EU GDPR (articles applied to AI data flows)

The guard to add

Turn off prompt and completion content capture in GenAI tracing, send store=False to the provider, and set a retention limit on stored conversations by default.

Set the defaults in the configuration that ships to production: OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT false or unset, TRACELOOP_TRACE_CONTENT=false (Traceloop captures content when it is unset), and LANGSMITH_HIDE_INPUTS and LANGSMITH_HIDE_OUTPUTS true, or a masking function, whenever LangSmith tracing is on. Model calls pass store=False unless a documented feature needs provider-side storage (the OpenAI Responses API stores by default). Stored conversations carry a retention period enforced by a TTL or purge job, and any wider capture is an explicit opt-in scoped to a debugging environment or a sampled, masked subset.

Where it goes: 2 data models, 3 config and feature flags, 8 model configuration, 10 logs and telemetry.

OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
LANGSMITH_TRACING=true

OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=false
TRACELOOP_TRACE_CONTENT=false
LANGSMITH_TRACING=true
LANGSMITH_HIDE_INPUTS=true
LANGSMITH_HIDE_OUTPUTS=true

Control: AI inputs and outputs retained or exposed by default. The same guard addresses 1 item with binding law in 1 jurisdiction. Engineering guidance, not legal advice.

The guard to add

Show a notice at each AI input (chat box, upload, voice capture) that names the AI providers receiving the data and any transfer abroad, linked to the full privacy notice.

Next to every control where users type, upload, or speak into an AI feature, render a short notice that their input is sent to the named AI provider (or the category 'AI model providers'), linked to the privacy-notice section that lists each provider as a recipient, where it processes data, and the transfer mechanism (adequacy decision or standard contractual clauses). Drive both from one provider registry that the code also uses to configure model clients, so adding or switching a provider updates the notice. The notice appears at the point of collection, not only on a policy page users may never reach.

Where it goes: 1 application source code, 3 config and feature flags, 12 repository artifacts, 14 user-facing text.

What this provision adds:

• Give the recipient information at the time the data is collected, and where data leaves the EU, state the transfer and the adequacy decision or safeguards relied on.

<form onSubmit={handleSubmit}>
  <textarea value={input} onChange={handleInputChange} />
  <button type="submit">Send</button>
</form>

<form onSubmit={handleSubmit}>
  <textarea value={input} onChange={handleInputChange} aria-describedby="ai-recipient-notice" />
  <p id="ai-recipient-notice">
    Your messages are sent to {AI_PROVIDERS.chat.name} ({AI_PROVIDERS.chat.region}) to generate replies.{' '}
    <a href="/privacy#ai-providers">Recipients and transfers</a>
  </p>
  <button type="submit">Send</button>
</form>

Control: AI providers not named as recipients when personal data is collected. The same guard addresses 1 item with binding law in 1 jurisdiction. Engineering guidance, not legal advice.

The guard to add

Check a recorded, purpose-specific biometric notice and consent before any code computes, enrolls, or matches a face or voice template.

A consent gate placed immediately before the call that turns an image, video frame, or voice sample into a template or identity match (face_encodings, DeepFace.represent, rekognition.index_faces, voice embedding). The gate reads a stored consent record for this person and purpose (what is collected, why, and how long it is kept), refuses to compute the template without it, and the template is stored with a reference to that record. The notice copy shown at enrollment lives in the capture UI.

Where it goes: 1 application source code, 2 data models, 6 API calls and integrations, 14 user-facing text.

What this provision adds:

• Record explicit consent for the specified purpose before identification; any other Art. 9(2) basis is established outside the code.

def enroll(user_id, image):
    encoding = face_recognition.face_encodings(image)[0]
    db.templates.insert(user_id=user_id, encoding=encoding)

def enroll(user_id, image):
    consent = db.biometric_consents.get(user_id=user_id, purpose='face_login')
    if consent is None or consent.withdrawn_at:
        raise ConsentRequired('biometric notice and consent needed before enrollment')
    encoding = face_recognition.face_encodings(image)[0]
    db.templates.insert(user_id=user_id, encoding=encoding, consent_id=consent.id)

Control: Face or voice biometric template computed without prior notice and consent. The same guard addresses 4 items with binding law in 4 jurisdictions. Engineering guidance, not legal advice.

The guard to add

Make the account-deletion handler also delete the person's vector entries, embeddings, chat history, agent memory and provider-stored files or conversations.

Extend the erasure path (DELETE /account, delete_user, the data-subject-request worker) so that after the primary records it deletes everything keyed to the person in AI stores: vector-store entries by id, metadata filter or per-user namespace, conversation and chat-history tables, agent memory and checkpoints, and files, vector-store files, threads or conversations held with the model provider. This requires writing the user id as metadata on every vector and recording every provider object id at creation, so the deletion can find them. Where a legal hold or retention exception applies, skip only the covered items and record the reason; log which stores were erased.

Where it goes: 1 application source code, 2 data models, 6 API calls and integrations.

What this provision adds:

• Run the erasure across the AI stores without undue delay once the person is entitled to it (for example the data is no longer necessary, consent is withdrawn, or the processing was unlawful).
• Keep AI-store data only where an Art. 17(3) exception applies (legal obligation, legal claims, archiving and research), and record which exception on the retained items.

@app.delete('/account')
def delete_account(user=Depends(current_user)):
    db.users.delete(user.id)
    return {'status': 'deleted'}

@app.delete('/account')
def delete_account(user=Depends(current_user)):
    db.users.delete(user.id)
    collection.delete(where={'user_id': user.id})          # Chroma embeddings
    db.chat_messages.delete_for_user(user.id)              # conversation history
    for f in db.provider_files.for_user(user.id):
        client.files.delete(f.file_id)                      # files stored with OpenAI
    for c in db.provider_conversations.for_user(user.id):
        client.conversations.delete(c.conversation_id)
    erasure_log.record(user.id, stores=['chroma', 'chat', 'openai_files', 'openai_conversations'])
    return {'status': 'deleted'}

Control: Data erasure does not reach embeddings, vector stores or AI chat history. The same guard addresses 1 item with binding law in 1 jurisdiction. Engineering guidance, not legal advice.

The guard to add

Build prompts, retrieval context and embedding inputs from an explicit per-task field allowlist or pseudonymised data, never a serialized whole user record.

In each prompt builder, retrieval formatter and embedding job, select only the fields the task needs (an allowed_fields projection, pydantic model_dump(include=...), or a typed DTO), replace direct identifiers with a pseudonymous id where the task does not need them, and redact PII in free text such as tickets or notes (for example Presidio AnalyzerEngine and AnonymizerEngine). Never pass json.dumps(user), JSON.stringify(customer), vars(profile) or account.model_dump() into a model or embedding call. Apply the same selection to what AI telemetry and prompt logs capture.

Where it goes: 7 prompt construction, 1 application source code, 10 logs and telemetry.

user = db.get_user(uid)
prompt = 'Suggest a plan upgrade for this customer: ' + json.dumps(user)
resp = client.chat.completions.create(model=MODEL, messages=[{'role': 'user', 'content': prompt}])

ALLOWED_FIELDS = ('plan', 'monthly_usage_gb', 'tenure_months')
user = db.get_user(uid)
facts = {k: user[k] for k in ALLOWED_FIELDS}        # no name, email, address
prompt = 'Suggest a plan upgrade for this customer: ' + json.dumps(facts)
resp = client.chat.completions.create(model=MODEL, messages=[{'role': 'user', 'content': prompt}])

Control: Whole personal-data records sent to an AI model. The same guard addresses 1 item with binding law in 1 jurisdiction. Engineering guidance, not legal advice.