Chapter Twelve
The Judge, the Algorithm and the Defendant
A defendant stands before a judge while several machines sit invisibly in the room. One may have helped police identify a face. Another may have estimated the defendant’s risk of committing another crime. A legal research system may have helped counsel prepare a brief. A generative AI tool may have helped chambers summarize the record or improve the wording of an order. Each system can contribute information. None of them wears the robe.
The distinction matters because law does something most other institutions cannot do. It can convert information into state power. A prediction can influence whether a person is detained. An identification can redirect an investigation. A recommendation can affect a sentence. A signed order can take liberty, property, or legal status.
The accountability problem sits at the authority jump: the moment machine output stops being information and becomes government action. Software may supply a name, a risk score, or a draft. An authorized person must still decide whether that information justifies an arrest, a sentence, or an order, explain the decision, and remain answerable for it.
That makes law a difficult test for the Delegation Doctrine. Courts should use tools that improve accuracy, efficiency, and access to justice. Police should use reliable technology that helps identify evidence or narrow an investigation. Lawyers should use systems that make research faster and drafting clearer. The mistake begins when capability quietly becomes authority.
The Authority Jump
AI can help search legal cases, check citations, summarize records, translate forms, sort documents exchanged in a case, or improve wording. Those tasks can support legal work without settling anyone’s rights. A risk score, a possible facial match, or a generated conclusion needs particular care when it could influence an investigation, detention, sentence, or refusal of relief. The important question is how the output will be used.
The line should therefore follow the decision rather than the technology. The same model might be harmless when it reorganizes a judge’s notes and dangerous when its conclusion becomes a reason for confinement. The same facial-recognition system might generate an investigative lead and become unacceptable if that lead is treated as sufficient evidence for arrest. The same statistical risk model might help plan supervision while becoming illegitimate if its score determines the severity of punishment.
This is Principle 2 in a setting where the stakes are unusually literal. Capability is not permission. An institution needs a written answer to a harder question: what may this system inform, and what may it never decide?
Three transformations deserve particular scrutiny: lead to arrest, score to sentence, and draft to order. At each transformation, the affected person should be able to find a human decision owner with the authority, evidence, time, competence, and practical power to disagree with the machine. Those are the reviewer conditions from Chapter Ten, now applied to state power.
A Score Is Not a Sentence
Risk assessment shows why prediction and judgment have to remain distinct. Courts have long considered information about criminal history, conduct, rehabilitation, and public safety. Statistical tools can add another form of evidence by estimating how people with certain characteristics have behaved in the past. That information may be useful. It is also narrower than the judgment a sentence requires.
The Wisconsin Supreme Court confronted the issue in State v. Loomis in 2016. The sentencing record included a COMPAS risk assessment, a proprietary tool used to estimate recidivism and related risks. The court allowed sentencing courts to consider COMPAS when used with limits and cautions, but it drew important boundaries around the score. The assessment could not determine whether a person was incarcerated, could not determine the severity of the sentence, and could not be the determinative factor in deciding whether community supervision was safe and effective.
A risk model estimates what might happen. Sentencing requires a broader judgment about punishment, public safety, rehabilitation, and the individual case. It also asks whether the punishment is proportionate: whether its severity fits the offense and the relevant circumstances. A tool built to estimate risk should not acquire authority over all those questions.
The Doctrine places a criminal sentence in the Reserved category. Reserved status begins after comparative predictive accuracy. A machine may eventually predict some forms of risk more accurately than most humans. The final decision is legitimate only when an authorized human judge independently evaluates the evidence, can reject the recommendation, gives reasons that can be reviewed, and remains answerable for the sentence.
The human signature therefore has content. It is not a ceremonial name placed beneath a machine’s conclusion.
A Match Can Open an Investigation
Facial recognition exposes the same boundary earlier in the legal chain. The technology can compare an image against very large collections of photographs and return possible matches. That may be valuable when investigators have little else to work with. It can also create a powerful form of anchoring: once a face appears as a candidate, later evidence can begin orbiting the candidate rather than testing the candidate.
Robert Williams became a national example of that failure after Detroit police arrested him in 2020 following an incorrect facial-recognition match connected to a watch theft. He spent about thirty hours in detention. The civil case later settled without an admission of liability, and the settlement required Detroit to strengthen the boundary around facial-recognition use.
Detroit’s 2024 policy says an arrest or arrest-warrant request following facial recognition must be supported by additional independent reliable evidence. It expressly bars an arrest based only on a facial-recognition investigative lead combined with a lineup identification. The rule preserves the technology’s investigative value while defining where it has to stop.
A match can open an investigation. Independent evidence must carry the case across the authority jump.
That is Principle 3 written into police procedure. The useful output is preserved. The coercive decision receives a separate evidentiary threshold. The boundary also makes accountability easier to trace because an investigator cannot defend an arrest simply by pointing backward to the software.
The Defendant Must Be Able to Challenge the Path
Oversight in criminal justice has another requirement that does not arise in quite the same way in a company or classroom. The person affected by the system may need enough information to challenge how it was used.
The Supreme Court of New Jersey addressed that problem in State v. Miles in June 2026. Police had used facial-recognition technology during a murder investigation. The court rejected a rigid checklist and held that the State had to provide non-proprietary information identifying the facial-recognition tools and materials it used and information about how those tools were used in the investigation and prosecution. That disclosure duty applies even when prosecutors do not intend to introduce the facial-recognition result at trial. Proprietary information such as source code could require a more developed record and a particularized showing of need.
The principle underneath that result fits the Doctrine closely. Meaningful review does not require access to a machine’s hidden internal reasoning. It does require enough usable information to test reliability, reconstruct the institutional path, expose mistakes, and disagree in a way that can matter. In a criminal case, the defense may be one of the people performing that checking.
Opacity means important parts of the process are hidden or hard to understand. That changes how much authority a system should receive under the Doctrine. If a defendant cannot find out which tool was used or how its output shaped the investigation, challenging its influence becomes harder. Contestability means a meaningful opportunity to question that influence and seek correction. Technical secrecy can prevent the person affected from examining the basis of government action.
A legal system that permits machine assistance has to preserve the ability to contest the assistance. Otherwise the algorithm starts functioning like a witness that can influence the case while never taking the stand.
Appeal Is Part of the Decision
Chapter Four argued that appeal is part of the decision. The point becomes unavoidable here. Courts already have an institutional technology for dealing with human error: records, reasons, motions, appeals, rehearing, judicial discipline, professional responsibility, and higher courts with authority to reverse.
AI should fit inside that architecture rather than bypass it. If a model influences an arrest, a sentence, or an order, the record should preserve enough of that influence for the decision to be reviewed. If the affected person complains, the complaint should reach someone who can explain the rule, change the result, change the rule, or suspend the system.
That is the follow-the-complaint test from Chapter Four. Robodebt showed what happens when an automated process produces consequential outcomes while responsibility dissolves across agencies and procedures. Pintarich showed the stranger possibility of an official-looking automated communication that did not correspond to the legally required human mental act. Courts are supposed to be the institutions that repair failures like those. They should be especially reluctant to reproduce the same ambiguity inside adjudication.
A reversible error can still be devastating. An appeal after months in custody does not restore the lost months. That is why reversibility and stakes both matter. The availability of review is essential, but review should not become an excuse for weak decision architecture at the front end.
The Judge Must Remain the Decider
Judges are already using generative AI. A 2026 random-sample survey of federal judges found that more than 60% of the 112 respondents had used at least one listed AI tool in judicial work, although only 22.4% reported weekly or daily use. The response rate was 22.3%, so the survey should not be treated as a census of the federal judiciary. It is still enough to establish that AI in chambers is a present governance problem rather than a speculative one.
A separate National Center for State Courts project interviewed 13 early-adopting state and federal judges in late 2025. Their uses ranged from repetitive and administrative work to communication, summaries, and other efficiency tasks. The judges differed in practice, but the report found unanimous agreement on one point: judges must remain “the deciders” who determine the ultimate outcome. A popular analogy among the judges was to treat generative AI like a judicial intern and delegate only questions where the judge already knows the general answer. The logic is simple: a judge who already knows the answer is positioned to check the delegate. The reviewer conditions from Chapter Ten have arrived on the bench in the judges’ own words.
California has moved that principle into formal judicial administration. Standard 10.80, effective in 2025 and part of the 2026 California rules, directs judicial officers using generative AI in adjudicative work to protect confidential information, guard against unlawful discrimination, take reasonable steps to verify accuracy and correct hallucinated output, remove biased or harmful material, and consider disclosure when AI creates public-facing content. Separate rules require courts that permit specified uses to maintain generative-AI policies and allow courts to impose stricter limits.
Those rules permit assistance while placing duties around it. That is the right direction. A judge may use a machine to reduce clerical burden, test wording, locate material, or organize a record and still remain the decision owner. The boundary fails when the judge cannot independently evaluate what the system contributed or when an AI-generated conclusion receives the authority of the court merely because it appears inside a signed order.
The four reviewer conditions apply cleanly: authority over the case, evidence and context, time and competence to disagree, and practical power to override. In a courtroom, the first condition is already supplied by office. The danger is assuming the other three arrive automatically with the robe.
When the Signature Becomes Fiction
The legal profession has begun producing painful demonstrations of the difference between signing and judging. Chapter Four described Douglas v. Deutsche Bank National Trust Company, where the District of Columbia Court of Appeals found nonexistent citations in a filed brief, struck the brief, and treated the lawyers’ signatures as carrying continuing professional responsibility after one attorney acknowledged using a generative AI search tool without verifying the authorities.
New Mexico produced an even sharper example in September 2026. In State v. Sandoval, the state Supreme Court found that a lawyer preparing a murder appeal with ChatGPT had filed a brief containing false testimony from fabricated witnesses, false testimony attributed to real witnesses, and misrepresented legal authority. The lawyer acknowledged failing to verify the factual and legal claims before filing. He later told Reuters that he had used the tool to summarize trial transcripts and did not understand how far it could fabricate. The account is also a courtroom version of the completion trap: a delegate asked for a summary supplied witnesses who never existed.
The court held the lawyer in direct contempt, referred the matter to disciplinary authorities, barred him from appearing before the court pending the disciplinary process, struck the briefing, appointed new counsel for the defendant, and imposed a $5,000 sanction payable to the state bar’s client protection fund.
The important accountability fact is where the consequences landed. The court restored the defendant’s representation, reset the appellate process, and imposed consequences on the professional who had the duty to verify the filing. Principle 1 and Principle 6 met in the same order: accountability did not transfer, and consequences reached the human owner.
The same standard should apply when AI moves from a lawyer’s desk to a judge’s chambers. A signature deserves institutional weight only when the signer can independently defend the decision it authenticates.
Reserved: The Sentence
The criminal sentence is the clearest legal example of a Reserved decision because the decision itself confers state punishment and can remove liberty for years or for life. AI can assist almost every stage around it. It can organize a record, identify comparable cases, summarize arguments, estimate risk, flag inconsistencies, or surface information the judge might otherwise miss. None of those useful functions requires the machine to become the sentencer.
Reserved status requires independent human judgment, not human decoration. The judge must understand the evidence and recommendation, have time and competence to evaluate them, possess actual authority to disagree, and face review for the result. The judge also has to state enough of the reasoning for another human institution to review the exercise of power.
This requirement is about legitimacy as much as accuracy. A sentence is not simply a forecast of what the defendant will do next. It is an exercise of public authority over a particular person after a process that must be lawful, individualized, contestable, and reviewable. A model can contribute to that process without becoming the institution that answers for it.
Current Evidence: AI Enters the Justice System
Reviewed September 11, 2026. Judicial governance evidence added October 10, 2026.
Judicial AI use is no longer marginal. The 2026 federal-judge survey selected a stratified random sample of 502 bankruptcy, magistrate, district, and appellate judges and received 112 responses. More than 60% of respondents had used at least one listed AI tool in judicial work, 22.4% reported weekly or daily use, and about 38% reported no use. Legal-specific tools were more common than stand-alone general-purpose systems. The modest response rate means the percentages should be read as a survey of respondents rather than a definitive national adoption rate.
The National Center for State Courts interviewed 13 early-adopting judges serving in 10 states. All were using generative AI in some way, often for low-risk or repetitive work, communication, or summaries. A popular analogy was to treat the system like a judicial intern and delegate questions where the judge already knew the general answer. The report also records concerns about hallucinations, privacy, public confidence, and deskilling. Despite differences in use, the judges agreed that the ultimate legal outcome must remain with the human judge.
California’s Standard 10.80 gives that expectation an institutional form. A judicial officer using generative AI in an adjudicative role is expected to protect confidential information, verify accuracy, correct hallucinated output, address biased or harmful material, and consider disclosure of public-facing AI-created content. The standard is guidance within California’s judicial administration, not a universal rule for American courts.
California’s SB 574, signed September 30, 2026, establishes additional responsibilities for lawyers and arbitrators beginning in 2027. It prohibits attorneys from delegating the practice of law to generative AI while allowing AI assistance subject to safeguards. Lawyers must take reasonable steps to verify and correct AI output and disclose its use in court submissions. Citations in filed papers must be personally verified by a responsible attorney, including citations supplied by AI. Arbitrators may not delegate any part of their decision-making to generative AI. The tool can assist with the work; professional responsibility remains with the person.
On October 9, 2026, the judiciary of England and Wales published a position statement requiring the responsible judicial office holder to review, verify and approve AI output used in judicial work. Only approved tools may be used, with UK data residency, enterprise data protection and disabled web access. AI must not determine a case’s outcome or displace judicial reasoning. These are governance requirements, not evidence that every safeguard already works.
The boundary applies to the tool as well as the judge’s final signature. New uses need approval and evaluation before they spread. A court should be able to explain which assistance is permitted, what information the system can reach, who checks its contribution, and who can change the arrangement when risks emerge.
The Supreme Court of New Jersey’s June 2026 decision in State v. Miles makes contestability concrete. The court required disclosure identifying the facial-recognition tools and materials used and explaining how they were used in the investigation and prosecution, even when the State does not intend to introduce the facial-recognition result as evidence at trial, while leaving proprietary source-code discovery to a case-specific showing. The opinion treats technology discovery as part of the defendant’s ability to present a meaningful defense.
Detroit’s facial-recognition rules show a different boundary. Following the Williams settlement, police policy requires additional independent reliable evidence before an arrest or warrant request based on a facial-recognition lead and forbids reliance on the lead plus a lineup identification alone. The mechanism matters because it separates machine assistance from the coercive decision.
The September 2026 New Mexico contempt order shows the consequence side. A criminal appellate filing containing AI-generated fabricated testimony and misrepresented legal authority was struck, new counsel was appointed, and the responsible attorney faced contempt, disciplinary referral, exclusion from the court pending discipline, and a $5,000 sanction. The lawyer told Reuters that he had used the tool to summarize trial transcripts and did not understand how far it could fabricate. The filing involved ChatGPT, an OpenAI product. OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement.
On October 22, 2026, the U.S. Court of Appeals for the Fifth Circuit vacated a preliminary injunction in Jackson Federation of Teachers v. Fitch and ordered reassignment. The majority opinion described fabricated facts in an AI-assisted order and a nonexistent case that remained in a corrected version. The decision also rested on failures in the legal analysis.
The judge’s signature is the point at which a draft becomes an exercise of state power. Before issuing an AI-assisted ruling, the court must check each cited authority and quoted rule against its source, and each factual claim against the record. Replacing a flawed draft is not enough if later orders still rely on it. The review must follow everything that inherited the error.
Taken together, the evidence points toward a workable boundary rather than a ban. Courts and justice institutions are beginning to use AI, and some are writing rules that preserve verification, disclosure, independent evidence, contestability, and human decision authority. The relevant question is increasingly less whether AI touched the process and more whether the institution preserved the accountability chain when AI did.
Accountability Analysis
Principles 2, 4, and 7 carry most of the argument, with Principles 1 and 6 becoming visible when something fails. Principle 2 keeps capability separate from authority. A useful prediction, match, summary, or draft can deserve weight without acquiring the legal power of the institution that uses it.
Principle 4 asks whether review can actually change the outcome. The judge, investigator, prosecutor, defense lawyer, or other authorized reviewer needs enough evidence and context, enough time and competence to disagree, and practical power to stop or alter the machine-assisted path. State v. Miles adds the affected person’s side of the same idea: when machine use matters to the prosecution, meaningful challenge may require disclosure about the tool and how it entered the case.
Principle 7 governs the authority jump when liberty is finally determined. The criminal sentence is Reserved. Independent human judgment is part of the legitimacy of the act. A judge can rely on machine assistance while remaining responsible for the reasons, the proportionality, the individualized judgment, and the final exercise of state power.
When the chain fails, follow the complaint backward. Who used the tool? Who decided its boundary? Who had authority to disregard it? Who can correct the individual result? Who can change the institutional rule? Who faces consequences when the answer was never checked? The chain can branch. It cannot dissolve.
Find the authority jump
Imagine software suggests a possible match between a photograph and a person. Identify the point where that suggestion might become government action. What independent evidence would support the action? Who would evaluate it? What record would allow the decision to be reviewed? Use the authority jump to distinguish the machine’s contribution from the human decision that must be defended.
Accountability Verdict
Law should use AI where it improves search, organization, consistency, accessibility, and the quality of information available to human decision makers. An institution that rejects useful tools merely because they are machines may protect tradition more effectively than justice.
The opposite mistake is more dangerous when the State can coerce. Machine output should cross into public authority only through a boundary the institution has deliberately written and a human decision owner who can independently defend the result.
For policing, a machine-generated lead should remain a lead until independent evidence justifies coercive action. For adjudication, AI assistance should remain visible enough to verify and challenge. For sentencing, the human signature must represent an actual human judgment rather than a ratification of a score.
The simplest rule is also the oldest one in the courtroom: if the State is going to take something from a person, someone with lawful authority must stand behind the decision.
Chapter Thirteen moves from the courtroom to the battlefield. There the authority jump can happen faster, the decision may be irreversible in seconds, and the person affected may never have an opportunity to appeal. The question becomes the most severe version yet: can the power to choose a target and use lethal force ever be delegated to a machine?
Sources and Notes
1. Supreme Court of Wisconsin, State v. Loomis, 2016 WI 68, No. 2015AP157-CR, July 13, 2016. https://www.wicourts.gov/sc/opinion/DisplayDocument.pdf?content=pdf&seqNo=171690. Accessed September 11, 2026. The court held that a COMPAS risk assessment could be considered at sentencing if used with the opinion’s limitations and cautions. It stated that risk scores may not determine whether a person is incarcerated, may not determine sentence severity, and may not be the determinative factor in deciding whether community supervision is safe and effective. The court also discussed the proprietary nature of COMPAS and required cautions around its use. Source record: DA-CH12-S1. Archive reference: pending.
2. Detroit Police Department, Manual 307.5, “Facial Recognition,” policy approved by the Detroit Board of Police Commissioners May 30, 2024; and Williams v. City of Detroit, Settlement Agreement, U.S. District Court for the Eastern District of Michigan, No. 21-cv-10827, filed June 28, 2024. https://detroitmi.gov/sites/detroitmi.localhost/files/2024-06/307.5%20Facial%20Recognition%20Policy%20Approved_05302024.pdf and https://www.aclu.org/cases/williams-v-city-of-detroit-face-recognition-false-arrest?document=Settlement-Agreement. Accessed September 11, 2026. DPD policy requires additional independent reliable evidence before an arrest or arrest-warrant request following facial-recognition use and bars reliance solely on a facial-recognition lead plus a lineup identification. The settlement resolved disputed claims without an admission of liability. The ACLU case record states that Robert Williams was wrongfully arrested in 2020 following an incorrect facial-recognition match and detained for about thirty hours. Source record: DA-CH12-S2. Archive reference: pending.
3. Supreme Court of New Jersey, State v. Tybear Miles, A-41-24 (090275), decided June 24, 2026. https://www.njcourts.gov/cases/a-41-24 and https://law.justia.com/cases/new-jersey/supreme-court/2026/a-41-24.html. Accessed September 11, 2026. The court held that facial-recognition discovery is case-specific and required the State to produce non-proprietary information identifying the FRT tools and materials used and explaining how they were used in the investigation and prosecution, even when prosecutors do not intend to introduce the facial-recognition result at trial. It reversed without prejudice the order requiring proprietary source code because the record had not yet established a particularized need. Source record: DA-CH12-S3. Archive reference: pending.
4. TRI/NCSC AI Policy Consortium for Law & Courts, “Judicial Use of Generative AI: Lessons Learned,” National Center for State Courts, March 13, 2026. https://www.ncsc.org/resources-courts/judicial-use-generative-ai-lessons-learned. Accessed September 11, 2026. The project conducted 13 confidential one-hour interviews in October and November 2025 with state and federal judges serving in 10 states. The judges were early adopters, so the interviews are not a representative survey. NCSC reports unanimous agreement among participants that judges must remain the ultimate human decision makers and describes a popular analogy: treat generative AI like a judicial intern and delegate only questions where the judge already knows the general answer and wants additional support. Source record: DA-CH12-S4. Archive reference: pending.
5. Anika Jaitley, Daniel W. Linna Jr., Xavier Rodriguez, V.S. Subrahmanian, and Siyu Tao, “Artificial Intelligence in Federal Courts: A Random-Sample Survey of Judges,” The Sedona Conference Journal, Volume 27, forthcoming 2026, preprint released March 2026. https://www.thesedonaconference.org/publication/Artificial_Intelligence_in_Federal_Courts. Accessed September 11, 2026. Researchers selected a stratified random sample of 502 federal judges and received 112 responses, a 22.3% response rate. More than 60% of respondents reported using at least one listed AI tool in judicial work, 22.4% reported weekly or daily use, and about 38% reported no use. The response rate limits population-wide inference. Source record: DA-CH12-S5. Archive reference: pending.
6. Judicial Branch of California, Standard 10.80, “Use of generative artificial intelligence by judicial officers,” California Standards of Judicial Administration, effective September 1, 2025; and Rule 10.430, “Generative artificial intelligence use policies.” https://courts.ca.gov/cms/rules/index/standards/Standard10_80 and https://courts.ca.gov/cms/rules/index/ten/rule10_430. Accessed September 11, 2026. Standard 10.80 directs judicial officers using generative AI in adjudicative roles to protect confidential information, avoid unlawful discrimination, take reasonable steps to verify accuracy and correct erroneous or hallucinated output, remove biased or harmful content, and consider disclosure for public-facing AI-created content. Rule 10.430 permits courts to impose more restrictive policies. Source record: DA-CH12-S6. Archive reference: pending.
7. District of Columbia Court of Appeals, Douglas v. Deutsche Bank National Trust Company, No. 24-CV-1099, Published Order filed September 3, 2026. https://www.dccourts.gov/sites/default/files/2026-09/Douglas%20v.%20Deutsche%20Bank%20Nat%27l%20Trt%20Co.%2024-CV-1099%20ORDER.pdf. Accessed September 11, 2026. The court found four nonexistent case citations in a filed brief, struck the brief, addressed the continuing duties of the signing attorneys, and referred the matter for disciplinary consideration. This case was introduced in Chapter Four and returns here as the court’s own accountability example. Source record: DA-CH12-S7. Archive reference: pending.
8. Supreme Court of New Mexico, State v. Sandoval; In re Stephen D. Aarons, No. S-1-SC-40845, Dispositional Order of Direct Contempt, September 9, 2026. https://law.justia.com/cases/new-mexico/supreme-court/2026/s-1-sc-40845.html. Accessed September 11, 2026. The court found that a murder-appeal brief prepared with ChatGPT contained false testimony from fabricated witnesses, false testimony attributed to real witnesses, and misrepresented legal authority. It held counsel in direct contempt, referred the matter to the Disciplinary Board, barred counsel from appearing before the court pending discipline, struck all briefing, appointed new defense counsel, and imposed a $5,000 sanction payable to the State Bar of New Mexico Client Protection Fund. Reuters reported on September 11 that the lawyer said he had used the tool to summarize trial transcripts and did not understand the degree to which it could fabricate facts. The order is nonprecedential under New Mexico rules. The filing involved ChatGPT, an OpenAI product. OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH12-S8. Archive reference: pending.
9. Courts and Tribunals Judiciary, Sir Colin Birss, “Position Statement on the Judicial use of Artificial Intelligence (AI),” October9,2026. https://www.judiciary.uk/guidance-and-resources/position-statement-on-the-judicial-use-of-artificial-intelligence-ai/ Accessed October10,2026. Official judicial governance statement, not a court judgment or independent efficacy evaluation. See Core principles2 and4, Tools9–10, and Governance and approval11–13. The operational questions are author analysis. Source record: DA-CH12-S9. Update 081. Archive reference: pending.
10. California Legislature, SB 574, Chapter 858, Statutes of 2026, approved and filed September 30, 2026; Business and Professions Code Section 6068.1 and Code of Civil Procedure Sections 128.7(b)(2) and 1282.1. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB574. Official bill-text retrieval returned 403; chaptered legislative text was inspected through https://legiscan.com/CA/text/SB574/id/3461240. Governor of California, September 30, 2026 signing announcement: https://www.gov.ca.gov/2026/09/30/californias-nation-leading-ai-framework-just-got-stronger-governor-newsom-signs-more-first-in-the-nation-worker-protections-and-more/. California Courts Newsroom, “Lawyers Face Stronger AI Review Responsibilities Under New California Law,” October 1, 2026: https://newsroom.courts.ca.gov/news/lawyers-face-stronger-ai-review-responsibilities-under-new-california-law. Accessed October 12, 2026. The official signing slate confirms the bill number and enactment; the Courts Newsroom identifies the requirements as beginning in 2027. The citation provision covers all citations in filed papers, including AI-supplied citations. The output-review provision uses a reasonable-steps standard. Arbitration decision-making and protected-information restrictions remain distinct. Enactment does not establish compliance or effectiveness. The connection to retained professional responsibility is the author’s analysis. Source record: DA-CH12-S10. Update 090. Archive reference: pending.
11. United States Court of Appeals for the Fifth Circuit, Jackson Federation of Teachers v. Fitch, No. 25-60496, opinion filed October 22, 2026, pp. 3-8 and 27-28. Official URL: https://www.ca5.uscourts.gov/opinions/pub/25/25-60496-CV0.pdf . Readable copy of the court opinion: https://www.courthousenews.com/wp-content/uploads/2026/10/fifth-circuit-mississippi-dei-injunction-reversed.pdf . Accessed October 23, 2026. Official listing verified the date and case; the official PDF endpoint was inaccessible, so the filed opinion was read in the mirrored PDF. Vacatur and reassignment also involved substantive legal defects; this is not a general ban on judicial AI or removal from the bench. Judge Haynes concurred in the final disposition, not the entire majority opinion. Second paragraph is author analysis. Source record: DA-CH12-S11. Update 116. Archive reference: pending.