GDPR, PIPL and your gateway traffic: a short field guide
2026-11-13
GDPR, PIPL and your gateway traffic: a short field guide
Two privacy regimes, one request path, and a lot of confident-sounding advice
that falls apart the moment you check the official texts. This is the version we
wished someone had handed us before we wired the first gateway.
The two questions that decide everything
1. Whose personal data is in the prompt? US/EU residents → GDPR. Mainland
residents → PIPL. Both in the same prompt → both apply, per field.
2. Where does the request physically go? GDPR cares about transfers out of
the EU/EEA; PIPL cares about exports out of the mainland. A domestic key on
an international route makes an EU user's data a mainland export — that is
exactly the S3 flow the gateway blocks at the code level (see the
GDPR, in the shape engineers need it
- No adequacy decision for China as of 2026-10. EU-to-China needs Article 46
safeguards: SCCs, a transfer impact assessment, and supplementary measures.
Standard contractual clauses are a template, not a strategy — the TIA is where
you find out the clauses are insufficient.
- Your prompt logging choice is a compliance surface. If the gateway keeps
prompt bodies, you are a processor of potentially special-category data, and
the retention clock starts on write. We default to not persisting prompts
(see the docs' data-residency section for the current state of each flag).
- Data subject requests (access / deletion) must reach the upstream that
actually holds the data. If your architecture logs prompts in three places,
you need a deletion story for all three.
PIPL, in the shape operators need it
- Export paths: security assessment, standard contract, or certification.
The mainland-specific part is that the volume and sensitivity of data
decides which path applies. Keeping pilot volumes small is a legitimate
compliance decision, not a hack.
- Separate consent for cross-border transfer of personal information is
required in the mainland flow (S2), and consent must be unbundled from other
terms. Bundle it and you have invented a new compliance problem.
- The entity question comes first. A mainland entity with an international
product is in the export regime by construction; an offshore entity for the
international business is not. Registering in the right place is cheaper than
every control in this list combined.
What we publish, and what we don't claim
We publish the boundary, not a promise. The cross-border page states exactly
which claims are confirmed (the three-flow architecture, the S3 code-level
block) and which are pending (international data residency pending Baidu's
written confirmation). If a vendor's dashboard says "compliant", ask them which
article, which transfer path, and which entity — the answer is usually thinner
than the badge.
One honest note about this post
We are not your lawyer, and this is not legal advice. It is a map of the
terrain our engineering team had to understand to build the gateway; the legal
review before launch is a separate, human step. See also
cross-border AI without the fear
for the architecture side.