Hirzel Dreyfuss & Dempsey, PLLC

NEWS AND INFORMATION

Patrick Dempsey Patrick Dempsey

AI Learned to Read Software. Your Business System Is Next.

AI can rebuild software from the outside, and now a franchise system too. What trade secret law, the FDD and Florida covenants still protect.

The short answer

In October we wrote a two-part series on what happens to software owners when AI agents can turn a shipped program back into readable code. Part 2 concluded that the contract and the server matter more than the binary.

Six weeks later, three things are worth reporting.

  1. There is still no lawsuit. As of this writing we have found no court case over AI-assisted decompilation or an AI-built software clone. Enforcement is happening through takedowns, letters and the AI companies' own guardrails.

  2. The one appellate decision that touches the question cuts against the copier. A federal appeals court said in September that making copying easy does not make it fair.

  3. The same method works on a business, not just a binary. An AI that can study a program from the outside and write a specification for rebuilding it can do the same with a franchise system. For franchise systems, federal law requires the franchisor to publish part of the outline.

This post covers the update briefly and then turns to the third point, which matters most to our franchise and business clients.

What changed in software since October

The AI companies became the first line of enforcement. On November 10, 404 Media reported that users of Anthropic's Claude and OpenAI's Codex were hitting safety blocks, and in some cases account limits, when they asked the tools to reverse engineer software, including on devices they own. One user post it quoted read, "Claude is refusing to decompile anything." Some users are moving to open models they can run themselves. The practical point is that the "contract" layer we described in Part 2 now includes the AI vendor's terms, not only the target software's license.

Takedowns continued. Trade press reported that Nintendo had several automated "recompilation" ports of its games removed from code-hosting sites in mid-November. We have not seen the notices, and no report confirms that AI was used to build those ports.

A federal appeals court drew a line between necessity and convenience. In Thomson Reuters v. ROSS Intelligence, the Third Circuit rejected a fair use defense for an AI legal research startup that copied Westlaw headnotes. The case is not about software. But the court addressed the software cases that reverse engineers rely on, including Sega v. Accolade and Sony v. Connectix. It read them as cases where copying was the only way to reach the unprotected functional parts of a program. Then it wrote: "Unlike necessity, ease is not a justification for copying."

That sentence is a problem for anyone whose argument is that AI made decompilation cheap. ROSS asked the full court to rehear the case on November 3, and several groups filed briefs in support. The decision does not bind courts in Florida. It is still the first appellate opinion to frame the issue this way.

"Clean room by AI" became a product, and no court has tested it. A clean room is a way of rebuilding something without copying it. One team studies the original and writes a description of what it does. A second team, which never sees the original, builds from that description. In March, a service called Malus began selling an AI version of this for open-source software. The same month, the maintainer of a widely used Python library released a new version under a more permissive license and called it a clean-room rewrite by AI. Critics answered that the AI model had been trained on the original code. That dispute never reached a court, and neither has any other.

The tools are good, and less finished than the demos suggest. A September study of AI decompilers found that output passing every available test still behaved differently from the original in about 5 percent of cases, and in up to 13 percent for one system. Code that looks done is not always code that is right.

The same method works on a business

Here is the part of this story that has had the least attention.

A clean-room rebuild of software starts with what anyone can observe: the screens, the documentation, the behavior. A business system has the same kind of public surface. An AI agent can read a competitor's website, menu, pricing, job postings, customer reviews, training videos and regulatory filings, and turn them into a working description of how the business runs.

For a franchise system, that public surface is larger than most owners realize. The FTC's Franchise Rule requires a franchisor to give prospects a Franchise Disclosure Document, and FDDs are collected on searchable websites. Item 11 requires the franchisor to disclose the table of contents of its operating manual and "the number of pages devoted to each subject."

That is an outline of the system, published by the system. One bakery franchisor's 2025 disclosure, for example, lists a 50-page operations manual by subject, with page counts for staffing, operations and maintenance, administration, products and recipes, and marketing. An AI cannot read the manual from that. It can see what the manual covers and in what proportion, and fill in the rest from everything else that is public.

What the law protects, and what it never did

None of this is new law. It is old law that just became cheap to use.

Learning from the outside is lawful. The federal Defend Trade Secrets Act says that "improper means" of acquiring a trade secret does not include "reverse engineering, independent derivation, or any other lawful means of acquisition." Florida's trade secret statute reaches the same place by a different route: information that is "readily ascertainable by proper means" is not a trade secret at all.

Taking from the inside is not. The operations manual itself, training materials, recipes and specifications, supplier terms and customer lists can all be trade secrets when the owner actually keeps them secret. A former franchisee, manager or employee who feeds the manual into an AI tool to build a competing business has not reverse engineered anything. That is ordinary misappropriation, usually combined with breach of the franchise agreement. As we explained in an earlier post, people who never signed the agreement can be enjoined for helping.

"It was all public" has limits. In Compulife Software v. Newman, the Eleventh Circuit, whose decisions bind federal courts in Florida, upheld a trade secret judgment against competitors who used a bot to pull millions of insurance quotes from a public website. Each quote was public. The compilation was not something a person could realistically collect by hand. The court was careful to say that automated collection can be perfectly legitimate, and it tied its ruling to the facts of that case. An AI agent that harvests a competitor's customer-facing system at scale is closer to those facts than a person taking notes in a store.

Copyright helps less than owners expect. A manual is copyrightable, but the protection is thin where the content is basic instruction. In a 2016 case over training manuals, a Colorado federal court found that a competitor had copied from the plaintiff's manuals and still rejected the copyright claim, because what was copied was too basic to protect.

The look of the place is a separate question. A competitor who uses AI to match a system's appearance runs into trademark and trade dress law, which protects distinctive look and feel. A clean room does not help with that.

What four 2026 franchise cases have in common

Four federal decisions this year involved franchisors trying to protect their systems after a franchisee left. We rely here on published summaries by franchise lawyers and, for the Florida case, on the court's docket.

  • New York. A court let trade secret claims go forward against a former franchisee who allegedly kept the operations manual and a printed client database after termination.

  • Arizona. A trade secret claim survived because the franchisor identified the secrets in its manual. A separate claim for conversion failed.

  • Virginia. A court dismissed trade secret claims that rested on a "confidential brand operations manual" and marketing strategies described only in general terms. It let claims about a specific pricing formula and a lead-tracking system proceed.

  • Florida. In King Lombardi Acquisitions v. Petricoff, a Southern District of Florida judge denied a franchisor's motion for a preliminary injunction on June 3, after a two-day hearing. According to a published summary of the order, the court found that the methods and forms the franchisor called proprietary were generic industry material, which is not a legitimate business interest under Florida's restrictive covenant statute. The docket shows the franchisor voluntarily dismissed the case later that month.

The common thread is specificity. The franchisors who could say exactly what was secret, and show that they treated it that way, kept their claims. The ones who pointed at "the system" did not. We wrote about how specific a trade secret complaint has to be earlier this year, and the question is not settled in this circuit.

AI raises the stakes on that question. When the observable half of a system can be rebuilt cheaply, the case turns on the half that cannot be observed, and on whether the owner can identify it.

Contracts: what is being added, and what will hold

Franchise lawyers have begun recommending new contract terms: bans on uploading system materials to outside AI tools, bans on using confidential information to train AI models, and return obligations that reach AI workspaces and chat histories. Those are sensible. A franchisee who has agreed not to put the manual into an AI tool, and does it anyway, has made the franchisor's case simpler.

There is a limit, and Florida law states it plainly. A restrictive covenant must protect a legitimate business interest. A clause that tries to stop a former franchisee from using AI to analyze public information protects nothing the franchisor owns, and King Lombardi shows how a Florida court treats a claim built on generic material. Florida's statute also presumes that a covenant against a former franchisee that is not based on trade secrets is unreasonable if it runs more than three years. The terms that decide these disputes are still the ordinary ones, drafted with care.

The leak that runs the other way

In July, Microsoft's chief executive, Satya Nadella, described what he called a reverse information paradox. "You essentially pay for intelligence twice," he wrote: once in money, and once in the proprietary knowledge a business feeds into an AI system to make it useful.

For a franchise system, that is a secrecy problem with legal consequences. Trade secret protection depends on reasonable efforts to keep the information secret. If franchisees and managers are pasting recipes, pricing models and manual sections into consumer AI tools, and no one told them not to, the franchisor will hear about it when it tries to enforce. We made a similar point about litigation documents and protective orders.

A short checklist

For franchisors and other system owners

  1. Write down what is actually secret, item by item. If you cannot list it, you cannot plead it.

  2. Read your own FDD the way a competitor's AI would. Know what Items 11 and 14 already disclose.

  3. Keep the manual in a system you control, with individual logins and access logs.

  4. Add AI terms to the franchise agreement and the manual, and extend them to franchisees' employees.

  5. Enforce the terms you have. Courts look at practice, not only paper.

  6. Invest in what cannot be copied by reading: the brand, supplier relationships, training and support.

For franchisees, former franchisees and competitors

  1. Build from what is public: the FDD, what a customer can see, and general industry knowledge.

  2. Keep anyone who had the manual out of the build, and keep the manual out of the AI tool. A file uploaded once stays in the record.

  3. Assume your prompts and agent logs will be read. They are discoverable, and they will either prove your clean room or disprove it.

  4. Do not automate bulk collection from a competitor's customer-facing systems without advice.

  5. Stay away from the name, the marks and the look.

  6. Check your post-term obligations before you start, not after the demand letter.

What to watch

  • The Third Circuit's ruling on the request to rehear ROSS.

  • The first lawsuit over an AI-built clone. We still expect one.

  • Whether any court addresses an AI clean room.

  • Whether franchisors begin adding AI-specific confidentiality terms to their agreements and disclosure documents. We did not find one in the disclosure documents we searched.

A note on our earlier posts

Part 1 named three of the ArtCraft project's AI-built alternatives to Adobe applications. The project launched with seven and has since added office applications.

Sources

Disclaimer

This post discusses publicly reported legal developments for general informational purposes. It is not legal advice, it does not create an attorney client relationship, and it does not reflect the firm's position in any pending matter. Outcomes depend on the specific facts and the governing law of the relevant jurisdiction.

Read More