If you took the California Bar Examination in February 2025, you would have encountered utter chaos. This was the California Bar Exam’s first attempt in decades at a hybrid in-person and remote delivery, and the result was an immediate collapse upon launch: test-takers were booted from the exam software, interfaces lagged and threw errors, and exam questions were even riddled with typos and logical flaws. The fuse for this debacle was a financial deficit. The California Bar Exam is administered by the State Bar of California—an administrative arm of the California Supreme Court—and is held twice a year. Historically, like most US states, California procured its multiple-choice questions from the National Conference of Bar Examiners (NCBE) through the Multistate Bar Examination (MBE). In 2024, facing a $22 million fund deficit, the State Bar terminated its contract for NCBE multiple-choice questions, hired Kaplan for $8.25 million to develop new questions, and brought in Meazure Learning to administer exam delivery.
Beyond the technical glitches, the subsequent revelations regarding the provenance of the exam questions triggered an even larger crisis of credibility. According to an inventory disclosed by the State Bar after the exam, out of the 171 scored multiple-choice questions, 100 were procured from Kaplan, 48 were pulled from the First-Year Law Students’ Examination (commonly known as the “baby bar” for students at unaccredited law schools), and the remaining 23 were generated directly using AI by ACS Ventures. ACS Ventures was the psychometric vendor hired by the State Bar, whose role was to use statistical methods to evaluate exam reliability and establish passing scores. A referee tasked with independently validating test quality had instead slipped 23 AI-generated questions into the official exam, without any prior report to the Committee of Bar Examiners, the Supreme Court, or the public.
This directly violated the terms of the engagement contract. The agreement between Kaplan and the State Bar explicitly stated that the involvement of generative AI should remain de minimis—that is, kept to an absolute minimum—and that AI tools could only be used to augment limited portions of existing human-authored work. Filings submitted by the State Bar to the California Supreme Court in April further confirmed that after receiving the AI-generated drafts, ACS performed only preliminary formatting and styling adjustments, without proofreading legal accuracy, screening for bias, or evaluating suitability for testing entry-level attorneys. Although the State Bar maintained that all questions were ultimately vetted by content validation panels and subject matter experts, that the combined multiple-choice section exceeded the psychometric reliability target of 0.80, and denied any conflict of interest, the California Senate Judiciary Committee still characterized the exam as an “unmitigated disaster” in its legislative analysis, while the author’s office explicitly noted that multiple questions were never substantively reviewed by practicing attorneys before being presented to examinees.
The fallout concluded with intensive remedies and litigation. The State Bar arranged for approximately 85 affected examinees to retake the exam in March; a review uncovered system export defects in the scratch paper notes of 431 unsuccessful candidates, leading to an immediate reversal that passed 8 of them; the California Supreme Court intervened urgently, ruling that the July 2025 examination return entirely to an in-person format and reinstate procurement of the NCBE question bank; and a class-action lawsuit filed by examinees against Meazure reached a settlement in August 2026.
Confronted with this crisis of credibility in testing, the California Legislature opted to enact a disclosure statute rather than issue a technological ban. AB 1651, chaptered as Chapter 116 of the Statutes of 2026, was authored by Assemblymember Diane Dixon, signed by the Governor on August 22, 2026, and adds § 6060.15 to the California Business and Professions Code, taking effect on January 1, 2028. The entire legislative process met no formal opposition, with the California State Assembly passing it unanimously by a vote of 72 to 0.
The core logic of the bill is built upon four tightly interlocking provisions. On definitions, § 6060.15(a)(2) defines AI-generated content as visual and text content generated entirely or partially by generative AI—partial generation counts, with no percentage threshold established. On scope of application, § 6060.15(b) and (c) delineate disclosure obligations across two categories of assets: first, all exam questions, practical skills tests, model answers, and grading rubrics developed by or at the explicit direction of the State Bar; second, official study and preparation materials compiled, endorsed, published, or distributed by the State Bar, covering sample questions, model answers, study guides, and explanations. On disclosure form, § 6060.15(e) mandates that AI disclosures for official examinations be posted on the State Bar’s website at least 60 days in advance, while prep materials must feature the disclosure statement directly on their covers. The central provision of the entire statute is concentrated in § 6060.15(d), which contains just a single sentence:
Subdivisions (b) and (c) apply regardless of whether the artificial intelligence-generated content is revised or reviewed by a natural person.
This rule establishes a clear principle: regardless of whether AI-generated content is subsequently revised or reviewed by a natural person, the disclosure obligation applies. In other words, proofreading and rewriting by human experts cannot wash away the material’s original AI provenance.
At the same time, the institutional boundaries of this legislation remain restrained. The statute contains no penalty clauses such as fines, establishes no specialized administrative enforcement mechanism, and grants no private right of action to examinees. During the bill’s legislative deliberations, its scope underwent a narrowing revision: the State Bar had expressed concerns to the committee that it had no way of knowing whether external contractors (such as the 100 questions previously delivered by Kaplan) utilized AI tools during their internal workflows, making mandatory disclosure unworkable if applied to such content; the bill’s author accepted amendments accordingly, restricting the statute to materials developed under the State Bar’s own direction.
AB 1651 also builds upon a predecessor statute. In SB 253, signed earlier on October 6, 2025, California had broadly required the examining committee to provide notice whenever AI was used in developing exam questions or grading, but left unspecified the party responsible for notice, the exact time window, and the concrete medium. AB 1651 supplied these operational details: specifying public vehicles such as the official website and document covers, locking in the 60-day pre-exam delivery milestone, extending oversight to study materials, and incorporating the cornerstone principle that human review does not exempt disclosure.
Shifting focus from licensing examinations to the daily practice of law, the very same profession operates under starkly contrasting rules regarding AI. In legal practice, attorneys are not required to proactively disclose their use of AI tools to clients or courts. The American Bar Association established this benchmark in Formal Opinion 512, issued on July 29, 2024: lawyers may utilize generative AI as a tool to assist their work, provided they independently verify the output and retain professional judgment. The entire ethical opinion holds attorneys fully responsible for the work product ultimately submitted, imposing no general duty to disclose the use of AI.
Judicial disciplinary cases corroborate this liability attribution mechanism. In Mata v. Avianca, heard in the U.S. District Court for the Southern District of New York, attorneys cited six bogus judicial opinions fabricated by ChatGPT in a brief submitted to the court. Presiding Judge Castel imposed a $5,000 sanction under Rule 11 of the Federal Rules of Civil Procedure. In his ruling, the judge pointed out that the duty of reasonable inquiry required by the procedural rules cannot be outsourced to a technological tool, and that an attorney’s signature without verification constituted sanctionable misconduct. The court punished the failure to fulfill the verification duty; the use of the AI tool itself was not sanctioned. The human attorney’s signature and affirmation complete the absorption of liability for the tool’s output.
Juxtaposing the rules across both domains makes the contrast unmistakable. In everyday practice, the profession follows the logic of use, verify, and take responsibility—human professional oversight negates the duty of disclosure. In licensure examinations, California AB 1651 makes clear that if AI is used in generating questions, it must be disclosed in advance—revisions and review by human experts cannot erase the original provenance. Within the exact same professional domain, everyday production and gatekeeping credentialing diverge along opposite paths of accountability.
Observing governance rules across industries dealing with AI, one cannot compress them into a single, monotonically increasing ladder of strictness. Outright bans and mandatory disclosures answer two independent sets of questions: the first ruler measures whether use is allowed, while the second ruler measures whether use must be disclosed. The T0 through T4 tier labels used in this article serve merely as reference coordinates to unpack mechanistic logic; they do not represent any official taxonomy, nor do they form a one-dimensional ranking of strictness.
The first ruler is whether use is allowed. Institutions situated at T0 (Pledge of Non-Use) require creators to provide written affirmations confirming that AI tools were not used at any stage. For example, ASME journals require submitting authors to confirm that AI was not involved in the development of original scholarly content; China’s software copyright registration, effective March 15, 2026, requires applicants to handwrite a pledge stating no AI was used, with dishonest statements leading to inclusion on a dishonest entities list. Institutions at T1 (Complete Ban) systematically bar AI involvement outright. The NCBE, the national testing organization, stated in its Summer 2025 Letter from the Chair that AI is not used in developing official exam questions or scoring for either the current bar exam or the NextGen UBE; the CFA Institute likewise states that it does not use AI to independently generate credentialing content, score exams, or make high-stakes certification decisions.
The second ruler is whether use must be disclosed, which unfolds only when AI use is permitted. Situated at T2 (Use with Accountability, No Disclosure) is the prevailing model in modern professional practice: AI may be used, but human professional judgment must review the output and assume full liability, generally with no requirement for external disclosure. The legal profession’s stance was set by ABA Formal Opinion 512; the American Academy of Actuaries similarly noted that actuaries cannot deflect responsibility onto model outputs without independent verification, anchoring professional liability squarely on the credentialed practitioner without imposing a blanket AI disclosure obligation toward clients.
T3 (Use with Conditional Disclosure) permits the use of AI, triggering disclosure only under specific conditions, and commonly includes a human review exemption. California’s AB 3030 mandates that clinical communications generated by generative AI directed to patients must include a disclaimer, but this disclosure requirement is automatically waived if the content is reviewed and read by a licensed health care provider. Article 50(4) of the EU Artificial Intelligence Act similarly provides that for text published with the purpose of informing the public on matters of public interest, disclosure is exempted where the content has been subject to human review or editorial control and a natural or legal person holds editorial responsibility.
At the other end of the second ruler lies T4 (Use with Mandatory Disclosure, No Human Review Exemption): if it was used, it must be disclosed, and human review does not eliminate this obligation. Much of academic publishing has settled here since 2023. The International Committee of Medical Journal Editors (ICMJE) requires authors to disclose the use of AI-assisted technologies upon submission; ICLR makes clear that any use of large language models must be disclosed; while most academic guidelines exempt basic language editing, ACM updated its policy in May 2026 so that writing assistance no longer requires mandatory disclosure. In statutory comparators, China’s Measures for the Labeling of AI-Generated and Synthetic Content, effective September 1, 2025, contains no human-review exemption clause across its entire text; California’s AB 2355 requires political advertisements containing AI-generated content to carry explicit labels.
Placing California’s AB 1651 within these two rulers makes its position evident. In the professional credentialing examination space, the NCBE and CFA follow T1 under the first ruler, excluding AI from exam development and grading workflows; the United States Medical Licensing Examination (USMLE), Uniform CPA Examination, Project Management Professional (PMP) certification, and NCLEX-RN licensing examination have no statutory AI exam question disclosure requirements. California’s AB 1651 follows T4 under the second ruler: AI use is permitted, but as long as it is directed by the State Bar, it must be publicly disclosed, and subsequent expert review cannot waive that requirement. It is not the world’s first T4—academic publishing led the way in 2023; but it is the first statutory T4 in the professional credentialing examination space, and within T4, it is even stricter than mainstream academia, containing neither a de minimis exemption nor a safe harbor for basic language and grammar polishing.
When drafting an internal AI policy for technical teams, the most common pitfall is attempting to cover all workflows with a single uniform rule. The first step in structuring a policy is to bifurcate workflows into everyday operational production (Practice) and credentialing checkpoints (Gate).
Practice accounts for the vast majority of an engineer’s regular time: completing code via Copilot or Cursor in pull requests, drafting internal RFCs, summarizing support tickets, or iterating on specific features. Here, the core focuses are deliverable quality and clear ownership, corresponding to T2: AI may be used, but must be human-verified, with you bearing full responsibility for what gets merged—structurally isomorphic to ABA Formal Opinion 512’s framing of everyday legal practice. Workflows of this type do not need an AI disclosure badge stamped on every single file.
Gate represents the checkpoints where an organization provides external or internal certification: publishing external launch notes, preparing SOC 2 audit materials, releasing model cards with benchmark figures, authoring technical blog posts that represent an official stance, take-home assignments for hiring, or promotion dossiers. Here, material provenance and institutional credibility are paramount; post-hoc human review cannot erase generative origin. The most frequent misstep is mapping the intuition of Practice onto Gate, assuming that because material was double-checked, AI needs no mention. This is precisely the path California’s AB 1651 rejected on the bar examination side.
Gate workflows must first choose a ruler: T1 (ban), ensuring no AI is used in this credentialing material, mirroring the approach of NCBE and CFA in high-stakes certification; or T4 (permit with mandatory disclosure), where human verification does not exempt disclosure. If T4 is chosen, the policy needs only adjust two knobs. Trigger threshold: whether all usage requires disclosure, or only substantial contributions trigger it, and whether spelling and grammar tools are exempt. Effect of human review: whether review waives disclosure, as in California AB 3030 or Article 50(4) of the EU AI Act; or whether provenance transparency must be maintained even after verification, as in California AB 1651 or China’s content labeling measures.
When drafting or revising a team’s AI policy, begin by taking inventory of deliverables: categorize most routine engineering into Practice—empower usage, make mergers accountable; isolate assets directly tied to external endorsements and internal certifications into Gate, configuring them with T1 or T4 as needed. Distinguishing between these two accountability logics keeps daily engineering moving while ensuring external commitments hold firm.