Data exposure can happen even when Claude behaves perfectly. If a tool call returns a full customer record including SSN and Claude faithfully includes it in a reply, that's not a model error — it's a design failure: too much sensitive data was allowed into the context in the first place. The fix is structural (trim the input/tool output before it reaches Claude), never just an instruction telling Claude not to repeat it — that's probabilistic, not a control.
6.3 — Three Regulatory Frameworks · Core
All three can apply to the same organization at once (a telehealth company with EU patients that also serves a federal agency faces all three). GDPR is explicitly extraterritorial — the trigger is the data subject's EU residency, never the company's home country.
6.4 — Human Review for Consequential Actions · Core
Human authorization is required before an irreversible or high-impact action executes — not after:
- Sending external communications (emails, contracts)
- Financial transactions/commitments
- Data deletion (irreversible by definition)
- Decisions affecting a person's opportunities: credit, hiring, benefits eligibility
For high-volume, low-risk, reversible work, sampled review is sufficient — per-action approval isn't required everywhere, only where the action is irreversible or consequential. Regulatory human-accountability requirements apply regardless of model accuracy — a 99%-accurate model still doesn't substitute for the mandated human sign-off on an adverse credit decision.
6.5 — Bias, Fairness, and Transparency · Core
Aggregate accuracy does not prove fairness. A 95%-overall-accurate system can be 71% accurate for one language or demographic group — always check per-slice, not just the total. Fairness needs consistent criteria across groups (different thresholds per demographic is disparate treatment, even with good intentions) and ongoing monitoring — a system fair at launch can drift unfair as usage patterns shift over months.
Transparency has three separate, all-required elements:
- Disclose AI involvement — presenting Claude as a human (a human name, photo, evasive answers to "are you AI?") is a violation on its own.
- Explain the decision — "the AI decided" is not an explanation; the actual factors must be articulable.
- Document known limitations so users know when to seek other help.
Disclosure alone is never sufficient — that's the standard exam trap in this section.
Exam reflexes for Domain 6
- Claude doesn't refuse a request → proves nothing about company policy compliance; check the AUP separately.
- Full customer record (with SSN) faithfully returned by Claude → not a model error; the design let too much data into context.
- BAA mentioned in a scenario → signals HIPAA specifically, not GDPR.
- "EU resident's data, non-EU company" → GDPR still applies; extraterritorial by design.
- Irreversible action (send, pay, delete) proposed with no human step → requires authorization before execution, always.
- 90%+ aggregate accuracy cited as proof of fairness → wrong; check per-slice before concluding anything.
- Different confidence thresholds by demographic group, even "for equity" → disparate treatment, not a fix.
- AI discloses it's AI but gives no reasoning for a decision → transparency incomplete; explainability is a separate requirement.