EDPS Generative AI Orientations: Evidence for EU Institutions
The EDPS generative AI orientations apply Regulation (EU) 2018/1725 to EU institutions, bodies, offices, and agencies. They call for documented roles, DPO involvement, risk assessment, data minimisation, traceability, security, and regular review. Runtime records can support that work, but they cannot decide legal basis or prove how a provider trained its model.

The EDPS generative AI orientations give EU institutions a specific operating context: Regulation (EU) 2018/1725, not a generic GDPR checklist. An institution has to identify its role, involve its DPO, assess high-risk processing, limit personal data, document the system lifecycle, and keep reviewing a changing service. Runtime evidence helps when it records the authenticated caller and approved model route, plus the policy and treatment of prompt data. Several controlling decisions still belong to the institution and its providers.
I would reject any readiness plan that turns this guidance into a screenshot of a settings page. A regulator needs decisions and evidence tied to an actual processing operation.
TL;DR
- The EDPS orientations address EU institutions under Regulation (EU) 2018/1725 and sit separately from the AI Act.
- Build an evidence chain for purpose and DPO advice, followed by DPIA review and minimisation. Include approved routes, retention and incidents.
- Routed HTTP records can show caller and destination, plus policy and redaction. They can also show the outcome. They cannot establish legal basis or provider training practices.
- Test one permitted request and one blocked request, then reconcile both with the institution's processing records.
The orientations set an EUI-specific scope
The EDPS generative AI orientations address EU institutions, bodies, offices, and agencies, which the document calls EUIs. The EDPS issued them in its role as data protection supervisor. It also says the guidance is without prejudice to the AI Act. That distinction keeps the legal analysis anchored to the right instrument.
For personal data processing by an EUI, the controlling text is Regulation (EU) 2018/1725. The orientations apply its principles across development and deployment, including inputs and outputs once a system is running. They also cover training data and inferred information. A provider's claim that a model uses anonymised or synthetic data calls for evidence about the controls behind that claim.
This scope creates a useful first field in the evidence register: processing stage. Training, validation, inference, output review, and later reuse should have separate owners and records. One "generative AI" row hides too much.
Accountability starts before the first prompt
The EDPS assigns the DPO an active advisory role. Independent advice requires enough information about the lifecycle and input and output mechanisms, together with personal data processing and the decision procedure. The controller should document its processes and update records of processing. As a best practice, it should also maintain a specific inventory of generative AI systems and applications.
A DPIA is required before processing likely to create a high risk to fundamental rights and freedoms. The orientations call for regular monitoring because a model can worsen an identified risk or create a new one. They also require the controller to seek DPO advice during the DPIA. The broader GDPR AI DPIA workflow offers a related operating pattern, but an EUI should cite Regulation 2018/1725 and the EDPS material in its own record.
The readiness file should therefore name the purpose and controller. It should preserve the DPO consultation and DPIA decision, including the date and trigger for review. Provider roles and transfers need their own entries. A gateway event can later join to this file through a protected system or use-case identifier. It cannot create the legal analysis after the fact. Nothing can.
Minimisation needs a route-level decision
The orientations state that personal data must be adequate, relevant and limited to what the purpose requires. That obligation covers testing, acceptance and production. The EDPS also calls for high-quality datasets, documented structure and intended use, and periodic review. For a third-party service, the EUI should assess the provider's treatment of minimisation too.
Inference adds a concrete decision point. Before an authenticated request leaves the institution, policy can evaluate its data class and approved purpose. It can remove an unnecessary identifier, replace a value with a token, block a restricted category, or reject an unapproved destination. The record should preserve the transformation type and policy version without copying every sensitive value into general telemetry.
Picture a procurement officer comparing two records on a laptop: the source case contains a citizen's full name, while the routed prompt carries PERSON-0427 and a protected case reference. That visual difference is evidence of a minimisation action. It still needs a documented purpose and authorised re-identification path, with a retention rule and proof that the token mapping is protected.
The operating record needs stable fields
A useful route record starts with the authenticated subject and role. Add the source application and approved use case, then the provider endpoint and model identifier. Record the data classification and minimisation action. Include the policy version, disposition, precise event time, and a request identifier that joins to the institution's processing record. Retention class and incident correlation complete the operational layer.
Those fields support several parts of the orientations. They show which route handled personal data and which rule applied. During an investigation, they can expose prompt injection and jailbreak attempts, plus suspected disclosure. They also create a bounded trace for rights handling when the institution must locate processing connected to an individual. The AI decision record pattern is useful here, provided the EUI adds its Regulation 2018/1725 purpose and role, followed by retention and rights metadata.
Keep the evidence layers separate. Provider documentation covers training sources and validation, including accuracy and model limitations. The institution's governance file covers lawful basis, DPIA conclusions, notices, controller and processor allocation, and transfer decisions. Runtime traffic records cover the request path that actually crossed the managed gateway.
A two-request test exposes missing evidence
Choose one approved EUI use case in a nonproduction environment. Send a synthetic request through an authenticated application session with a marker such as EUI-ORIENT-118. The request should contain one field that policy permits and one unnecessary identifier that policy removes. Verify the destination and model, then the policy version and transformation. Check the result and event time separately. Reconcile the request identifier with the AI inventory and DPIA record.
Run a second request against an unapproved provider endpoint. The expected result is a block with the same identity and use-case context, plus the policy rule that caused the decision. Hand both records to the DPO or an internal reviewer. Without asking an engineer to reconstruct console logs, the reviewer should identify the processing purpose and governing assessment, plus the owner, retention rule, and technical outcome.
A direct provider call outside the managed route should produce no gateway event. Record that result as a coverage gap for endpoint and network controls, plus procurement or account controls. The AI governance framework can hold that wider control map. The gateway evidence remains deliberately narrower.
The proof boundary stays narrow
Runtime records can support an EUI's account of who sent personal data, where it went, what policy applied, and which minimisation or blocking action occurred. They can also support incident reconstruction and retention controls for the routed event. That is useful evidence, especially when it is independent of the calling application's write path.
The same records cannot establish the lawfulness of training data. They cannot prove a provider's controller or processor role, validate consent, determine a transfer mechanism, test bias across the model lifecycle, or satisfy a data subject request by themselves. DPO advice and DPIA conclusions remain institutional decisions. Accuracy testing and model documentation need evidence from their owners. So do notices, rights handling, and provider oversight.
The orientations are technologically neutral and expressly avoid prescribing specific technical measures. Treating one gateway log as proof of compliance would contradict that design. The defensible claim is smaller: route evidence supplies one verifiable part of a larger Regulation 2018/1725 accountability record.
DeepInspect
DeepInspect is a stateless proxy for authenticated HTTP traffic deliberately routed between EUI users or agents and LLM endpoints. Before forwarding, it can evaluate application-supplied identity and use-case context, plus the destination and data classification. Policy can permit, redact, reroute, or block the request, and response inspection can apply a separate decision before content returns.
Each routed decision can produce a signed, tamper-evident record with identity and route, followed by model, policy version, data treatment, disposition, and time. Those fields support the operating evidence described above. DeepInspect does not determine lawful basis, approve a DPIA, validate provider training data, fulfil individual rights, or prove off-route processing. Local models and direct provider access remain outside its coverage. So do provider-side training and downstream business decisions. If traffic does not travel through the authenticated HTTP proxy, DeepInspect does not see or control it.
Book a technical deep dive at deepinspect.ai.
Frequently asked questions
- Do the EDPS orientations apply to private companies under the GDPR?
The orientations are written for EU institutions, bodies, offices, and agencies under Regulation (EU) 2018/1725. Private companies may find the operational questions useful, but they should map their duties to the GDPR and the competent supervisory authority's guidance. An EUI should preserve the Regulation 2018/1725 context rather than relabel the document as general GDPR guidance.
- Does every EUI generative AI use require a DPIA?
The EDPS ties the requirement to processing likely to result in high risk to individuals' rights and freedoms. The controller should assess that threshold before processing and seek DPO advice. New technology is a relevant factor. The decision and its rationale belong in the governance file, along with a review trigger when the system or its risks change.
- Should an EUI store complete prompts and responses?
That choice depends on purpose, necessity, risk, retention, and the rights process. Full content can help reconstruction while creating another sensitive store. A protected source reference and classification, together with a transformation record and digest, may provide enough route evidence when authoritative content remains available elsewhere. The controller should document and test the chosen design.
- Can route evidence prove that a provider excluded prompts from training?
A route record can show what the institution sent and which provider endpoint received it. Provider-side reuse requires separate evidence such as contractual terms, technical configuration, service documentation, and assurance material. The institution should reconcile that evidence with its role allocation, transfer analysis, and procurement controls.