{"resourceId":"ccj-public-defender-genai-case-202609","versions":[{"version":"external-726c08f1687b5690843bd840801cfb341630be47db14496f7f5d75e5b3cd1c04","resource":{"id":"ccj-public-defender-genai-case-202609","title":"CCJ defender case study allows general-purpose AI only when confidentiality preconditions enable Standard Deploy","organization":"Council on Criminal Justice Task Force on Artificial Intelligence","sector":"Indigent defense / courts-adjacent practice","geography":"United States (hypothetical large urban public defender office)","publishedAt":"September 2026","publicationDate":null,"eventDate":null,"sourceName":"CCJ","sourceLabel":"Council on Criminal Justice case study","sourceUrl":"https://counciloncj.org/assessing-ai-in-action-a-case-study-on-public-defender-use-of-general-purpose-ai-tools","evidenceClass":"standards-guidance","outcomeClass":"emerging","topics":["knowledge-work","data-security","governance-procurement","accessibility-workforce"],"finding":"September 2026 CCJ case study examines general-purpose AI for public defender research, drafting, discovery analysis, and case prep in a large urban office. Classification is low risk and substantial opportunity—triggering Standard Deploy—only if privacy/confidentiality preconditions are met: enterprise-tier or air-gapped tools, no-training-on-customer-data terms, MDM and network blocks against consumer AI for client work, and lawyer-in-the-loop verification. Without those preconditions, recommendation flips to Generally Avoid because consumer accounts create a privilege-loss pathway (citing Heppner). Notes ~300 judicial AI standing orders (as of March 2026) and large hallucination-case counts as operational context. Not an RCT of defender outcomes.","sledRelevance":"Useful for public defender offices, court IT, and county counsel setting acceptable-use policy: distinguishes efficiency opportunity from privilege risk. Cross-relevant to prosecutor offices with similar confidentiality duties. Interpretation; jurisdiction-specific disclosure matrices still required.","evidence":"Opened original HTTPS HTML. Executive summary and risk section state the binary privacy precondition; implementation details specify MDM, domain blocking, and Acceptable-Use Policy (Appendix J).","architectureImplications":"Prefer enterprise SSO, document-management integration, air-gap where needed; block consumer AI domains on office networks; log approved-tool use; keep verification and court-disclosure checklists in filing workflow.","governanceImplications":"Adopt Acceptable-Use Policy defining permitted/prohibited uses and data-entry bans; maintain court-rule matrix; measure citation-flag catch rates and prohibited-tool use; reassess before renewal or on model/policy changes.","securityPrivacyImplications":"Confidentiality is the controlling risk: contractual confidentiality, no training on customer data, bind subprocessors, data residency where required; personal consumer accounts for client work are a primary failure mode.","caveats":"Hypothetical office; opportunity claims reference secondary efficiency reports; criminal-defense-specific accuracy benchmarks remain limited; Heppner and standing-order counts are evolving.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: Sell confidentiality architecture and acceptable-use design, not generic ChatGPT seats. The customer problem is caseload pressure plus shadow use of consumer AI that can waive privilege or create sanctions exposure. Bound engagement to selecting enterprise or air-gapped options, writing policy, and a verification-time pilot. Do not promise fewer ineffective-assistance claims or faster dispositions without measured pilots.","engineering":"Engineering framing: Fit is enterprise-tier or air-gapped tooling only—not consumer chat accounts on personal devices. Integration work centers on SSO, DLP or network controls that block consumer AI domains, and document-management hooks so verification and court-disclosure checklists sit in the filing path. Prerequisites include negotiated no-training-on-customer-data terms, FedRAMP or equivalent where required, and operational detection of prohibited tool use. Proof of value is net time savings after verification plus zero confidentiality incidents—not raw prompt throughput.","delivery":"Delivery framing: Deliver policy, CLE-style training, a court-disclosure matrix, MDM and network controls, and a pilot measuring verification time and citation errors caught. Ownership sits with a named managing attorney plus IT security. Acceptance criteria include zero approved client work on consumer accounts and net time benefit after verification—or narrowed tool scope. Risks include staff bypass on personal devices and hallucination reaching filings despite training."},"roleDetails":{"sales":{"customerProblem":"Under-resourced defender offices need efficiency without destroying privilege or inviting sanctions from unverified AI output.","stakeholders":["Chief public defender","IT/security lead","Training/CLE coordinator","County counsel","Line attorneys and investigators"],"discoveryQuestions":["Are consumer AI domains blocked on office networks and managed devices?","Does the contract forbid training on customer data and bind model subprocessors?","How will verification time be measured against claimed savings?","Which judges' standing orders require AI disclosure in this venue?"],"valueHypothesis":"Enterprise-controlled AI plus verification workflows can reallocate attorney time to counseling and investigation if—and only if—confidentiality preconditions hold.","potentialEngagement":"Acceptable-use policy + enterprise tool selection + 60–90 day pilot with verification-time and citation-error metrics.","unsupportedClaims":["Guaranteed case-outcome improvement","That consumer ChatGPT is safe for client facts","That RAG legal tools eliminate hallucination","One disclosure rule covering all assigned courts"]},"engineering":{"fit":"Fit for offices that can field MDM, SSO, and enterprise AI; unfit if staff must use personal consumer accounts for client work.","architectureIntegration":"Enterprise AI behind SSO into DMS; network DNS/proxy blocks for consumer AI; optional air-gapped research appliance for sealed matters.","prerequisites":["Contractual confidentiality and no-train terms","MDM on office devices","Acceptable-use policy with consequences","Filing checklist hooks for court disclosure"],"constraints":["Budget limits on legal-AI enterprise pricing","Weak state/local-law model performance","Frequent foundation-model migrations","Detection gaps for personal-device bypass"],"security":"Treat prompts containing client facts as privileged ESI; DLP; eDiscovery-ready logging; prohibit sealed-record entry into non-approved systems.","proofOfValue":["Net time savings after verification","Citation-error catch rate in attorney review","Zero confirmed privilege incidents","Compliance rate with court disclosure rules"]},"delivery":{"work":"Write Acceptable-Use Policy, deploy MDM/network controls, train on hallucination and ethics, maintain court-rule matrix, run pilot.","dependencies":["IT security capacity","CLE budget","Vendor public-sector pricing","Supervising attorney bandwidth"],"ownership":"Named managing attorney accountable for AI practice policy, with IT owning technical controls.","skillsAdoption":"Role-specific training for research, drafting, discovery review; emphasize lawyer-in-the-loop and jurisdiction disclosure.","governanceCheckpoints":["Precondition checklist before Standard Deploy","Pilot go/no-go on confidentiality/sanctions","Reassessment before contract renewal","Triggered review on vendor privacy-term changes"],"acceptanceCriteria":["No approved client work on consumer accounts","Disclosure checklist used on AI-assisted filings","Pilot shows net time benefit or tool scope narrowed"],"risks":["Personal-device shadow AI","Hallucinated citations reaching court","Privilege waiver under Heppner-line facts","Verification effort erasing savings"]}},"retrievedAt":"2026-09-17T03:15:10Z","enrichedAt":"2026-09-17T03:20:00Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Training and CLE are core controls; understaffed offices may lack verification capacity, which itself weakens the opportunity case.","procurementImplications":"Level 1 terms only if Standard Deploy applies: enterprise/air-gap requirement, hallucination methodology disclosure, migration notice, confidentiality and no-train clauses, termination for rights concerns.","operatingModelImplications":"Shift from informal individual AI use to office-controlled tooling with continuous monitoring of prohibited-tool use and disclosure volume.","sourceVerification":{"openedUrl":"https://counciloncj.org/assessing-ai-in-action-a-case-study-on-public-defender-use-of-general-purpose-ai-tools/","referenceExcerpt":"low risk and substantial opportunity, as long as privacy preconditions are met, which triggers “Standard Deploy.”","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}