Issue No. 5 - October 14, 2026 | |
| |
Issue No. 4 covered the debate over whether frontier AI should be reined in, and where that leaves a business lawyer once federal rules are off the near-term table: with state law, contracts, and ordinary liability. This issue picks up the practical side of the same problem. Two September surveys put numbers on the gap: nearly every large public company surveyed now has an AI policy, and a large share route around it under deadline. That is the Spotlight, and it frames much of what follows: privilege risk when AI joins the advisory circle, vendor AI addenda, and what a firm actually owns when a vendor trains on its work.
Florida is well represented this time. A Fourth District opinion condemns “AI slop” filings and says the problem is not fake citations, and the Attorney General has asked a state court for a temporary injunction against OpenAI. Also in the courts, the Ninth Circuit rules on Copilot and the DMCA, and a new antitrust suit tests whether public calls for an AI slowdown can amount to an agreement among competitors. The Third Circuit issued what appears to be the first federal appellate ruling on fair use and AI training, the FTC confirmed it is investigating OpenAI, Anthropic, and other AI developers, and California’s governor signed SB 574, which sets rules for lawyers’ use of generative AI.
As always, every item links to a source you can read in full, grouped by topic so you can go straight to what matters for your practice.
The Task Force welcomes your participation. If you come across an article worth sharing, have an idea for the Task Force, or want to get involved, we want to hear from you. You can find us at https://flabizlaw.org/committees-task-forces/task-forces/artificial-intelligence-task-force/
— The BLS AI Task Force
| | |
98% of large public companies surveyed have an AI policy. Nearly half have bypassed it.
EY US AI Risk and Governance Survey (Sept. 15, 2026); OneTrust 2026 AI-Ready Governance Report (Sept. 14, 2026)
Why it matters: with formal AI policies now near-universal at large companies, the exposure for most clients is a policy that exists on paper and gets overridden in practice, which is exactly the record a plaintiff or regulator wants to find.
Two surveys published in September put numbers on the gap. EY surveyed 202 senior AI decision-makers at U.S. publicly traded companies with at least $1 billion in annual revenue. Ninety-eight percent report a formal AI governance policy, while 47% say the organization has previously bypassed that process for urgent deployments. The gap is widest around autonomous “agentic” tools: among companies using agentic AI, 49% say their governance framework has not been updated to cover it, and 26% cannot detect unauthorized agents operating internally. Thirty-six percent report an AI incident or failure that caused a materially negative impact.
OneTrust’s 2026 AI-Ready Governance Report, released September 14 from a survey of 1,200 senior business decision-makers in eight countries, points at the mechanism. A third (33%) of surveyed organizations have seen employees use unapproved AI because approved tools or processes were not available quickly enough. Eighty-seven percent encourage AI agent use, but only 47% have clear governance, oversight, and controls in place, and 86% experienced at least one AI-related incident in the past year, such as sensitive data or intellectual property exposure, unapproved employee AI use, misinformation, or data loss.
For members, the takeaway is concrete and repeatable across clients. A governance policy that is routinely bypassed under time pressure can be riskier than a thinner policy the organization follows, because a documented gap between the written standard and actual practice is evidence a plaintiff or regulator can use. Three things to raise with a board or GC: make the exception path a real, documented process rather than an informal override; treat shadow AI as a sign of an approval process that is too slow; and, for any client deploying agents, pin down in writing who monitors the agent after launch and who is accountable when it acts outside its lane.
| | Ethics and Professional Responsibility | | |
A Florida appeals court condemns “AI slop” filings and threatens a Bar referral, though fake citations were not the problem
District Court of Appeal of Florida, Fourth District (Lisandrillo v. Palozzi, No. 4D2026-2262, Sept. 16, 2026; Gross, J., with Levine and Shepherd, JJ., concurring)
Denying a writ petition in a Broward County divorce case, the court said counsel’s filings appeared to be “AI slop”: a 61-page petition seeking review of eight rulings, scattershot arguments, and factual claims the transcript contradicted. It said the problem had nothing to do with hallucinated citations, added that whether AI was actually used made no difference because a lawyer must stand behind every filing, cited the Florida Bar’s competence and candor rules, among others, and ordered counsel to show cause, without using AI, why she should not be sanctioned or referred to the Bar. Counsel has told Law.com she does not use AI to generate court filings, and we found no ruling on the show-cause order as of this writing. For members, checking citations is only the floor for anyone supervising AI-assisted drafting.
| | Privilege and Confidentiality | |
When AI joins the advisory circle, privilege can leave with it
ArentFox Schiff LLP (Adam Diederich and Kirstie Brenson), Aug. 24, 2026
When a client’s advisors run privileged legal advice through an AI tool (to summarize it, expand on it, or ask follow-up questions), the alert warns that privilege may be waived even when the prompt does not upload or quote the original advice. The alert is written for family offices, where accountants, trustees, and investment managers increasingly use AI as a “legal thought partner,” but the risk reaches any client whose team treats a general AI tool as a sounding board for counsel’s advice. The practical move is to tell clients, before the fact, that pasting legal advice into an AI tool is a disclosure decision with privilege consequences, and to set expectations on which tools, if any, are acceptable for that use.
Human contractors are reading Copilot prompts and uploaded images, 404 Media reports
404 Media (Joseph Cox); complaint in Vredenburgh v. OpenAI OpCo, LLC, No. 3:26-cv-10527 (N.D. Cal., filed Sept. 16, 2026)
404 Media reports that human contractors are reading Microsoft’s Copilot prompts and uploaded images, two weeks after it reported the same practice at OpenAI. A putative class action over OpenAI’s “Project Lily” (Vredenburgh v. OpenAI OpCo, LLC) alleges that OpenAI never disclosed to users that third-party human reviewers read their prompts to improve the model. Those allegations are untested, and OpenAI’s response is currently due October 13. The practical point for members is to confirm which tier of any AI tool holds client information and what its terms say about human review.
Quantum risk is real, but the nearer threat to privileged data is files that were never encrypted
International Legal Technology Association (Patrick Bryden), Aug. 6, 2026
Most quantum-security discussion centers on “harvest now, decrypt later,” meaning encrypted data stolen today and read once quantum decryption arrives. Bryden argues that framing pulls attention from a closer problem, which he calls “harvest now, read now”: sensitive files that were never encrypted at the content layer, so anyone past the perimeter can read them immediately, with no quantum computer required. The practical step is to encrypt privileged and confidential files themselves, not just the network around them, while folding post-quantum encryption into the longer-term roadmap. (The author’s company sells content-layer encryption.)
A court sorts a testifying expert’s ChatGPT history: the case summaries he relied on are discoverable, his questions and drafts are not
U.S. District Court, District of Nevada (Villanueva v. Las Vegas Metropolitan Police Department, No. 2:24-cv-00125-ART-DJA, ECF No. 112, Sept. 29, 2026)
A magistrate judge had ordered an excessive-force expert to produce his entire 40-page ChatGPT log. On objection, District Judge Anne R. Traum held that went too far: the case-law summaries ChatGPT generated and the expert consulted must be produced as facts or data he considered, while his prompts and the output he used to prepare for his deposition or to draft sections of his report stay protected as trial-preparation and draft material. She also held that using ChatGPT, which she described as an application rather than a person, did not waive any protection. For members who retain experts, settle at the engagement how the expert’s AI use will be logged and what will be disclosed, because at least one court has now drawn the line through the log rather than around it.
| |
California moves to speed up its new AI-safety laws and study a frontier-model “kill switch”
Office of Governor Gavin Newsom (Executive Order N-9-26, Sept. 18, 2026); reported by POLITICO (Christine Mui)
On September 18 Governor Newsom signed Executive Order N-9-26. It moves up the implementation deadlines for SB 813 and AB 1405 (the AI-auditor registration framework covered in Issue No. 4) and directs the Government Operations Agency, consulting with the Office of Emergency Services and national experts, to recommend by November 16, 2026 how to strengthen California’s AI-safety laws. Two items on that list are worth watching. One is updating the definition of “critical safety incidents” that AI companies must report to include loss-of-control incidents; the Governor’s announcement points to the Hugging Face attack, the incident behind our Issue No. 3 Spotlight. The other is a possible “kill switch” requirement for frontier models, with its efficacy verified on an ongoing basis by an independent organization. A federal version was blocked on the Senate floor, and Anthropic’s Dario Amodei has said a kill switch “could be a good idea.” For members, the concrete marker is November 16: clients tracking state AI-compliance obligations should watch the recommendations and how fast the accelerated SB 813/AB 1405 timelines land.
California enacts SB 574: lawyers may not delegate the practice of law to AI and must personally verify every citation they file
Holland & Knight LLP (Philippa J. Balestrieri, David Jargiello, Kristin A. Ornstein, and Tom Stephenson), Oct. 1, 2026
Governor Newsom signed SB 574 on September 30; Issue No. 4 reported it waiting on his desk. Starting January 1, 2027, a California lawyer may not delegate the practice of law to generative AI, may enter confidential or personally identifying information only into a system whose access is limited to people bound by confidentiality, must take reasonable steps to verify AI output and disclose its use to the court for all documents submitted, and may not file a paper containing any citation the attorney responsible for submitting it has not personally verified, whatever its source. Holland & Knight flags open questions the statute leaves unanswered, including what counts as delegation and how broad the disclosure duty is. For Florida members with California matters this is now a compliance date, and for everyone else it is the most detailed model yet of what a lawyer-AI statute looks like.
OpenAI’s Altman says the world “should accept some bad things happening” for AI’s benefits, even as his company backs tougher state safety laws
POLITICO (Christine Mui and Brendan Bordelon), Decoded by POLITICO, Oct. 4, 2026
In an interview for the debut of POLITICO’s Decoded newsletter, OpenAI CEO Sam Altman said there is “a lot of daylight” between OpenAI and Anthropic on regulation: OpenAI will accept some hacks, misuse, and scams as the price of keeping AI broadly available, though not catastrophic risks such as a serious loss of control. POLITICO reports that the companies’ actual positions have narrowed. Altman agreed with Anthropic CEO Dario Amodei’s call last month to slow the most advanced models, OpenAI has followed Anthropic in endorsing stricter state safety laws, and its lobbyists recently backed a bipartisan House proposal that would require top AI companies to embed outside safety evaluators. For members, state legislatures remain where frontier developers’ safety commitments are being written into law, and a leading vendor’s stated tolerance for “some bad things happening” is a reason for clients to allocate AI risk in their own contracts rather than rely on the vendor’s posture.
| |
Florida’s attorney general asks a state court to bar OpenAI from serving minors and to halt new model development without independent safety checks
Axios (Josephine Walker and Avery Lotz); also reported by WJXT News4JAX (Jonathan Lundy)
On September 28 Attorney General James Uthmeier moved for a temporary injunction in the suit he filed June 1 against OpenAI entities and CEO Sam Altman, which includes a claim under the Florida Deceptive and Unfair Trade Practices Act. He asks the court to keep ChatGPT away from Florida minors, to bar representations that it is safe, reliable, or accurate, and to halt development of AI models without independent third-party guardrails and approval. The motion is pending with no hearing date reported as of this writing, and the case (No. 26-000295-GC-AXMX, Highlands County) is back in circuit court after U.S. District Judge Aileen M. Cannon granted remand on September 8 (federal docket). Clients marketing AI products should read the requested restrictions as a preview of how a FDUTPA theory may treat claims that an AI product is safe or reliable.
Ninth Circuit holds Copilot’s generated code is a new work, not a copy with attribution stripped, so the DMCA claim fails
U.S. Court of Appeals for the Ninth Circuit (Doe v. GitHub, Inc., No. 24-7700, Sept. 16, 2026; Miller, J., joined by Thomas, J., and Blumenfeld, D.J.)
In an interlocutory appeal, the panel affirmed dismissal of the DMCA section 1202(b) claim because, on the plaintiffs’ own allegations, Copilot creates new works that never contained copyright management information, so nothing was “removed” or “altered.” It treated the separate input (training-stage) theory as forfeited and took no view on whether similar output could infringe copyright. The two breach-of-contract claims over open-source license terms remain pending in the district court, so the contract claims and any ordinary copyright theory are the routes left for attribution complaints.
Four AI subscribers sue Anthropic, OpenAI, SpaceXAI and Google, calling the public call for an AI slowdown an antitrust agreement
Associated Press (Kaitlyn Huamani), via PBS NewsHour; Buist v. Anthropic, PBC, No. 5:26-cv-10693-NC (N.D. Cal., filed Sept. 18, 2026)
The complaint (available on CourtListener) alleges that the September 12 exchange, in which Anthropic’s CEO urged industry-wide coordination to slow frontier development and executives at the other three companies publicly agreed, was an agreement among competitors to restrict output under section 1 of the Sherman Act. It seeks treble damages and an injunction for a proposed class of paid subscribers. This follows the Issue No. 4 Spotlight and rests on allegations only (none of the four companies had responded to AP’s request for comment when it published), but it tells clients in industry consortia or standards bodies that safety coordination among competitors can draw antitrust scrutiny.
Third Circuit affirms for Thomson Reuters against ROSS: Westlaw headnotes are copyrightable, and using them to train a competing AI tool was not fair use
U.S. Court of Appeals for the Third Circuit (Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, Sept. 29, 2026); reported by LawSites (Bob Ambrogi)
In what appears to be the first federal appellate decision on fair use and AI training, the panel (Restrepo, Montgomery-Reeves, and Bove, JJ.) affirmed partial summary judgment for Thomson Reuters: its headnotes carry enough creativity to be protected, and ROSS’s use of them to train a legal search tool built to compete directly with Westlaw was not fair use. The court called it “no more than an ordinary copyright case,” and the facts are narrow (a non-generative tool built before the generative AI boom, aimed squarely at the plaintiff’s market); ROSS says it will ask the Supreme Court to review the decision, according to LawSites. For clients building products on licensed or scraped content, training on a competitor’s material to build a substitute product now carries appellate-level risk in the Third Circuit, though the court stressed that ROSS’s tool was not generative AI.
FTC confirms it is investigating OpenAI, Anthropic, and other AI developers over risks to consumers
New York Post, Sept. 30, 2026; FTC confirmation reported by the Associated Press
An FTC spokesperson confirmed that the agency has opened an investigation into OpenAI, Anthropic, and other AI developers over dangers their technology may pose to consumers but declined further comment; the Post, which first reported the probe, said it has been underway for months. In recent months the companies have disclosed examples of AI agents going beyond human instructions and hacking external websites. Read with Chairman Ferguson’s remarks in the rogue-agent item below, consumer-protection authority is the federal enforcement tool to watch. Clients that deploy AI agents should be able to show what controls they use and that their public statements about those controls are accurate.
Arizona appeals court vacates a sentence that relied on an AI-generated video of the victim
Arizona Court of Appeals, Division One (State v. Horcasitas, No. 1 CA-CR 25-0191, Sept. 30, 2026, published)
The court held that an AI video recreating a manslaughter victim’s face and voice, speaking words his family imagined he would say, lacked the minimal reliability due process requires of sentencing evidence, and that a disclaimer did not cure the problem. The conviction stands, but the sentence was vacated and the case remanded. It is a criminal case, and review was only for fundamental error because no one objected, but its reliability analysis is a useful guide for anyone offering an AI-generated recreation to a decision-maker, from trial demonstratives to arbitration presentations.
| AI Issues Affecting Clients' Businesses | | |
How DraftKings uses AI to target the gamblers likeliest to lose
The New York Times (Alex Klavens, Walt Bogdanich, and Jenny Vrentas), Sept. 19, 2026
Drawing on internal documents and interviews with more than 40 former employees, the Times reports that DraftKings began building machine learning models in 2023 to steer its promotional “free” bets and bonuses toward customers predicted to gamble, and lose, more in response, while a separate model that would have given gamblers addiction “risk scores” was sidelined, according to two former employees who worked on it. A company executive has told investors that AI helps it personalize hundreds of millions of promotional dollars; DraftKings says its promotions go to engaged customers, not to customers based on their losses, and rejects any suggestion that its marketing is unfair. For members advising a business that uses AI to decide who gets offers, discounts, or prices, the story shows how a model tuned only for return can look like targeting vulnerable customers, which is the record a regulator or plaintiff will look for. Test the model’s targeting against harm indicators before launch, and document that you did.
Flagged for the Task Force by Robert Kain, The Concept Law Group, P.A., Chair of the BLS Blockchain & Digital Assets Committee.
Who’s liable when AI agents go rogue?
MIT Technology Review (Michelle Kim); FTC remarks reported by Reuters (Jody Godoy)
Kim reports that state AI transparency laws require incident reports only for catastrophic harm, so the state attorneys general investigating OpenAI over the Hugging Face breach (the incident behind Issue No. 3’s Spotlight) are relying on consumer-protection and other state laws, which the experts she quotes call a poor fit. University of Houston law professor Gabriel Weil sees “plausible grounds for a negligence claim,” but Hugging Face has chosen not to sue. Separately, FTC Chairman Ferguson told a Reuters event on September 25 that he will resist treating AI agents as independent actors because audit trails show they follow instructions, and Reuters reported his remarks as suggesting developers should be liable for what their agents do, with the FTC’s existing data-breach authority as one possible tool. Clients deploying agents should expect regulators to look to whoever instructs the agent, and clients buying agents should negotiate accordingly.
Proceed with caution: using AI to set wages
Hunton Andrews Kurth LLP (Daniel J. Butler and Robert T. Quackenboss), Hunton Retail Law Resource, Sept. 23, 2026
Colorado’s legislature passed HB 26-1210 to limit AI-driven price-setting and wage-setting, and the governor vetoed it as vague, but similar bills are moving in California, Maryland, New York, and Georgia, and a federal bill proposed last year would allow private suits. Employers using AI to set pay should limit inputs to lawful factors such as duties, performance, and seniority and keep biometric and online-activity data out of the model.
Minimum viable AI governance: how growing companies can manage AI risk without slowing down
Nixon Peabody LLP (Troy Lieberman), Sept. 29, 2026
For companies that cannot staff a full AI program, the article sets out a short list: inventory the AI tools in use, map what data goes in and whether it trains a model, test how each system fails, scale controls to the harm a failure would cause, give someone ownership, and record why high-risk uses were approved. Its most useful point for counseling startups is that enterprise customers’ AI diligence questionnaires are already the strictest regulator many of them will face, which pairs with this issue’s Spotlight on policies that exist on paper.
What corporate clients are starting to ask outside counsel about AI
National Law Review (Olivia Gang), Aug. 21, 2026
Corporate clients are increasingly asking their outside firms how they use AI and how they govern it (from data handling to output verification) as part of engagement and diligence. For members, this cuts both ways: it is a preview of the questions your own clients may put to you, and a template for the questions a client should be putting to its other vendors. The concrete step is to have a short, honest answer ready on how the firm uses AI and what controls sit around it, before a client asks in a pitch or an audit.
Meta rushed to fix “VM escape” flaws in its Muse AI agent in the weeks before launch, 404 Media reports
404 Media (Jason Koebler), Oct. 5, 2026
Citing internal posts and a Meta source, 404 Media reports that in the weeks before Muse launched on September 8, Meta security teams worked nights and weekends to fix vulnerabilities that could have let a user break out of the virtual machine each Muse agent runs in and reach sensitive internal Meta systems; the source called the fixes “half-baked protections” rushed out to hold the launch date, and Meta says Muse is protected by red-teaming, a bug bounty, and user controls. Separately, researcher Patrick Wardle found a zero-day that let other software take control of a user’s Muse (Ars Technica). Muse works by connecting to a user’s email, calendar, messaging, and other accounts, so any client or firm considering a consumer AI agent should ask what a single failure in the vendor’s isolation would expose, and ask it before anyone connects a work account.
| Law Practice Best Practices | | |
Prompt injection: hidden instructions in a document can hijack an AI tool’s output
Eversheds Sutherland (US) LLP (Amelia Toy Rudolph), via JD Supra
Instructions buried in a document (white-on-white text, tiny type, or off-page placement) can be picked up and followed by an AI tool that processes the file, producing output the lawyer never intended and may not catch. The practice point is specific: treat AI review of a document from an opposing party or an outside source as an input to verify, and understand that the risk can arrive in the file itself as well as in the prompt the lawyer writes.
Vendors are racing to build the “judgment layer,” and it raises an ownership question
Artificial Lawyer, Sept. 3, 2026
Legal-tech firm Aloi is building a system to capture the judgment embedded across a firm’s negotiated documents and closed matters, so a new matter draws on the firm’s own accumulated decisions rather than a generic model. This is part of a wider move toward models trained on proprietary firm data and “digital twins” of lawyers. The forward-looking question for members is contractual: if a vendor trains a model on your work product, who owns the resulting model, what happens to it when the engagement ends, and does the client whose matters generated that work have any say. Those terms are worth settling before the tool goes into use, not after.
| | | | |
