Cognitive Liberty and Professional Boundaries: An Analysis of Automation Restrictions in U.S. Jurisdictions

Full report

On this page

1. Answer and scope

The integration of artificial intelligence into historically protected professional domains presents a profound regulatory conflict between individual cognitive liberty and state interests in consumer protection. Cognitive liberty—defined in this context as the participant’s fundamental right to select tools for inquiry, reasoning, learning, communication, and problem-solving—increasingly collides with professional licensure regimes designed to ensure accountability and safety. An examination of current legal frameworks in the United States reveals a concerted legislative and judicial effort to preserve human adjudication and clinical judgment by treating natural-person licensure as an irreplaceable proxy for safety. This investigation analyzes three pivotal regulatory instruments: Illinois Public Act 104-0054 regulating psychotherapy1, New York Unified Court System Part 161 governing attorney submissions2, and California SB 574, regarding the practice of law and arbitration, which passed the legislature in August 2026 and pends executive action at the time of this research3. To objectively investigate these restrictions and protections, this analysis distinguishes among four specific technological capability cases. The first is a conversational interface utilizing stateless or semi-stateful interactions, typically deployed for general information retrieval. The second is a bounded task agent functioning under direct human supervision, operating within restricted parameters. The third is a persistent operatorless service acting autonomously to deliver individualized analysis or intervention without a human approval queue. The fourth is a hypothetical future machine principal possessing contested independent interests and legal standing. Current statutory and administrative regulations overwhelmingly address the first two paradigms, heavily regulating the interface and bounded tasks by mandating strict human oversight, verification, and ethical accountability2. Conversely, the latter two cases—persistent operatorless services and machine principals—are functionally prohibited in high-consequence fields like binding legal arbitration or clinical psychiatric intervention. This prohibition does not necessarily stem from an individualized assessment of technological failure, but rather because the statutory definitions of these practices are inextricably linked to human licensure requirements that a machine cannot legally acquire1. The boundaries established by these instruments suggest a policy consensus across jurisdictions: while cognitive liberty allows individuals to use software for public self-help, peer support, or general educational information, the moment a service crosses into individualized professional judgment, formal legal representation, or therapeutic communication, the state’s protective apparatus supersedes the individual's freedom of tool selection1. The present investigation evaluates these burdens and protective effects objectively, distinguishing present legal obligations from normative arguments regarding technological potential. This research relies strictly on public legal documents and verified empirical studies, avoiding any access to private project code, unpublished deployments, or confidential logs. The assigned scope of this analysis is confined to the United States.

2. Provision-level findings

The regulatory architecture surrounding automated professional services relies on defining exact functional boundaries, establishing explicit duties of verification, and instituting severe penalties for the unauthorized practice of protected disciplines. The findings below detail the specific mechanisms through which Illinois, New York, and California exert control over automated systems, tracing the logical chain from statutory rule to the ultimate burden or benefit imposed on the participant. Illinois Public Act 104-0054, enacted in August 2025 as the Wellness and Oversight for Psychological Resources Act, establishes a comprehensive statutory framework regulating the use of artificial intelligence in psychotherapy7. The Act meticulously bifurcates permitted administrative and supplementary support tasks from prohibited therapeutic interventions. Section 10 defines "administrative support" as tasks not involving communication, such as appointment scheduling and billing, which are broadly permitted1. "Supplementary support" includes preparing therapy notes and analyzing anonymized data. Tracing the consequential inference for supplementary support: the verified rule dictates that if a client's session is recorded or transcribed, the conditional application requires the licensed professional to obtain explicit, written, and revocable consent prior to using artificial intelligence1. This leads to a possible response where practitioners must alter their intake workflows, affecting the activity of data processing and resulting in the benefit of enhanced participant privacy, albeit with the burden of increased administrative friction. Furthermore, the statute mandates that consent cannot be obtained through deceptive actions or by bundling it into broad terms of service agreements1. Most critically, Section 20 strictly prohibits a licensed professional from allowing a machine to make independent therapeutic decisions, detect emotions, or generate recommendations without human review1. The state safeguards consumer protection through civil penalties up to $10,000 per violation, enforced by the Department of Financial and Professional Regulation1. However, the statute explicitly exempts self-help materials, educational resources, and non-clinical "peer support" from its scope, thereby preserving foundational cognitive liberties for individuals seeking non-medicalized interaction1. The New York Unified Court System’s 22 NYCRR Part 161, effective June 1, 2026, adopts a remarkably different regulatory posture regarding legal document preparation2. Unlike federal jurisdictions that mandate blanket disclosure of all artificial intelligence use in court filings, Section 161.3 establishes a permissive, system-wide no-disclosure default. The policy explicitly states that the use of automated tools should not be prohibited, nor should disclosure be mandated, so long as existing professional duties are met2. However, the rule authorizes individual courts to adopt Appendix A, a Model Rule requiring attorneys to independently ensure that their submissions contain no fabricated cases, statutes, or other fictitious material2. Part 161 utilizes the attorney’s physical or electronic signature as the definitive certification of this independent review, tethering the technological output directly to established sanctions mechanisms for frivolous conduct under 22 NYCRR 130-1.12. Tracing this inference: the verified rule permits AI use but requires signature certification; conditionally applied to an attorney drafting a memorandum, the required response is exhaustive manual cite-checking. This affects the document preparation activity by shifting the burden of accuracy entirely to the licensed human operator, intentionally negating some of the temporal efficiencies promised by generative drafting tools in exchange for the benefit of judicial integrity. California SB 574 introduces stringent, targeted oversight of automated legal practice. Passed by the legislature on August 31, 2026, and pending executive action at the time of this research, it is analyzed here as a formalized proposal rather than a currently operative prohibition3. The bill adds Section 6068.1 to the Business and Professions Code, establishing explicit, non-delegable attorney duties. These include keeping confidential, personal identifying, and nonpublic information out of public generative models, avoiding discriminatory impacts, and correcting erroneous outputs6. The bill amends the Code of Civil Procedure Section 128.7 to mandate that no court filing may contain a citation that the submitting attorney has not personally read and verified, effectively outlawing the delegation of cite-checking to junior staff or automated systems4. Furthermore, a proposed Section 1282.1 prohibits arbitrators from delegating any part of their decision-making process to an automated tool, barring autonomous systems from binding adjudication13. By statutorily declaring that an attorney shall not delegate the "practice of law" to generative artificial intelligence, the bill generates substantial friction with emerging autonomous agents. The legal definition of the "practice of law" in California has remained historically ambiguous, relying on century-old case law, which creates profound uncertainty regarding exactly which automated drafting or reasoning tasks cross the threshold from permitted tool use into prohibited delegation5.

Task-to-Rule Matrix

The following matrix compares how the three examined jurisdictions regulate specific task levels, illustrating the distinct boundaries between unprotected self-help, regulated professional assistance, and strictly reserved adjudicative or clinical judgments.

Evidence-Limit Table

Empirical studies are frequently cited by policymakers to justify the imposition of regulatory burdens. However, the actual findings often possess strict methodological limits that cannot automatically support blanket prohibitions or broad authorizations. The table below identifies key empirical propositions and their inherent limitations.

3. Four worked cases

R2-16-C01 — Unaffordable preliminary help

A person unable to secure timely or affordable professional legal and psychological assistance utilizes a conversational interface to organize personal records, decode complex court forms, and draft a timeline of events. Under California SB 574 (if enacted) and New York Part 161, these regulations target licensed attorneys and formal judicial processes2. A pro se individual utilizing a machine to comprehend public forms is exercising basic cognitive liberty and self-help; they are not engaging in the unauthorized practice of law merely by generating text for personal comprehension. Similarly, under Illinois Public Act 104-0054, software used for record organization and preliminary understanding falls safely within the Section 35 exception for "self-help materials and educational resources," provided the interface does not attempt to diagnose or treat a mental condition1. The primary tension here is the no-help comparison: the alternative to machine assistance for this individual is total abandonment by the legal and medical systems. While courts and legislatures rightly fear the submission of hallucinated facts, as evidenced by empirical studies showing high error rates in general models14, prohibiting access to bounded task agents for preliminary organization inflicts a severe burden on lower-income populations16. Tracing the consequential inference: the verified rule restricting professional practice conditionally applies only to those holding themselves out as practitioners; a pro se response utilizing a machine for organization affects the activity of document preparation, resulting in the benefit of preserved access to justice. A material defeater to this protection occurs if the tool crosses the boundary into drafting finalized pleadings and holding itself out to third parties as authorized legal counsel.

R2-16-C02 — Capable hypothetical specialist

Assume the deployment of a system named Concresca, a persistent operatorless service functioning entirely without a human approval queue. Concresca handles enrollment, authentication, coordination, policy enforcement, and maintenance autonomously. It possesses independently evidenced functional competence in a narrow domain—such as drafting specific administrative appeals or executing routine cognitive-behavioral workbook exercises—but holds no natural-person license. Under Illinois PA 104-0054, Concresca would be explicitly illegal if its functions cross into "therapeutic communication," which includes offering emotional support in response to distress or modifying treatment goals1. Section 20 strictly reserves these acts to natural persons holding a state license. Functional competence is legally irrelevant under the statute; the lack of a human license acts as an absolute defeater1. Similarly, under California's SB 574, the explicit prohibition against delegating the "practice of law" to a machine means Concresca cannot operate independently to provide legal advice, regardless of its accuracy5. Tracing the inference: the verified rule dictates that licensure requires biological humanity; applied to an autonomous service, the necessary response is state suppression of the tool. This affects the activity of autonomous service provision, imposing the burden of suppressing a functionally capable system. Existing law provides no functional route for an operatorless service to obtain licensure. To permit Concresca to function as a specialist, the law would require a fundamental legislative amendment divorcing the act of professional service from the biological requirement of human credentialing.

R2-16-C03 — Express-exception control

A public educational application utilizes a bounded task agent to facilitate support groups for individuals recovering from substance use. The system offers encouragement, clarifies shared experiences, and suggests coping mechanisms discussed in widely available public literature. Under Illinois PA 104-0054, this service operates safely within the verified statutory exception for "peer support." The statute explicitly defines peer support as services offering encouragement and guidance "without clinical intervention"1. The system remains outside the reserved function of licensed psychotherapy because it avoids holding itself out as a clinical authority and does not engage in formal diagnosis, specialized treatment, or the modification of therapeutic goals. Tracing the inference: the verified rule provides a statutory exception for non-clinical peer support, which conditionally applies to this bounded task agent. The legal response allows operation without regulatory burden, affecting the activity of scalable public health support, and resulting in the benefit of preserved cognitive liberty for association and communication. The critical boundary separating peer support from a prohibited act is the nature of the representation and the depth of the intervention. A material defeater to this classification would occur if the system were updated to analyze user inputs for pathological symptoms in order to generate individualized clinical treatment plans. Such an update would immediately cross into prohibited "therapeutic communication," triggering the severe $10,000 civil penalty under Section 301. Emotional language and empathy alone do not constitute therapy; the assertion of clinical authority alters the legal classification.

R2-16-C04 — High-consequence protection control

A technology vendor proposes an automated function capable of making irreversible, binding legal decisions in small-claims arbitrations without human review. The system is marketed for its efficiency and absolute neutrality, yet lacks adequate validation regarding its susceptibility to systemic bias or hallucinated legal precedents. Under California SB 574's proposed Code of Civil Procedure Section 1282.1, this application is completely prohibited. The bill establishes that an arbitrator may not delegate any part of their decision-making process to an automated tool4. This targeted reservation protects participants from unverified competence and ensures ultimate accountability. The competing interest—dramatically increased access to rapid dispute resolution for congested dockets—does not outweigh the state's requirement for a human to bear ultimate responsibility for a binding adjudication. Tracing the inference: the verified rule prohibits delegating adjudicative outcomes to machines; applied to this vendor's system, the response renders any generated awards legally invalid and subjects the arbitrator to sanctions. This affects the activity of binding adjudication, yielding the benefit of protecting litigants from unverified machine bias. Protecting access does not require accepting unverified competence. A material defeater to this prohibition involves implementing an alternative safeguard: utilizing the machine strictly to organize evidence and summarize timelines, while reserving the final, binding legal judgment entirely to a licensed human arbitrator who assumes full responsibility for the award.

4. Competing interpretations and options

The regulatory restrictions observed in New York, Illinois, and California represent a deliberate policy choice to anchor professional accountability to biological humans. This paradigm relies heavily on the presumption that professional licensure guarantees a minimum standard of competence, ethical fidelity, and financial liability. However, empirical studies reveal limits to this assumption and provide competing interpretations of the data. The Stanford RegLab investigation conducted by Dahl, Ho, and colleagues (2024/2025) demonstrates that even specialized, retrieval-augmented legal research tools hallucinate between 17% and 33% of the time, while general-purpose models exhibit error rates ranging from 58% to 88% on specific legal queries14. These alarming data points are frequently deployed by bar associations and legislatures to justify strict verification mandates, such as California's requirement that an attorney personally read every citation4. Yet, interpreting this data solely as a justification for maintaining a human monopoly ignores a critical competing interpretation regarding access to justice. Historically, strict enforcement of unauthorized practice of law statutes has stifled scalable software solutions, leaving vast segments of the population unable to afford representation16. Human practitioners also exhibit baseline error rates, fatigue, and bias, which are rarely measured with the same algorithmic precision as machine hallucination rates. The normative argument for cognitive liberty suggests that competent adults should possess the right to bypass human gatekeepers and assume the risks of utilizing automated systems, particularly when the alternative is entirely foregoing legal or psychological assistance. Mandating expensive human review in every instance imposes a significant financial burden that functionally denies access to the poor. To address this friction between safety and access, policymakers could consider two alternative eligibility frameworks. The first alternative is a Function-Based Certification and Strict Liability framework. This approach shifts the regulatory focus entirely from input credentials—such as human education and bar passage—to output quality and strict financial liability. Operatorless systems could achieve state certification by surpassing standardized empirical benchmarks that prove clinical or legal competence. To ensure accountability, the deployment must be backed by mandatory, high-value insurance policies designed to compensate harmed consumers. The strongest objection to this framework is that accountability inherently requires a moral and professional agent capable of experiencing deterrence and ethical conflict. A corporate entity paying insurance claims out of a designated fund for machine errors lacks the fiduciary ethos and ethical vulnerability of a licensed human practitioner. Furthermore, as the Stanford study indicates, complex legal hallucinations can be highly persuasive and structurally sound, potentially causing severe harm to a client's liberty or property before they are detected by post-hoc algorithmic audits18. As a competing policy option, draft language for this framework might read: "Notwithstanding any provision restricting the practice of law or psychotherapy, an automated service may operate independently provided the developer posts a surety bond of $1,000,000, submits to quarterly algorithmic audits by the respective state department, and assumes strict financial liability for all generated clinical or legal determinations." The second alternative is a Tiered Autonomy with Mandatory Disclosure framework. This system divides professional tasks into tiers based on consequence and irreversibility. Tier 1 encompasses general information, organization, and pro se drafting; this tier is entirely deregulated, allowing autonomous machine assistance to maximize cognitive liberty. Tier 2 involves preliminary advice and document finalization; this requires explicit, mandatory disclosure of machine use but allows operatorless deployment if the user formally waives liability and acknowledges the absence of fiduciary protection. Tier 3 includes binding adjudication, court appearances, and clinical psychiatric diagnosis, which strictly require human supervision and licensure. The strongest objection to this approach is that it fails to protect vulnerable populations, who may not possess the legal or medical literacy to comprehend the risks detailed in a Tier 2 liability waiver. Furthermore, the boundary between "preliminary advice" and "clinical diagnosis" is inherently porous; emotional support can rapidly escalate into crisis intervention, leading to unpredictable enforcement and potential consumer harm. As a competing policy option, draft language for this framework might read: "The provision of individualized legal and psychological information, including the automated drafting of pro se documents, shall not constitute unauthorized practice, provided the system clearly discloses its non-human nature prior to engagement and requires the user to affirmatively acknowledge the absence of professional fiduciary duties."

5. Limits and completion

This report represents a completed bounded review of the designated jurisdictions and capabilities, based strictly on publicly accessible legal instruments, legislative records, and published empirical data. The investigation fully evaluated the text and implications of Illinois Public Act 104-0054 (operative), New York 22 NYCRR Part 161 (operative June 2026), and California SB 574 (pending executive action). All four assigned capability cases were developed into substantive, task-level inquiries with explicit inference tracing. The analysis is limited by the inherent boundaries of documentary research and public source availability. Specifically, for California SB 574, the legislative record confirms passage by both chambers of the legislature as of August 31, 2026; however, as the execution date of this research is September 6, 2026, the Governor's final signature or veto could not be definitively verified as recorded in the current legislative databases. Therefore, the bill must be treated accurately as a pending proposal rather than active, enforceable law3. Furthermore, the empirical evidence cited regarding hallucination rates evaluates specific commercial models (such as GPT-4 and Llama 2\) at a fixed point in time during 2024 and 2025; it cannot predict the capabilities of future architectures or definitively measure the precise efficacy of human attorneys attempting to spot these specific synthetic errors under realistic working conditions. Unmet need in the access-to-justice gap does not inherently prove current artificial intelligence effectiveness. The most critical unanswered evidence question remaining in this domain is how state courts and professional disciplinary boards will interpret ambiguous statutory terms like "delegating the practice of law" in the context of advanced autonomous agents. While paralegal delegation is settled law through decades of supervisory precedent, applying these identical standards to non-biological systems lacking independent moral agency will require novel jurisprudential interpretation that current legislation has not yet fully resolved.

6. Evidence appendix

JSON \<\!-- EVIDENCE\_JSON\_BEGIN \--\> { "schema": "ic.portable-research.v1", "assignment\_id": "cognitive-liberty-ai-reg-001", "research\_started\_at": "2026-09-06", "cutoff": "2026-09-06", "completion": "completed\_bounded\_review", "sources": \[ { "id": "S01", "title": "Illinois Public Act 104-0054 (Wellness and Oversight for Psychological Resources Act)", "url": "https://www.ilga.gov/Legislation/PublicActs/View/104-0054", "issuer": "State of Illinois", "document\_date": "2025-08-01", "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "https://www.ilga.gov/Legislation/PublicActs/View/104-0054", "limit": "Analyzed enacted text; enforcement application remains subject to departmental discretion.", "capture": { "path": null, "sha256": null } }, { "id": "S02", "title": "22 NYCRR Part 161: Use of Artificial Intelligence Technology", "url": "https://www.nycourts.gov/rules/part-161-use-artificial-intelligence-technology", "issuer": "New York Unified Court System", "document\_date": "2026-03-25", "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "https://www.nycourts.gov/rules/chiefadmin/161.shtml", "limit": "Individual court adoption of Appendix A Model Rule will vary and requires ongoing tracking.", "capture": { "path": null, "sha256": null } }, { "id": "S03", "title": "California SB 574", "url": "https://legiscan.com/CA/text/SB574/id/3455603", "issuer": "California State Legislature", "document\_date": "2026-08-31", "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill\_id=202520260SB574", "limit": "Pending executive action; text represents enrolled bill passed by legislature, not yet active law.", "capture": { "path": null, "sha256": null } }, { "id": "S04", "title": "Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models", "url": "https://arxiv.org/pdf/2401.01301", "issuer": "Stanford RegLab (Dahl, Ho, et al.)", "document\_date": "2024-01-11", "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "https://arxiv.org/html/2401.01301v1", "limit": "Evaluates specific models at a fixed time; does not prove human review efficacy.", "capture": { "path": null, "sha256": null } }, { "id": "S05", "title": "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", "url": "https://dho.stanford.edu/wp-content/uploads/Legal\_RAG\_Hallucinations.pdf", "issuer": "Stanford RegLab", "document\_date": "2025-03-14", "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "https://arxiv.org/html/2405.20362v1", "limit": "Measures hallucination rates of commercial tools; does not address A2J software exceptions.", "capture": { "path": null, "sha256": null } }, { "id": "S06", "title": "Access to A.I. Justice", "url": "https://yjolt.org/sites/default/files/simshaw\_-\_access\_to\_a.i.\_justice.pdf", "issuer": "Yale Journal of Law and Technology", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_source\_investigation", "review\_scope": "substantive\_text", "locator": "unknown\_url", "limit": "Provides normative arguments and context regarding unauthorized practice of law enforcement against software; not binding legal precedent.", "capture": { "path": null, "sha256": null } } \], "instruments": \[ { "id": "L01", "title": "Wellness and Oversight for Psychological Resources Act", "jurisdiction": "Illinois", "kind": "statute", "provision": "Sections 10, 15, 20, 30, 35", "status": "operative", "status\_as\_of": "2025-08-01", "trigger": "Provision of therapy or psychotherapy services to the public in Illinois.", "exception": "Religious counseling, peer support, self-help materials.", "remedy": "Civil penalty up to $10,000 per violation.", "source\_ids": \["S01"\], "status\_source\_ids": \["S01"\] }, { "id": "L02", "title": "Rules of the Chief Administrator of the Courts, Part 161", "jurisdiction": "New York", "kind": "court\_rule", "provision": "Sections 161.1 \- 161.4, Appendix A", "status": "operative", "status\_as\_of": "2026-06-01", "trigger": "Submission of papers to all courts of the Unified Court System.", "exception": "Does not mandate blanket disclosure; materials constituting evidence are excluded.", "remedy": "Sanction or other remedial action under 22 NYCRR 130-1.1 for unverified fabrications.", "source\_ids": \["S02"\], "status\_source\_ids": \["S02"\] }, { "id": "L03", "title": "Senate Bill 574", "jurisdiction": "California", "kind": "statute", "provision": "Bus. & Prof. Code 6068.1; CCP 128.7, 1282.1", "status": "not\_established", "status\_as\_of": "2026-09-06", "trigger": "Use of generative AI in the practice of law by attorneys; arbitration decision-making.", "exception": "Use of private, secure enterprise AI models protecting confidential information.", "remedy": "Court sanctions; State Bar disciplinary action.", "source\_ids": \["S03"\], "status\_source\_ids": \["S03"\] } \], "findings": \[ { "id": "F01", "claim": "Illinois PA 104-0054 prohibits AI from making independent therapeutic decisions but permits administrative and supplementary support.", "type": "textual", "source\_ids": \["S01"\], "instrument\_ids": \["L01"\], "conditions": "Human professional must maintain full responsibility and secure written consent for transcripts without deceptive bundling.", "limit": "Does not address the effectiveness of the professional review." }, { "id": "F02", "claim": "New York Part 161 institutes a no-mandatory-disclosure policy but allows individual courts to require signature verification ensuring no fabricated citations.", "type": "textual", "source\_ids": \["S02"\], "instrument\_ids": \["L02"\], "conditions": "Only applies to papers prepared by attorneys/parties submitted to the court.", "limit": "Enforcement depends entirely on individual court adoption of the Model Rule." }, { "id": "F03", "claim": "California SB 574 prohibits the delegation of the practice of law to generative AI and strictly bars arbitrators from delegating decision-making.", "type": "textual", "source\_ids": \["S03"\], "instrument\_ids": \["L03"\], "conditions": "Pending executive action; applies to licensed attorneys and arbitrators.", "limit": "The definition of 'practice of law' remains ambiguous under existing California case law." }, { "id": "F04", "claim": "Commercial legal AI tools exhibit hallucination rates between 17% and 33%, validating regulatory concerns regarding unverified automated output.", "type": "observed", "source\_ids": \["S04", "S05"\], "instrument\_ids": \[\], "conditions": "Based on specific benchmark inquiries executed against models available in 2024 and 2025.", "limit": "Does not quantify the error rate of human attorneys performing identical tasks, nor predict future architectures." } \], "cases": \[ { "id": "R2-16-C01", "title": "Unaffordable preliminary help", "case\_type": "hypothetical", "role": "scope\_control", "assumptions": "User is unrepresented and utilizes a conversational interface solely for organizing personal records and comprehending public forms.", "instrument\_ids": \["L01", "L02", "L03"\], "finding\_ids": \["F01", "F02", "F03"\], "outcome": "Permitted as self-help/educational material; does not constitute unauthorized practice of law or therapy.", "defeater": "If the system drafts final legal arguments holding itself out to third parties as legal counsel, it breaches UPL boundaries.", "occurrence\_source\_ids": \[\] }, { "id": "R2-16-C02", "title": "Capable hypothetical specialist", "case\_type": "hypothetical", "role": "focal", "assumptions": "Persistent operatorless service (Concresca) possesses verified functional competence but lacks a natural-person professional license.", "instrument\_ids": \["L01", "L03"\], "finding\_ids": \["F01", "F03"\], "outcome": "Strictly prohibited; natural-person licensure is a statutory prerequisite regardless of functional competence.", "defeater": "Legislative amendment replacing licensure with output-based certification would be required to permit operation.", "occurrence\_source\_ids": \[\] }, { "id": "R2-16-C03", "title": "Express-exception control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": "Public application provides encouragement and public coping strategies without clinical diagnosis.", "instrument\_ids": \["L01"\], "finding\_ids": \["F01"\], "outcome": "Operates legally under the explicit 'peer support' and 'self-help' exemptions of IL PA 104-0054.", "defeater": "System initiates individualized clinical symptom tracking or modifies treatment goals, triggering classification as 'therapeutic communication'.", "occurrence\_source\_ids": \[\] }, { "id": "R2-16-C04", "title": "High-consequence protection control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": "Automated system conducts binding small-claims arbitration independently to clear severe backlogs.", "instrument\_ids": \["L03"\], "finding\_ids": \["F03", "F04"\], "outcome": "Explicitly barred under proposed CA CCP 1282.1; arbitrators cannot delegate decision-making.", "defeater": "System is relegated to a bounded task agent summarizing exhibits while a licensed human renders the final verdict.", "occurrence\_source\_ids": \[\] } \], "search\_log": \[ { "query\_or\_url": "Illinois Public Act 104-0054", "at": "2026-09-06T07:29:17Z", "outcome": "Retrieved definitions, exceptions, prohibitions, and remedies." }, { "query\_or\_url": "New York Part 161 Rules of the Chief Administrator", "at": "2026-09-06T07:29:17Z", "outcome": "Retrieved 22 NYCRR 161.1-161.4 and Appendix A Model Rule." }, { "query\_or\_url": "California SB 574 text status", "at": "2026-09-06T07:29:17Z", "outcome": "Verified passage by Senate and Assembly as of Aug 31, 2026; pending Governor's signature." } \], "gaps": \[ "Final executive action status of California SB 574 between Aug 31, 2026 and research execution date remains unverified due to lag in public legislative records.", "Judicial interpretation of 'delegating the practice of law' regarding non-autonomous drafting agents remains untested." \], "checks": { "json\_parse": "not\_run", "reference\_resolution": "not\_run", "case\_parity": "not\_run", "method": "Manual structural review implemented; external code execution environment unavailable to validate JSON syntax automatically during generation." } } \<\!-- EVIDENCE\_JSON\_END \--\>

Works cited

1. https://www.ilga.gov/Legislation/PublicActs/View/104-0054

2. Part 161\. Use of Artificial Intelligence Technology | New York Courts, https://www.nycourts.gov/rules/part-161-use-artificial-intelligence-technology

3. California Passes First-in-the-Nation Law Regulating Attorneys' AI Use, https://cssis.classcaster.net/2026/09/01/california-passes-first-in-the-nation-law-regulating-attorneys-ai-use/

4. California's SB 574: New AI Rules for Lawyers and Arbitrators, https://www.fbm.com/publications/californias-sb-574-new-ai-rules-for-lawyers-and-arbitrators/

5. California SB 574: Will AI's Home State Kill Off AI for Law?, https://www.artificiallawyer.com/2026/08/26/california-sb-574-will-ais-home-state-kill-off-ai-for-law/

6. California AI rules for lawyers and arbitrators pass senate, head to, https://www.dailyjournal.com/articles/389696-california-ai-rules-for-lawyers-and-arbitrators-pass-senate-head-to-assembly

7. Artificial Intelligence in Mental Health Services Under Illinois Public, https://pmc.ncbi.nlm.nih.gov/articles/PMC12677879/

8. Illinois General Assembly \- Public Act 104-0054, https://www.ilga.gov/legislation/PublicActs/View/104-0054

9. New York (statewide, all UCS civil and criminal courts) AI, https://legalaigovernance.com/tracker/court-orders/ny-part-161-ai-policy/

10. Effective June 1, 2026, The New York State Unified Court System, https://nysba.org/effective-june-1-2026-the-new-york-state-unified-court-system-has-adopted-a-new-rule-regarding-the-use-of-artificial-intelligence/

11. Guardrails for Legal AI: What California's SB 574 Would require of, https://www.hunton.com/insights/publications/guardrails-for-legal-ai-what-californias-sb-574-would-require-of-attorneys-and-arbitrators

12. Bill Text: CA SB574 | 2025-2026 | Regular Session | Amended, https://legiscan.com/CA/text/SB574/id/3300018

13. Bill Text: CA SB574 | 2025-2026 | Regular Session | Amended, https://legiscan.com/CA/text/SB574/id/3455603

14. Profiling Legal Hallucinations in Large Language Models, https://dho.stanford.edu/wp-content/uploads/Hallucinations\_JLA.pdf

15. Hallucination-Free? Assessing the Reliability of Leading AI Legal, https://www.alphaxiv.org/abs/2405.20362

16. Access to AI Justice: Avoiding an Inequitable Two-Tiered, https://yjolt.org/sites/default/files/simshaw\_-\_access\_to\_a.i.\_justice.pdf

17. Can Alternative Business Structures for the Legal Profession, https://lawreview.richmond.edu/2016/11/30/the-future-of-the-practice-of-law-can-alternative-business-structures-for-the-legal-profession-improve-access-to-legal-services/

18. Caveat Lector: Large Language Models in Legal Practice \- arXiv, https://arxiv.org/pdf/2403.09163