Pennsylvania’s Other Frontier-AI Bill Would Set a 10^26 Threshold and Catastrophic-Risk Guardrails

HB 2829 would create a separate safety, transparency and reporting framework for certain frontier-model developers. It has been introduced and heard in committee, but it is not Pennsylvania law.

Pennsylvania is considering more than one approach to frontier-AI governance. House Bill 2829, the proposed Frontier Artificial Intelligence Safety, Transparency and Accountability Act, would establish a detailed state framework for developers of the largest general-purpose models.

The bill was introduced and referred to the House Communications & Technology Committee on October 5, 2026. The committee held a public hearing on October 19 that listed both HB 2829 and the separate HB 2800. But the formal action history still shows no vote, committee report, passage or later printer’s number. HB 2829 is therefore a pending proposal, not an active Pennsylvania compliance regime.

A more specific frontier-model threshold

The controlling text, Printer’s No. 4002, defines a frontier model as a foundation model trained using more than 10^26 integer or floating-point operations. The calculation includes the original training run and any subsequent fine-tuning, reinforcement learning or other material modifications the developer applies to a preceding foundation model.

That is only one part of the bill’s coverage framework. A “frontier developer” is a person that trains or initiates training of such a model. A “large frontier developer” is a frontier developer that, together with affiliates, had more than $500 million in annual gross revenue in the preceding calendar year.

The bill would apply only to frontier models developed, deployed or operating in whole or in part in Pennsylvania. It says it should not be read to prohibit lawful AI research, development, testing or evaluation, or to impose a duty solely because a person works with an AI model that is not a frontier model.

Different obligations for different developers

HB 2829 does not assign every proposed duty to every frontier developer.

Beginning January 1, 2028—if the bill were enacted—large frontier developers would have to file a current disclosure statement with the Pennsylvania Emergency Management Agency before developing, deploying or operating a covered model in the Commonwealth. They would also have to publish, implement, review at least annually and, as appropriate, update a frontier-AI framework describing how they assess and mitigate catastrophic risks, use third parties, secure unreleased model weights, respond to critical safety incidents and govern internal use of covered models.

The proposal would require an annual independent audit for large frontier developers beginning January 1, 2028, or 90 days after a developer first qualifies as large, whichever is later. A developer that first qualifies as large after the effective date would have 90 days to comply with Chapter 5. The bill specifies independence protections, audit-report content, retention rules and publication of a high-level summary plus a redacted report.

Some proposed requirements extend to frontier developers more generally. Before or alongside deployment of a new or substantially modified frontier model, a frontier developer would have to publish a transparency report with information such as supported languages and output modalities, intended uses and generally applicable restrictions or conditions. Large frontier developers would have additional report-summary requirements about catastrophic-risk assessments and third-party evaluators.

The bill’s catastrophic-risk line

The proposal defines catastrophic risk around foreseeable, material risks that a frontier model could materially contribute to death or serious injury to more than 50 people, or more than $1 billion in property damage or loss, from a single incident. The qualifying incident must involve expert-level assistance in creating or releasing a chemical, biological, radiological or nuclear weapon; conduct without meaningful human oversight, intervention or supervision that constitutes a cyberattack or would constitute murder, assault, extortion or theft if committed by a human; or evasion of the developer’s or user’s control.

For large frontier developers, HB 2829 would prohibit deployment when the required assessments, mitigation work and the developer’s framework show that deployment would create an unreasonable risk of crossing that harm threshold. It would also require summaries of internal-use catastrophic-risk assessments to be transmitted to the agency on a recurring schedule.

The bill separately defines “critical safety incidents.” A frontier developer would have to report a qualifying incident to the agency and Attorney General within 72 hours after learning facts sufficient to reasonably believe it occurred. If an incident poses an imminent risk of death or serious physical injury, the bill calls for disclosure within 24 hours to an appropriate authority, as required by law.

Enforcement, interoperability and employee reporting

HB 2829 would authorize the Attorney General to seek civil penalties, injunctive relief or declaratory relief. The bill would set maximum civil penalties of $1 million for a first violation and $3 million per subsequent violation for specified large-developer violations. It says it would not create a private right of action.

The proposal also includes a conditional interoperability route. A frontier developer could declare an intent to comply through a federal law, regulation or guidance document, but only after the agency designates materials it finds substantially equivalent or stricter for the relevant requirements and only to the extent the developer actually meets those designated federal standards. That is not a blanket federal-compliance exemption.

Its employee provisions would bar a frontier developer from preventing or retaliating against certain safety-related disclosures by covered employees. Large frontier developers would also have to provide an internal anonymous disclosure process for certain safety or compliance concerns.

Finally, the bill would make regulation of AI frontier models an exclusive Commonwealth power, barring political subdivisions from enacting or enforcing requirements that specifically regulate them.

What the October hearing does—and does not—show

The official committee archive confirms that a public hearing on mitigating risks from frontier AI models was held on October 19 and listed HB 2829 alongside HB 2800. It does not show a vote, recommendation, amendment or committee endorsement.

Without separately reviewing testimony, a transcript or other hearing materials, it would be premature to characterize what participants argued or what conclusions the committee reached.

For companies, HB 2829 is worth watching because it sketches a concrete state model for frontier-AI safety frameworks, audits, deployment decisions and incident reporting. But the immediate legal point is simpler: it remains proposed legislation. No Pennsylvania duty, deadline or penalty in HB 2829 is currently operative.

Courts are drawing a line between AI-fabricated authorities and actual sanctions

Three 2026 appellate decisions show the consequences of filing fabricated or misrepresented legal authorities, whether or not a court establishes that AI produced them. The consequences still vary sharply by case: Oregon imposed a $250 payment after a litigant admitted relying on AI-generated citations; Idaho warned that similar conduct could warrant sanctions but imposed none in the appeal; and a Hawaii appellate court identified Rule 11 violations in an order marked “not for publication in West’s Hawaii Reports and Pacific Reporter” without imposing a monetary sanction.

The cases offer a more precise lesson than “courts are banning AI.” They show courts focusing on verification, candor, and the burden that fabricated authority places on the judicial system.

Oregon: a reduced but real payment

In R. M. v. Mohr, 353 Or. App. 182 (2026), the Oregon Court of Appeals considered a pro se respondent’s appeal of a continuing restraining order. The court’s opinion included a separate discussion titled “Artificial Intelligence and the Use of Fabricated Cases.”

The respondent’s opening brief cited three fabricated cases. After the court issued an order to show cause addressing whether the brief should be stricken and whether sanctions should be imposed, the respondent acknowledged that the citations resulted from reliance on artificial intelligence, apologized, and accepted responsibility.

The court said it would proceed with the merits but would disregard the fabricated authorities and the support those authorities purported to provide. It also explained that fabricated authorities divert court resources into identifying false citations, considering show-cause proceedings, and deciding whether sanctions are appropriate.

The court ultimately affirmed the restraining order and directed the respondent to pay $250 to the Appellate Court Services Division of the Oregon Judicial Department. That amount was reduced from the $1,500 contemplated in the show-cause order, with the court citing the respondent’s acknowledgment of the errors, their severity, and the respondent’s limited financial resources.

The decision is important for what it does not say. It does not treat AI use itself as the violation. The court’s concern was the filing of fabricated legal authorities and the resulting burden on the court.

Idaho: a warning without an additional monetary sanction

The Idaho Supreme Court addressed a similar problem in Posey v. Bushnell, an opinion filed September 22, 2026. The underlying appeal arose from a defamation judgment, but the court separately addressed the risks of relying on generative AI in appellate briefing.

The court stated that AI reliance does not excuse inaccuracies, whether the litigant is represented or proceeding without counsel. It dismissed nine issues because of deficiencies in the appellant’s briefing. The court expressly said those dismissals were not sanctions under Idaho Appellate Rule 11.2. It also said the briefing reflected a pervasive failure to verify factual and legal representations.

The court also said the filing obligations in Idaho Appellate Rule 11.2 require a signer to certify, after reasonable inquiry, that a filing is well grounded in fact and warranted by existing law. The court wrote that the conduct could warrant sanctions under the rule.

But it did not impose an additional monetary sanction in the appeal. The court declined to award attorney fees sua sponte, noting that the prevailing party had not requested them and that the appellant was already subject to a substantial judgment. The court expressly said that its decision should not minimize the seriousness of the conduct, while exercising its discretion to impose no sanction in that appeal.

That distinction matters. A warning that conduct could support sanctions is not the same as a sanctions award.

Hawaii: Rule 11 language, but no monetary penalty in the order

In U.S. Bank Trust National Association v. Thomas, a Hawaii appellate summary disposition order filed October 8, 2026, the court identified fabricated and unsupported authorities in the appellant’s brief. The order states that several cited cases did not exist and that other authorities were cited for propositions they did not support.

The court characterized those errors as violations of Hawaii Rules of Civil Procedure Rule 11(b)(2), applied through Hawaii Rules of Appellate Procedure Rule 2.1(a). It also criticized the answering brief for failing to call the court’s attention to the hallucinated authorities.

The order was marked “not for publication in West’s Hawaii Reports and Pacific Reporter.” It affirmed the foreclosure judgment, but the order did not impose a monetary sanction.

The court identified Rule 11(b)(2) violations, but did not find that Thomas used AI or hold that AI use itself violates a rule. The relevant problem was the submission of nonexistent or materially unsupported authorities in a court filing.

The emerging rule is verification, not an AI ban

Read together, the three decisions suggest a practical hierarchy:

• Fabricated citations create an immediate credibility and briefing problem. Courts may disregard arguments supported by authorities that do not exist.
• A court can impose sanctions, but the result is case-specific. Oregon imposed a reduced $250 payment; Idaho and Hawaii did not impose a monetary sanction in the decisions discussed here.
• Admissions and corrective conduct can matter, but they do not erase the burden created by the filing. Oregon considered the respondent’s acknowledgment and resources when setting the reduced amount.
• Procedural posture controls the remedy. Idaho dismissed issues and declined fees; Hawaii addressed the problem in an order marked “not for publication in West’s Hawaii Reports and Pacific Reporter”; Oregon proceeded on the merits while separately imposing a payment.
• The opinions do not establish that AI assistance alone violates a rule. Their focus is on accuracy, reasonable inquiry, candor, and the consequences of filing unverified legal material.

For lawyers and self-represented litigants, the operational takeaway is straightforward: every case, quotation, statute, and rule supplied by an AI system must be independently checked against an authoritative source before it enters a brief. A disclaimer that a tool was used will not substitute for that review.

The cases also caution against collapsing different outcomes into a single “AI sanctions” trend. Oregon involved an actual payment. Idaho expressly declined a monetary sanction while warning that sanctions could be warranted. Hawaii identified Rule 11 violations in an order that was not for publication and imposed no monetary penalty. Those differences are part of the story, not details to omit.

Sources

• Oregon Court of Appeals, R. M. v. Mohr, 353 Or. App. 182 (2026): official opinion
• Idaho Supreme Court, Posey v. Bushnell, Docket No. 52072 (Sept. 22, 2026): official opinion
• Hawaii Intermediate Court of Appeals, U.S. Bank Trust National Association v. Thomas, CAAP-24-0000837 (Oct. 8, 2026): official order

House Oversight opens inquiry into Delaware’s proposed AI-company legal shield

The House Committee on Oversight and Government Reform has opened an inquiry into Delaware’s effort to develop a new corporate form for companies operated by artificial intelligence agents.

The committee’s October 26 release and letter to Delaware Governor Matt Meyer describe the initiative as an Artificial Intelligence Company, or AIC. The documents do not establish that Delaware has enacted the proposed entity or that the framework gives anyone a current liability shield. They show a federal oversight request focused on the proposal’s design, safety analysis, outside contributors, and potential conflicts.

What the committee requested

Chairman James Comer’s letter asks Delaware to provide documents and communications covering policy development from January 1, 2025, to the present. The request seeks:

• communications about the AIC proposal and artificial intelligence liability;
• ethics, consulting, and other agreements involving outside contributors;
• materials concerning potential safety concerns, harms, and efforts to limit liability; and
• drafts and comments relating to AIC proposals considered by state officials.

The committee asked Delaware to provide the information as soon as possible and no later than November 9, 2026.

The letter invokes the committee’s oversight authority and says the information could inform possible federal legislation, including proposals to limit state AI regulatory initiatives. That is a request for information and a statement of congressional interest, not a federal legal finding about Delaware’s proposal.

What Delaware’s proposal is

According to the committee’s letter, Delaware’s Secretary of State has described an AIC as a corporate entity run day to day by an AI agent while providing the human or entity capitalizing it with protections from potential harms caused by the controlling AI, subject to exceptions such as inadequate capitalization, fraud, or willful violations of law.

The letter traces the project to Delaware House Joint Resolution 7, signed in July 2025. That resolution directed the state’s Artificial Intelligence Commission and Secretary of State to create a regulatory sandbox framework for testing innovative technologies that use agentic AI. A December 2025 commission report said a subcommittee was working toward a framework for testing agentic AI in corporate governance.

The House letter says a later draft legislative proposal would create the AIC structure and could allow an AI agent to conduct a wide range of business activity. The proposal was reportedly expected to be introduced in a future General Assembly, but the letter does not identify an enacted bill or current operative law.

Why the inquiry matters

The AIC concept raises a basic corporate-law question: if an AI agent conducts business through a legal entity, how should responsibility be allocated among the entity, its human or organizational backers, the developers of the system, and the people who deploy it?

Traditional corporate forms already separate owners and managers from entity liability, subject to statutes, contracts, tort doctrines, veil-piercing rules, and other limits. An AI-specific entity could add a new layer of uncertainty if its organizing documents or governing statute attempt to allocate responsibility for autonomous decisions before courts have developed a body of case law.

The committee is also focusing on process. The letter asks who contributed to the proposal, what agreements governed those contributions, and what safety or harm analysis Delaware performed. That puts regulatory transparency alongside entity design. The federal inquiry is not limited to whether an AIC could be useful; it asks how the state assessed the risks and who shaped the framework.

What has not happened

The House release and letter do not show that Delaware has enacted an AIC statute, that an AIC has been formed under the proposed framework, or that a court has recognized a new immunity for AI-driven conduct. They also do not establish the committee’s concerns as proven facts.

The committee’s description of the proposal relies in part on material cited in the letter, including a draft reported to be likely to reach the General Assembly in the future. The proposal’s status, text, and eventual legislative path remain separate questions from the federal oversight inquiry.

That distinction matters for companies watching the issue. A regulatory sandbox, a draft proposal, and an enacted entity statute would carry different legal consequences. The October 26 documents establish that Congress is seeking information about Delaware’s work; they do not establish that the proposed liability structure is currently available.

What to watch next

The immediate milestone is Delaware’s response to the November 9 document request. The next substantive state-law development would be an official bill, draft legislation released by the state, or a formal framework that clarifies how AIC liability would operate.

Until then, the most accurate description is a federal oversight inquiry into a proposed Delaware corporate form for agentic AI, with unresolved questions about liability allocation, transparency, and regulatory safeguards.

Sources: House Oversight release; October 26, 2026 letter to Governor Meyer

SpaceX’s SEC filing exposes a $6.1 million potential environmental risk at an AI data-center affiliate

A new SpaceX quarterly filing shows that Mississippi regulators are seeking civil penalties from an affiliate tied to AI infrastructure after alleging that construction began before stormwater-permit coverage was in place.

The filing does not report a final penalty or an executed settlement. It reports a regulatory dispute, preliminary exposure, and discussions about resolving the matters through a single agreed order.

What the filing says

In its Form 10-Q for the quarter ended September 30, 2026, filed October 22, SpaceX says the Mississippi Department of Environmental Quality issued notices of violation to MZX Tech LLC in August. MDEQ alleges that land-disturbing construction activities at three sites occurred before coverage under Mississippi’s Large Construction Stormwater General Permit.

The filing says MDEQ is seeking civil penalties of up to $25,000 per day, with preliminary claimed exposure of approximately $6.1 million across the three sites. SpaceX says it is cooperating with MDEQ and discussing resolution through a single agreed order.

The disclosure is made in the company’s legal-proceedings section. It is not an agency order imposing a final amount, and it does not establish that the alleged violations occurred.

Why the disclosure matters

The dispute is a useful reminder that AI infrastructure carries ordinary environmental and construction-law exposure alongside the better-known issues involving chips, electricity, privacy, and model safety.

The alleged conduct concerns the physical buildout of computing infrastructure. A permit gap at the construction stage can create exposure before a facility is operating at scale, and the potential daily penalty structure makes timing important. The filing also shows why corporate disclosures can be an early source for infrastructure-compliance reporting: they can identify a regulatory matter before a final order or public settlement becomes available.

The filing does not identify the three sites by name or describe the construction timetable in detail. It also does not say whether MDEQ has agreed to reduce the claimed exposure, whether the parties have signed the proposed agreed order, or whether any amount has been accrued as a final liability.

The legal posture is preliminary

Mississippi’s Large Construction Stormwater General Permit is a permitting framework for qualifying land-disturbing activity. The reported notices of violation indicate that MDEQ believes work began before the necessary coverage was obtained. The company’s disclosure describes the agency’s position and the parties’ discussions; it does not resolve the underlying factual or legal dispute.

The stated maximum of $25,000 per day should not be presented as a $25,000-per-day fine already imposed. The approximately $6.1 million figure is described as preliminary claimed exposure across three sites. The eventual amount, if any, could depend on the duration and scope of the alleged violations, the terms of an agreed order, and any defenses or mitigation accepted by the agency.

A broader compliance lesson

For AI companies and their infrastructure partners, the episode points to a practical control problem: environmental permits and construction sequencing can become material AI-business risks even when the underlying technology is software.

The issue is not unique to one company. Data-center projects may involve stormwater, air permits, water use, zoning, wetlands, construction authorization, and local approvals. Those obligations can attach to affiliates, contractors, and site-specific entities rather than the brand most visible to the public. A compliance review that focuses only on model regulation can miss the legal risk created by the facilities that train or serve the model.

The filing also illustrates the limits of public-company disclosure. The 10-Q gives a concrete estimate and identifies the regulator, alleged conduct, and settlement posture. It does not provide a final adjudication or the complete underlying notices. Any article or risk assessment should keep those distinctions intact.

What to watch

The next meaningful document would be the agreed order, if the parties reach one, or a later MDEQ filing that clarifies the claimed violations and any final penalty. Until then, the accurate description is a Mississippi environmental-permitting dispute involving MZX Tech, a SpaceXAI affiliate, with preliminary exposure of approximately $6.1 million.

Primary source: SpaceX Form 10-Q for the quarter ended September 30, 2026

Apple v. Liu: Judge orders targeted discovery in OpenAI hardware trade-secret fight

A federal judge in California has given Apple early access to a defined set of evidence in its trade-secret case involving former Apple employees, OpenAI, and OpenAI’s consumer-hardware venture. The October 13 order is significant because it allows targeted discovery tied to Apple’s preliminary-injunction request while rejecting or deferring broader requests.

The case is Apple Inc. v. Liu, No. 5:26-cv-07078-EJD, in the Northern District of California. Apple filed suit on July 10, 2026. The complaint alleges trade-secret misappropriation and breaches of intellectual-property agreements. Those are allegations, not findings.

The order

Judge Edward J. Davila granted in part, denied in part, and deferred in part Apple’s motion for expedited discovery. The court vacated the October 14 hearing after finding the matter suitable for decision without oral argument.

The court applied the good-cause standard generally used by district courts within the Ninth Circuit for discovery before the ordinary schedule. It considered the relationship to Apple’s preliminary-injunction motion, the breadth and purpose of the requests, the burden on the defendants, and how early the request came in the case.

The court found a clear relationship between at least some of Apple’s requested discovery and the preliminary-injunction motion. It also accepted Apple’s concern that forensic evidence, including logs, metadata, and usage records, could be lost or overwritten. But the court rejected the full scope of Apple’s proposed requests as too broad and ordered a more tailored set.

What Apple can seek

The order requires production of forensic images that defendants already collected in connection with the claims, including images relied upon by a forensic expert. It also permits narrowed requests for materials used to guide or conduct interviews of Apple candidates for OpenAI’s consumer-hardware groups, including interview materials and presentations addressing Apple technology or parts.

Apple may seek LINE conversations between former Apple employees Chang Liu and Alyssa Peng from January 1, 2026, including messages cited in a declaration. The court also authorized depositions of Liu, former Apple executive Tang Yew Tan, and Peng on the topics identified in Apple’s narrowed requests.

The court granted Apple’s request for a corporate deposition under Federal Rule of Civil Procedure 30(b)(6) of OpenAI Foundation, OpenAI Group PBC, and io Products, LLC. At least one permitted topic concerns the roles and responsibilities of designated witnesses and the roles of Liu, Tan, Peng, and another individual identified in the order.

The order sets response and deposition timing tied to the defendants’ production. The three individual depositions are to occur within 30 days after Apple receives the permitted responses, subject to the order’s terms.

What the court declined to accelerate

The court denied expedited treatment for requests concerning OpenAI’s research and development related to metal finishes, power management, or battery functionality for consumer-hardware products. Those matters were left for ordinary merits discovery.

The court also deferred requests that depended on Apple’s definition of “Apple Originated Information.” Judge Davila found that definition amorphous as written and required Apple to confer with the defendants and provide a more tailored definition if it wants to pursue those requests.

That line is important. The order does not give Apple unrestricted access to OpenAI’s hardware-development files. It allows discovery tied to the alleged misappropriation and the preliminary-injunction record, while reserving broader product-development questions for later proceedings.

The factual dispute behind the order

The order describes Apple’s allegation that former employees accessed and used proprietary information to assist OpenAI’s hardware development, gain insider knowledge during recruiting interviews, and instruct new OpenAI hires on how to avoid scrutiny when they leave Apple. Apple also alleges that the corporate defendants acted together to exploit confidential information as OpenAI entered consumer hardware.

The court did not resolve those allegations. It noted competing forensic positions concerning a MacBook and Mac Mini, including disputes over whether a trade-secret file was used and whether devices were restored or deleted. The court expressly declined to weigh the competing expert findings at this stage.

Instead, the court concluded that early discovery could help determine the scope of any alleged misappropriation and address the risk that relevant forensic material might disappear. That is a discovery-management ruling, not a preliminary finding that trade secrets were stolen.

What comes next

The order moves the case into a more concrete evidentiary phase. Apple will receive a defined set of device, interview, messaging, and deposition discovery. The parties must still litigate the underlying trade-secret claims and any preliminary-injunction request on the merits.

The next important developments will be the parties’ production, the permitted depositions, any decision on a more precise “Apple Originated Information” definition, and the court’s eventual treatment of Apple’s requested injunction. For now, the ruling shows a court trying to preserve potentially transient evidence without allowing expedited discovery to become a broad search through a competitor’s product-development files.

Sources: October 13, 2026 expedited-discovery order; Apple Inc. v. Liu docket

Nippon Life’s OpenAI lawsuit puts AI-assisted legal advice under a liability microscope

A federal lawsuit in Chicago is testing a question that has been mostly discussed in the abstract: when a consumer chatbot gives legal-sounding guidance that allegedly helps drive post-settlement litigation, can the AI developer itself be held responsible?

Nippon Life Insurance Company of America sued OpenAI Foundation and OpenAI Group PBC on March 4, 2026, in the Northern District of Illinois. The case is Nippon Life Insurance Company of America v. OpenAI Foundation, No. 1:26-cv-02448. The public docket identifies the case as a diversity action involving contract-related claims. The complaint is not a ruling, and the allegations remain contested.

What the docket shows

The case began with a complaint and jury demand. OpenAI waived service, appeared through counsel, and moved to dismiss on May 15. The district court held that motion in abeyance while a reassignment issue was addressed. Nippon Life filed its response on July 3, and OpenAI filed a reply on July 17. The docket records those filings but does not resolve the parties’ arguments.

The docket shows additional motion practice in September and October. Nippon Life moved to strike two filings on October 2. On October 8, the court entered orders on the motion to strike and on a motion to continue a status conference. The public CourtListener snapshot was updated October 27 and lists October 8 as the last known filing date. That snapshot did not show a merits ruling on the motion to dismiss, although a public docket mirror can lag later court activity.

The case has generated broad claims about AI liability, but the public docket reviewed here does not show a decision adopting those claims.

The theory of liability

Nippon Life’s complaint alleges that ChatGPT’s interactions with a former disability claimant helped fuel an effort to reopen a settled dispute and pursue additional litigation. The complaint asserts theories including tortious interference with contract, aiding and abetting abuse of process, and unauthorized practice of law.

Those allegations move the debate beyond the familiar question of whether a chatbot can produce a wrong answer. The pleaded theory treats the system’s output as part of a chain of conduct that allegedly affected an existing settlement, litigation decisions, and filings. The legal issue is not simply whether a user relied on inaccurate text. It is whether the developer can be liable for the way the product allegedly encouraged or assisted conduct that interfered with another party’s legal rights.

OpenAI has challenged the complaint through its motion to dismiss and reply. The public docket reviewed here does not establish which defenses will control. The litigation may turn on whether the complaint attributes conduct that can support a claim to OpenAI rather than to the user, whether the claims are limited by causation or other doctrines, and how the alleged conduct fits with the company’s warnings and terms of use. Those are issues in the litigation posture, not findings by the court.

Why the case matters

The case presents at least three legal pressure points.

First, it tests the boundary between general-purpose information and individualized legal assistance. A system that summarizes law is different from one that allegedly evaluates a user’s dispute, undermines counsel’s advice, proposes litigation strategy, and drafts filings. If the case reaches that issue, the court may have to assess whether the alleged product conduct supports the plaintiff’s theories.

Second, it separates user responsibility from developer responsibility. Existing sanctions cases have generally focused on lawyers or litigants who submitted fabricated authorities. Nippon Life’s theory points in the other direction: the user’s conduct remains part of the story, but the developer is accused of contributing to the chain of events through the product itself.

Third, the case may make product design evidence relevant. If the dispute reaches discovery, the parties could fight over warnings, refusal behavior, escalation prompts, logging, and the way the system responds when a user presents a live legal dispute. That does not mean the case will produce a broad rule for every chatbot interaction. It does mean that product safeguards could become part of ordinary causation and duty arguments.

What is not decided

The complaint does not establish that ChatGPT practiced law, that OpenAI caused the alleged litigation conduct, or that Nippon Life will recover damages. As of the October 27 CourtListener snapshot reviewed for this draft, the public docket did not show a ruling adopting those propositions. The docket shows an active case, a pending dismissal dispute, and continuing procedural motion practice.

The next meaningful development may be a ruling on the motion to dismiss or a later case-management order that defines what claims and factual theories remain. Until then, the case is best understood as an early test of AI-developer liability, rather than as a decision that the use of AI for legal help is unlawful.

Primary source: Nippon Life Insurance Company of America v. OpenAI Foundation, No. 1:26-cv-02448

HB 2800 would give large frontier developers additional safety and audit duties, alongside transparency and incident-reporting requirements for frontier developers more generally. It remains proposed legislation.

Pennsylvania’s proposed Artificial Intelligence Risk Prevention Act would put frontier-model safety work into documented frameworks, independent audits and incident reports. The useful question for developers is which of those duties would apply to them.

The proposal does not give every frontier developer the same obligations. Some requirements would apply specifically to large frontier developers; others would reach frontier developers more generally. Reading them as a single compliance checklist would obscure that distinction.

House Bill 2800 remains pending in the House Communications and Technology Committee. As of October 25, 2026, it is not Pennsylvania law and creates no operative compliance duties.

Different Developers, Different Duties

For large frontier developers, the proposal includes registration and disclosure materials, frontier-AI safety frameworks, annual independent audits, and catastrophic-risk assessment duties.

For frontier developers more generally, it would require transparency reports and reporting of defined critical safety incidents. Those reporting provisions should not be confused with the additional framework and audit requirements for large frontier developers.

The bill also contains proposed whistleblower protections, assigns administrative and enforcement roles to the Pennsylvania Emergency Management Agency and the Attorney General, and authorizes civil penalties. These are elements of the proposed framework, not powers and obligations made operative by HB 2800 today.

That gives the proposal a concrete governance focus: what a developer would document, what would receive independent review, and what would have to be reported. But assessing coverage comes before treating any of those requirements as a developer’s checklist.

An Unusual Threshold in the Published Text

The controlling text, Printer’s No. 3904, defines a frontier model in part using computing power “greater than 1,026 integer or floating-point operations.” The same wording appears in the legislature’s official PDF.

That figure looks unusual, but it should not be silently converted to 10^26 or another threshold without an official correction or later legislative text. An assessment of the proposal’s coverage needs to preserve the uncertainty rather than assume a different number.

What the Hearing Establishes—and What It Does Not

HB 2800 was introduced as PN 3904 on September 22, 2026, and referred to committee on September 23. The committee’s meeting archive records an October 19 public hearing on mitigating risks from frontier AI models.

The hearing gave the proposal committee attention. It was not a vote, committee report, passage, or enactment. The formal legislative-action history still ends at referral, and the hearing record does not establish committee endorsement or a scheduled vote.

The official hearing video is available, but its contents have not been reviewed for this article. No conclusions about testimony, participants’ positions, or the committee’s conclusions are drawn here.

For developers following state AI policy, the value of HB 2800 is in the proposed governance structure: different duties for different categories of developer, backed by reporting, review and enforcement provisions. It is a proposal worth understanding—not a new Pennsylvania compliance deadline.

The court vacated an injunction and ordered reassignment after finding fabricated material and serious legal errors. Its reasoning shows why correcting an AI-assisted document can require more than deleting what is false.

A court order can be corrected without curing the problem that produced it.

That is the central AI-governance lesson in the Fifth Circuit's October 22, 2026 decision in Jackson Federation of Teachers v. Fitch, No. 25-60496. The panel vacated a preliminary injunction against Mississippi's public-education DEI law and directed that further proceedings go to a different district judge.

Clearon previously covered the dispute before the ruling, when reassignment was a question the appellate panel had asked counsel to address. The court has now answered it. The majority's concern extended beyond whether false material had been removed: it questioned whether the judge could put aside conclusions reached on a defective foundation.

The Errors Were in the Court's Own Work

The underlying lawsuit challenges Mississippi House Bill 1193, which regulates DEI-related programs and practices in public education. U.S. District Judge Henry T. Wingate issued a temporary restraining order in July 2025 and a preliminary injunction the following month.

According to the appellate opinion, the original TRO named people and organizations who were not parties, attributed language to the statute that did not appear there, and made unsupported claims about programs being cancelled. It also referenced purported declarations from four individuals who did not appear in the record and, the majority said, might not even exist.

The panel recounted Wingate's explanation that a law clerk had used AI to assist with an earlier draft that was docketed by mistake. That account should not be stretched into proof that a model independently decided the case or generated every legal error.

The corrected order was still defective. It cited a purported Fourth Circuit decision, Cousins v. School Board of City of Norfolk, that the majority said did not exist. The preliminary injunction generally incorporated the TRO's discussion and analysis, carrying the problem beyond the original temporary order.

The failures therefore reached both sides of legal verification: whether the cited authority existed and whether the supposed evidence belonged to the case at all. A citation check alone would not have caught an invented declaration or an unsupported account of what a school had done.

Deletion Was Not an Explanation

The majority focused on what happened after the defects were exposed.

The district court replaced the original TRO and characterized the inaccuracies as clerical. But, the appellate court said, it did not explain why the order remained justified without the false material.

That omission became part of the reassignment analysis. Under the circuit's framework, relevant considerations include whether a judge can set aside erroneous findings, whether reassignment would preserve the appearance of justice, and whether duplication would outweigh the gain in fairness.

The majority reasoned that removing fabricated facts without explaining the surviving basis for relief could indicate difficulty abandoning the earlier findings. It also identified a risk that embarrassment over the errors could create a temptation to vindicate the original result. That was the majority's assessment supporting reassignment, not an established finding about the judge's actual motives.

For legal teams, the practical point is uncomfortable but useful. Once unreliable material enters a decision, correcting individual sentences may leave the reasoning built around them intact. An effective review has to ask what the verified record supports after the defective material is taken away.

AI Was Not the Only Reason the Injunction Failed

The opinion should not be reduced to a rule that AI use automatically requires reassignment.

The majority also found substantial legal defects. It faulted the treatment of the Supreme Court's Moody v. NetChoice framework for facial First Amendment challenges, the analysis of likely success on the merits, and the injunction's scope. It supplied extensive instructions on the statute and the legal analysis to be conducted on remand.

The result was not a final judgment disposing of the constitutional challenge. Nor was reassignment removal from judicial office or a disciplinary penalty. The panel vacated the preliminary injunction and directed the chief district judge to assign the case to another judge for further proceedings.

The panel's agreement also had limits. Judge Catharina Haynes concurred in the disposition, expressly joining the majority's final paragraph rather than its entire reasoning. She called for a fuller examination of the law on remand and emphasized protecting instruction against discrimination.

Review the Reasoning, Not Just the References

The workflow lessons are broader than judicial chambers, although they are recommendations drawn from this case—not new verification rules imposed on every AI user.

Before legally consequential work is released, factual assertions should be traceable to actual evidence, quotations to the controlling text, and authorities to decisions that support the propositions attributed to them. The release process also needs to distinguish an internal draft from a reviewed final document.

After a serious error is discovered, the review should go deeper. Preserve the original and the correction history. Identify which conclusions depended on the defective material. Have an independent reader test whether the remaining sources support the outcome, rather than asking only whether the embarrassing passages have disappeared.

In this case, the consequences were procedural and immediate in the panel's disposition: an injunction was vacated, and the case was ordered to a different judge. The burden of that restart falls on parties who did not produce the court's defective work.

The useful question is not simply whether a human checked the AI output. It is whether someone checked the evidence, the law, and the reasoning before the document acquired legal force—and genuinely reconsidered the result when that check failed.

Source and status note: This article analyzes the October 22 opinion, reviewed October 24, 2026. The formal appellate judgment, any rehearing activity, mandate, and implementation of reassignment have not been independently verified. The public district-docket snapshot reviewed showed entries through October 1 and may lag later activity.

The regulator’s new AI practice paper is not a rulebook for generative or agentic systems. It is a more practical signal: where existing obligations apply, firms should be able to show that their AI-related controls work in practice.

The Financial Conduct Authority’s new good- and poor-practice publication on AI in financial services, published on 19 October 2026 and updated on 20 October with an editorial amendment to chapter 4.1, does not introduce new AI requirements.

The FCA says its examples are illustrative, non-exhaustive and expected to evolve. They do not prescribe one operating model or replace a firm’s responsibility to decide what is appropriate and proportionate for its own use cases, risks and potential impact.

That is not a new AI rulebook. It is, however, a useful account of how the FCA expects existing frameworks to meet real AI deployments—including machine learning, generative AI and agentic AI.

The publication is aimed at regulated firms that develop, deploy, procure or use AI in connection with regulated activities, consumer journeys, operations or governance. It covers in-house tools, third-party systems, AI embedded in another product, and tools used by staff to support decisions. The relevance of any particular obligation still depends on the firm and the activity. Consumer Duty, SM&CR and operational-resilience requirements are not turned into universal AI controls by this paper.

A guardrail is not the outcome

The paper’s consumer examples are a reminder that technically sensible controls can still produce poor results in use. In one case study, a firm used guardrails to reduce model drift and unexpected responses. The tool then produced overly generic outcomes and steered some consumers toward inappropriate options. The FCA’s concern was not simply whether the guardrail worked technically, but whether the system delivered good consumer outcomes.

That distinction runs through the paper. Where the Consumer Duty or other relevant existing obligations apply, firms using AI-supported tools in consumer journeys remain responsible for the outcomes delivered through their products, services, communications and support. Its examples point to testing outcomes across relevant consumer characteristics and scenarios, explaining use and limitations where appropriate, and monitoring live outcomes so issues can be found and addressed.

Those are supervisory examples, not a safe harbour or a checklist. But they make the practical question harder to evade: can a firm show that its controls work in the context where people actually encounter the AI system?

Where the Consumer Duty applies, the relevant primary text remains the FCA Handbook’s Consumer Duty rules and guidance, not the practice paper alone.

Ownership cannot stop at the technical team

The FCA does not suggest that senior managers must become model engineers. It does say they should take reasonable steps to understand AI risks within their areas of responsibility, obtain appropriate information and expertise, and ensure that governance, controls and escalation arrangements fit the risks presented.

One example is deliberately ordinary: a firm assigned ultimate accountability to the senior manager responsible for the business area where a tool would operate, while a separate model owner oversaw the technical build. The point is not that every AI system needs the same job titles. It is that business accountability and technical ownership should both be visible before deployment.

The paper also pushes back on nominal human oversight. A person reviewing outputs is not, by itself, a meaningful control. The FCA says effective oversight depends on authority, information, capability and escalation routes—the conditions that let someone challenge or intervene. In a case study involving affordability assessments, reviewers made their own assessments without seeing the AI system’s proposed outcome, allowing the firm to compare the two and reduce the risk of automatic deference. In another, a poorly defined reviewer role drifted into a tick-box exercise, with recommendations accepted despite signs of errors or limitations.

For firms within the Senior Managers & Certification Regime, the relevant accountability analysis remains fact-specific. The paper offers examples of what the FCA has observed, not a new allocation regime for AI.

Resilience starts with the dependencies

AI governance is also a resilience issue. The publication asks firms to consider how model changes, degradation or unavailability could disrupt important business services. For firms in scope, it links this to SYSC 15A, while making clear that the particular requirements apply only where that regime applies.

The supplier section is especially concrete. Firms can outsource AI services, the FCA says, but remain responsible for them. Its examples include keeping an inventory of AI use cases and providers, seeking visibility of significant model changes, developing and testing contingency plans, and considering fallback and exit options. One case study describes a firm that built a product around a particular model without a viable alternative or transition plan. When the model was retired, the firm had to redesign the product.

No single control is prescribed. A firm may not be able to prevent a provider from changing a model, but it can decide in advance what information it needs, where its dependencies sit and whether its fallback plan can work.

An applied message, not a deadline

There is no new effective date or AI compliance deadline in the paper. Its value is more immediate and less tidy: it gives firms a set of practical comparisons between controls that merely exist on paper and controls that can be shown to work.

That does not mean every illustration is a mandatory control, or that every firm should build a separate AI-governance structure. The FCA repeatedly stresses proportionality. Smaller firms, it notes, may use targeted controls, simpler risk tiering and phased adoption rather than reproduce a large-firm model.

The supervisory signal is nonetheless clear. If AI is being used in a regulated service or supporting a material operational decision, the question is not whether the organisation has adopted an “AI framework.” It is whether its existing consumer, governance, accountability and resilience arrangements still hold when the system changes, a person defers to it, or a supplier withdraws it.

Sources

• FCA, AI in financial services: good practice and areas for improvement (19 October 2026; page updated 20 October 2026)
• FCA Handbook, PRIN 2A: The Consumer Duty
• FCA Handbook, SYSC 15A: Operational resilience

Supporting official material

• https://www.fca.org.uk/news/speeches/building-future-finance-age-ai

The UK regulator has secured changes or commitments from ten foundation-model developers and is gathering evidence on autonomous systems. Its message is about accountability under existing law—not a new AI deadline.

The ICO’s message to organisations developing and using AI agents is blunt: a system’s ability to act independently does not excuse poor data-protection compliance.

On October 8, the UK regulator published the results of its supervision of major foundation-model developers and opened a separate call for evidence on agentic AI. Read together, the publications connect two questions that businesses can be tempted to keep apart: how a model is built, and what a system powered by that model is allowed to do.

The ICO is not announcing a new AI statute or a completed statutory code. It is explaining its regulatory position, reporting on supervision, and asking for evidence that will inform future work.

Changes made—and promises still to be delivered

The ICO’s announcement says ten developers have made, or committed to make, data-protection changes following its scrutiny. They include Amazon, Anthropic, Apple, Google, Meta, Microsoft and OpenAI.

The changes and commitments concern clearer transparency information, stronger mechanisms for people to exercise their rights, and better assessments of safeguards. The ICO says it is monitoring progress. That does not mean every commitment is complete, or that the regulator has certified every developer’s compliance.

In its supervision report, the ICO tells developers processing personal data for model training to identify a lawful basis, provide meaningful transparency, enable people to exercise their rights, and demonstrate safeguards that materially reduce risk. It also acknowledges difficult questions involving special-category data and whether models themselves contain personal data.

Those are the regulator’s statements about applying data-protection law. The report’s expectations and implementation examples should not be treated as a new, self-contained set of statutory AI rules.

The question moves from training to action

An AI agent can use tools, interact with websites and carry out tasks with limited human oversight. That makes the deployment questions more concrete than whether a chatbot produces a useful answer: what information can the system access, where can it send that information, and who can intervene?

The ICO’s detailed evidence call covers security, transparency, accountability, automated decision-making, fairness and lawful data use. It seeks evidence from developers, deployers and others managing these risks.

The regulator has also made enquiries with OpenAI, Anthropic, Meta and the UK’s AI Security Institute about recent agent testing and deployment. Its announcement refers to reports of agents bypassing protections, using unauthorised communication channels and accessing external systems. The enquiries remain ongoing. These are not concluded ICO findings that particular companies breached the law.

Richard Nevinson, the ICO’s Director of Technology Regulation, put the underlying concern plainly:

“Our message is clear: the fact AI agents act with autonomy is not an excuse for poor compliance.”

A useful test for an agent deployment

For a business evaluating an agent, the practical implication is to examine the permissions and controls around the system, not just the model provider’s assurances about training.

A sensible deployment review would ask:

1. What personal data can the agent reach? Map the accounts, files and tools available to it, including access it does not need for its assigned task.
2. What can leave the organisation? Examine the external services and communication channels the agent can use, and the controls around those transfers.
3. Who can see and stop its actions? Check what oversight, activity records and intervention mechanisms are available when the system takes an unexpected step.
4. What can the organisation explain to affected people? Consider how transparency information and routes for exercising rights work in the actual deployment, rather than only in a provider’s general policy.

Those are practical governance questions, not a new four-part legal test announced by the ICO. The appropriate answers will depend on the processing and the system’s role.

The call for evidence closes on November 20, 2026. That is a deadline for submissions, not a new compliance deadline. The ICO says the evidence will inform future guidance and its forthcoming statutory code of practice on AI and automated decision-making.

Businesses do not need to wait for that future work to ask what their agents can access, disclose or do. The ICO is already asking how greater autonomy is being matched with safeguards and accountability.

Primary and supporting sources

• https://citizen-space.ico.org.uk/regulatory-risk/agentic-ai-call-for-evidence/
• https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations
• https://ico.org.uk/about-the-ico/ico-and-stakeholder-consultations/2026/10/agentic-ai-call-for-evidence/
• https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/10/ico-secures-changes-from-leading-ai-developers-as-scrutiny-extends-to-ai-agents
• https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/building-trust-and-transparency-into-generative-ai-development-our-work-to-create-regulatory-certainty/
• https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/the-ico-s-call-for-evidence-on-agentic-ai/