influenced a person's decision can be reconstructed afterwards from the audit trail (DIS-GL-013). This is what makes the machine-generated label in P3 more than cosmetic: the label tells a user in the moment, and the log tells an investigator later. Where a change to a model, a prompt or a threshold alters a capability's behaviour, that change is a release with a change record, so the version named in the log is meaningful rather than approximate.

---

#5. Training data: what the hospital's data is never used for

Pensieve does not use a hospital's data to train, fine-tune or improve any model offered to another hospital, or to Edsol Edtech Pvt. Ltd. itself. This holds whether the data is identified, de-identified or aggregated, and it is not relaxed by permission. It is a contractual undertaking in DPA-GL-001, and the full position, including the treatment of any external model provider, is DIS-GL-027 Section 4 and Section 5. A model available to one hospital does not see, learn from, or produce an output influenced by another hospital's data, in any deployment model.

Pensieve Labs adds one undertaking of conduct to the contract: no Pensieve Labs employee may ask a hospital to relax this as part of a commercial negotiation (POL-GL-060 clause 1.4).

There is a second, independent reason for the commitment. Using hospital data to develop or validate a model is biomedical research under the applicable Indian ethical guidance, which would require ethics-committee review and, in most framings, participant consent. That reasoning sits alongside the data-protection analysis and is recorded in DIS-GL-027 Section 8. Pensieve Labs does not do it.

---

#6. Fairness and equity in an Indian clinical setting

Indian care is multilingual and served across a wide range of literacy. A model trained mainly on English text, or on one population, under-serves patients who speak another language or read at a lower level, and a model trained on operational data reflects the patterns in that data. Pensieve Labs does not treat a capability validated in one language or one population as validated everywhere.

The clinical boundary in Section 1 is what keeps the consequence of this proportionate. Because every capability that uses a model is administrative, the failure mode of a biased model is an inaccurate billing code, a wrong forecast or a misrouted document, corrected by the person who decides and logged either way. It is not a clinical harm. That is a deliberate design choice, not an accident of scope: keeping artificial intelligence away from clinical decisions keeps the equity failure mode financial and operational rather than clinical (DIS-GL-027 Section 9.5).

Pensieve Labs applies, voluntarily, the algorithmic due-diligence discipline that Indian law places on a significant data fiduciary, although Pensieve Labs is not so designated: every capability is assessed for whether its output could disadvantage a person, and the assessment is recorded (DIS-GL-027 Section 8). The non-discrimination and equity commitment this rests on is POL-GL-069. Stated honestly, Pensieve Labs does not yet publish bias-testing results for these capabilities. Section 10 records that as a commitment to be met with the accuracy figures, not as a control in place today.

---

#7. Explainability, honestly

A model's output is not explainable line by line. Where a capability suggests a code or flags a claim, Pensieve shows the evidence the capability drew on, but it does not produce a formal account of the model's internal reasoning, and Pensieve Labs does not claim otherwise. This is precisely why every capability has a human decision point rather than an explainability guarantee: the person can see the source and the suggestion and decide, which is a stronger control than a claim of full transparency the technology cannot honestly support (DIS-GL-027 Section 9.2).

---

#8. The regulatory horizon

Pensieve Labs states its regulatory position so a hospital's legal function does not have to reconstruct it. Two sentences frame it. India has no binding, horizontal artificial-intelligence statute as at 03 August 2026. Pensieve, because it takes no clinical decision, sits outside the high-risk and medical-device regimes that would otherwise apply.

#8.1 India

Governance runs through existing law, the Information Technology Act, 2000 and the Digital Personal Data Protection Act, 2023, together with voluntary guidance. The Ministry of Electronics and Information Technology published the India AI Governance Guidelines on 05 November 2025, setting out seven guiding principles and a largely voluntary, graded approach; NITI Aayog's Responsible AI principles (2021) are also voluntary. Pensieve Labs aligns to these as principles, not as binding law, and does not describe them as obligations they are not.

The DPDP Act regulates the personal data an artificial-intelligence system uses, not the automated decision as such. It contains no right equivalent to a right against solely automated decisions; its nearest provision is Section 8(3), under which a Data Fiduciary must keep personal data that is likely to be used to make a decision affecting a person complete, accurate and consistent. The Digital Personal Data Protection Rules, 2025 commence in phases, and the data-principal rights and Board enforcement commence in the third phase, expected in 2027 [UNVERIFIED: the exact third-phase commencement date is not settled in the sources, which place it in 2027]. Throughout, the hospital is the Data Fiduciary and Pensieve Labs is its Data Processor: the hospital holds these duties and Pensieve Labs assists it (DPA-GL-001; POL-GL-053).

One point is stated to head off a common overstatement. The labelling duty for synthetically generated content in the 2026 amendments to the Information Technology intermediary rules binds intermediaries and the users who post such content publicly, not a business-to-business hospital-software processor. Pensieve labels machine-generated output at the point of use as a matter of principle (P3), not because that rule compels it to.

#8.2 The medical-device boundary

The reason Pensieve stays outside medical-device regulation in every market is intended purpose: it performs no diagnostic, monitoring, scoring or treatment function on an individual patient. The Indian position under the Central Drugs Standard Control Organisation, the European position under Regulation (EU) 2017/745 Rule 11, and the equivalents in Australia and the United Arab Emirates, are set out with the named provision in each in DIS-GL-028. This statement does not restate them. It records that the clinical boundary in Section 1 is what keeps every one of those positions true.

#8.3 European Union

Under the EU Artificial Intelligence Act, Regulation (EU) 2024/1689, an artificial-intelligence system becomes high-risk only in one of two ways: it is, or is a safety component of, a product that must undergo third-party conformity assessment as a medical device (Article 6(1) with Annex I), or it performs a use listed in Annex III, which in healthcare means chiefly determining eligibility for healthcare services or emergency patient triage. Pensieve does neither. Being used in a hospital does not, by itself, make a system high-risk, and Pensieve Labs does not describe its Platform as high-risk artificial intelligence.

The Act's transparency duties in Article 50 apply from 02 August 2026, and the labelling in P3 already meets their intent for the administrative outputs Pensieve produces. The high-risk obligations were deferred by the Digital Omnibus on artificial intelligence, in force 27 July 2026: standalone Annex III high-risk systems now apply from 02 December 2027, and product-embedded high-risk systems from 02 August 2028. Pensieve Labs notes the deferral so that no reader mistakes the general application date of the Act for the date its high-risk rules take effect.

---

#9. Frameworks Pensieve Labs takes as reference points

Pensieve Labs is informed by the following instruments. It is aligned to them as reference points; it does not claim to be compliant with, or certified or conformity-assessed against, any of them. Edsol Edtech Pvt. Ltd. holds no artificial-intelligence certification of any kind (POL-GL-132 Section 6.3).

Citing a framework is not conformity. Pensieve Labs writes "informed by" and "aligned to", and reserves "certified" and "compliant" for a certificate that exists (WPR-GL-005 Section 6).

---

#10. Maturity: enforced by a control today, or a commitment

The honest division between what a control enforces now and what remains a commitment.

A change that would weaken P1, P2 or P5 is not a change to this statement; it would be a change to what Pensieve Labs is, and it would be published, with its reasoning, before it took effect.

---

#Complaints

A person may complain about an artificial-intelligence capability directly to info@pensievelabs.org, whether or not they are a Pensieve Labs customer. The Grievance Redressal Policy (POL-GL-066) governs how the complaint is received, investigated and answered. Where the person is a patient, the hospital is the Data Fiduciary and the request is made to the hospital, which Pensieve Labs assists under DPA-GL-001.

---

#Related documents

---

#Change history

---

---

STM-GL-034 v1.0.0 | Last Modified On 03 August 2026 | Review due 03 February 2027 | Published at https://trust.pensievelabs.org