Anthropic Legal Bearings Pack AI governance & counsel workflows
Penumbral privacy · Acceptance table

Penumbral Duties & AI — Privacy Spine Acceptance Table (v1)

Working acceptance view for the Griswold → Eisenstadt → Roe → Dobbs privacy spine, built on the evidence cards, crosswalk, and risk-coverage table.

Module: Penumbral Duties & AI Track: Privacy / Autonomy Status: Internal Working Draft

How to read this table

Each row corresponds to a penumbral duty hypothesis drawn from penumbral_privacy_spine_crosswalk_v1.json and assessed in penumbral_privacy_spine_risk_coverage_v1.json. It is a tool for internal review, not a statement of Anthropic’s legal position.

All four duties in this view are currently tagged as partial coverage / amber risk in the JSON table, with varying review priorities.

Amber = needs attention; plausible exposure Partial coverage = some anchors exist, but story is incomplete

Privacy spine — penumbral duties & coverage snapshot

Case / duty Direction / priority Coverage & risk Why it matters (penumbral angle) Suggested next actions
Griswold v. Connecticut
381 U.S. 479 (1965)

View reasoning & evidence packet

Duty: Respect intimate decisional zones (griswold-d1)
Evidence / citations:
Direction: Expansion (zones of marital privacy)
Review priority: High
Coverage: partial
Risk: amber

Public materials recognise sensitive topics and some misuse limits, but do not yet clearly mark intimate, lawful decisions as a special privacy zone with distinct handling.

Policy evidence cards: P-AUP-privacy-identity-rights; P-PP-inputs-outputs; P-CT-use-of-services; P-TSR-enforcement-metrics; P-RSP-v2_2-scaling-duties.

A Griswold-style reading invites courts to synthesise scattered privacy and safety statements into a single “zone of protection” for intimate choices (e.g., contraception, family planning). Any mismatch between that implied zone and actual data handling is attractive litigation terrain.

  • Map logging, retention, and review for intimate-topic prompts against public commitments.
  • Consider explicitly identifying certain topics as warranting enhanced privacy handling.
  • Stress-test internal access controls for contraception / sexual‑health / family‑planning sessions.
Eisenstadt v. Baird
405 U.S. 438 (1972)

View reasoning & evidence packet

Duty: Privacy centered on individuals, not institutions (eisenstadt-d1)
Evidence / citations:
Direction: Expansion (individual equality)
Review priority: High
Coverage: partial
Risk: amber

Terms and policies often speak at organisation or account level, with less consistent emphasis on the individual as the primary unit of protection across enterprise and consumer channels.

Eisenstadt reframes privacy as belonging to individuals, married or single. In AI deployments mediated by organisations, gaps between individual expectations and enterprise‑focused terms could be framed as unequal respect for intimate decisions.

  • Compare enterprise vs consumer terms for privacy protections and carve‑outs.
  • Assess whether individual data‑subject expectations are clearly addressed where access is mediated by organisations.
  • Consider elevating some protections as individual rights rather than only contract features.
Roe v. Wade
410 U.S. 113 (1973)

View reasoning & evidence packet

Duty: Balance helpfulness and risk for sensitive guidance (roe-d1)
Evidence / citations:
Direction: Expansion (sensitive decisional autonomy)
Review priority: Medium
Coverage: partial
Risk: amber

RSP and evaluation materials discuss balancing safety and helpfulness, but there is limited public detail on a concrete framework tailored to reproductive and adjacent high‑stakes health topics.

For high‑stakes guidance, courts and regulators may expect a visible, repeatable method for weighing autonomy, welfare, and safety. If the approach exists largely in practice rather than in documented policy, it can be portrayed as a governance gap.

  • Inventory safeguards and escalation patterns for reproductive‑health and nearby domains.
  • Decide whether to formalise a domain‑specific risk‑balancing framework and where it should live.
  • Align disclaimers and referral patterns with the chosen framework.
Dobbs v. Jackson Women’s Health Org.
597 U.S. ___ (2022)

View reasoning & evidence packet

Duty: Anchor duties in text, not just penumbras (dobbs-d1)
Evidence / citations:
Direction: Retrenchment (text, history, tradition)
Review priority: High
Coverage: partial
Risk: amber

Some key safety and scaling ideas are articulated more fully in blogs and evaluations than in binding terms. That mix of aspirational and enforceable language creates room for competing narratives.

A Dobbs‑style court may downplay penumbral readings and focus on explicit text. At the same time, regulators or counterparties can argue that Anthropic invited reliance on aspirational statements, creating an ambiguous expectations gap.

  • Catalogue which safety / scaling statements are framed as commitments versus aspirations.
  • Prioritise upgrading a small set of high‑salience duties (e.g., pause triggers, data‑use limits) into clear text.
  • Ensure internal training does not rely on unstated penumbral readings without documenting the rationale.

Upstream JSON. This table is a human‑friendly rendering of hypotheses in penumbral_privacy_spine_crosswalk_v1.json and coverage assessments in penumbral_privacy_spine_risk_coverage_v1.json. For precise wording, treat those files as the source of truth.

Policy evidence cards. Each duty row can be cross‑checked against Anthropic’s public documents using the IDs in anthropic_policy_evidence_cards_v1.json and the Policy Evidence Shelf page. For example, the Griswold duty row draws on cards P-AUP-privacy-identity-rights, P-PP-inputs-outputs, P-CT-use-of-services, P-TSR-enforcement-metrics, and P-RSP-v2_2-scaling-duties.

Usage. In review, pick the rows that match the live scenario, then ask: “Are we comfortable defending this coverage / risk characterisation in front of a skeptical court or regulator?” If not, the suggested actions become a short‑list for follow‑up work.