On September 30, 2026, a developer named Brandon Thomas committed 42,241 lines of code in a single push. Within a week, his project, ArtCraft, had released free counterparts to Photoshop, Illustrator, Premiere and four other Adobe applications, followed by Word, Excel and PowerPoint, with Anthropic’s Claude credited as co-author on most commits. “Software is over,” he announced.
For decades, companies such as Adobe and Microsoft have built multibillion-dollar businesses around software that once required enormous engineering teams and years of development. Developers are now asking AI agents to build competing applications using little more than natural-language instructions. Thomas’s projects aim to reproduce familiar word-processing, spreadsheet and presentation capabilities using open-source code. These projects remain at an early stage and should not be mistaken for complete replacements for their commercial counterparts. The implications are nonetheless significant. If an individual can instruct an AI agent to build an application resembling Photoshop, Word or Excel, what exactly prevents that individual from distributing or commercializing it? More importantly, can Adobe or Microsoft stop them?
The answer depends on what was taken and how. Copyright does not grant ownership of every function a program performs, but new code can infringe patents, copied assets can infringe copyright, and the process used to study a product can breach a licence. AI is accelerating development without collapsing these distinctions.
I.The Mechanics: What “Reverse Engineering” Means in Practice
Reverse engineering is the process of starting with a finished product and working backward to determine how it was designed or how it operates. In software, this ordinarily means analysing a program’s behaviour or compiled code to understand its functionality, usually without access to the vendor’s source code. Reverse engineering can encompass several distinct activities, and the distinctions among them matter because the development process may itself involve legally significant copying:
- Black-box observation. An agent can rebuild functionality without opening the original program’s code. The agent consults public documentation, supplies inputs and records outputs, then writes code, runs tests and revises its mistakes, refining the code until the results match. This is often better described as AI-assisted reimplementation than as extraction of the original software.
- Decompilation. Tools such as REA (“Reverse Engineer Anything”) connect agents to decompilers that convert machine code into approximate pseudocode. A decompiler attempts to reconstruct readable instructions from the distributed program, but it ordinarily cannot recover all of the original names, comments and structure.
- Source-based rewriting. Rewriting from exposed source code differs again. The agent restates the original code, often in another programming language. This is paraphrase with the original open on the desk, and it is where the chardet and Claw-Code disputes reside. Changing the programming language or asking a model to paraphrase does not establish independent creation.
Thomas claims, although the claim has not been independently verified, that ArtCraft was built using the first method. The project’s contributor guidelines target one-to-one parity with Photoshop and prohibit copying Adobe’s code, but the project also used scripts to drive the actual Photoshop, generating test files against which its output is verified pixel by pixel. Photoshop thus served as the answer key. The word “everything” in the title of this piece is, for now, an overstatement: PhotoCraft’s own tally placed it at roughly 45% of professional requirements in mid-October, and logic running on a vendor’s servers cannot be read from a desktop copy at all. The trajectory, however, matters more than the present state of the art. A recent paper by Ilan Mandel, James Grimmelmann and three co-authors contends that existing law tolerates reverse engineering because its difficulty separates legitimate study from misappropriation. That filter is eroding.
i.Clean Room Development
A “clean room” is a development process designed to show that new software was created independently of an existing program. One team studies the original and records its functional requirements in a specification, and a separate team, with no exposure to the original’s protected code, writes the new software from that specification alone. Evidence that software was developed in a clean room can strongly support a finding that it was independently created and not copied from the original.
AI blurs the separation on which a clean room depends. A single agent, or agents sharing the same context, may both study the original and write the replica, and because a model may have been trained on the original code or related material, it is difficult to show that the implementing side truly had no exposure to it. Instructions forbidding copying do not prove that an agent obeyed them, although uncertainty about training data does not itself prove that an output infringes. Logs, tool access, supplied files and code comparisons therefore become important evidence, and a developer should preserve an honest development history rather than attach the “clean room” label to a repository after the event.
II.What Remedies Are Available?
Against this background, an incumbent such as Adobe or Microsoft may look to copyright, patent and contract law, each of which turns on different facts.
i.Copyright Protects Only Expression, Which May Extend Beyond Literal Code
In the United States, under 17 U.S.C. § 102(b), copyright does not extend to ideas, processes, systems or methods of operation. An image editor does not infringe merely because it offers layers, masks and colour correction. Protected expression can, however, include original code and, where sufficiently original, non-literal structure, graphics, icons, templates and explanatory text. Writing fresh source code is therefore important evidence, but it is not a complete defence to every copyright claim.
Adobe could nonetheless argue that the newly developed software infringes the “look and feel” of its proprietary programs. In particular, the placement of user-interface elements and operational logic, such as sequence-specific keyboard shortcuts, within Thomas’s AI-generated software closely mirrors that of the programs it seeks to replicate. Although copyright protection may extend to a specific, creative arrangement or compilation of data and user-interface designs, a compelling counterargument exists: these interface placements and functional configurations may simply reflect de facto usability standards. In other words, their positioning may be driven by established user expectations and habits rather than by unlawful copying, and may therefore fall within the doctrine of scènes à faire, qualify as fair use (fair dealing in Canada) or reflect standard industry practice.
The U.S. Supreme Court’s decision in Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021), is instructive. The Court held that Google’s reuse of Java’s API declarations, which allowed programmers to carry familiar commands over to a new platform, was fair use. A clone that reads QuickBooks or Photoshop files may find support in that reasoning, at least where the copying is limited to what compatibility requires. The Court did not, however, declare APIs universally unprotected or authorize the wholesale duplication of commercial applications. The purpose, amount and market consequences of copying still matter, and a developer whose product substitutes directly for the original cannot simply rely on the outcome in Google without analysing the facts.
Canadian law adds a further caution. Cinar Corporation v. Robinson, 2013 SCC 73, adopts a qualitative and holistic assessment of substantial taking that can weigh the overall combination of elements. A product engineered for one-to-one parity may therefore be more exposed in Canada than in the United States, although the Court left open the possibility that abstraction-filtration-comparison may assist in cases involving computer programs.
The analysis changes where an agent must defeat licence checks, encryption or similar controls to reach the material it studies, because liability may then arise from the circumvention itself, and not only from copying. In the United States, 17 U.S.C. § 1201(a)(1)(A) prohibits circumventing a technological measure that effectively controls access to a copyrighted work, subject to a narrow exception in § 1201(f) for a person with a lawful right to use a program who circumvents solely to identify the elements needed for an independently created program to interoperate with other programs. Canada’s Copyright Act follows a similar structure. Section 41 defines “circumvent” and “technological protection measure,” and paragraph 41.1(1)(a) prohibits circumventing a measure that controls access to a work. Section 41.12 then permits circumvention of a measure protecting a lawfully obtained computer program in order to make that program, or a device in which it is embedded, interoperable with another program, device or component, or to obtain the information needed to do so. That exception is unavailable, however, to a person who infringes copyright in the process.
ii.Patent Law Can Protect Even Against Independent Creation
Patent law affords software developers such as Adobe a stronger avenue of recourse against those producing functional replicas of their products, particularly because independent creation is not a defence to patent infringement. Established software enterprises typically hold extensive portfolios of presumptively valid utility and design patents covering proprietary applications such as Adobe Acrobat and Photoshop. Consequently, if a competing program practises any patented design, process or functionality, the patent holder may have a viable claim for infringement. Because direct infringement under 35 U.S.C. § 271(a) is a matter of strict liability, liability attaches even if an AI system independently generates the underlying code without access to Adobe’s proprietary source code, rendering even a rigorous clean-room process ineffective as a legal shield. Nor is free distribution an exemption: section 271(a) reaches anyone who makes or uses the claimed invention without authority.
Canadian law is to the same effect. Section 42 of the Patent Act grants the exclusive right of making, constructing, using and selling the invention, and Monsanto Canada Inc. v. Schmeiser, 2004 SCC 34, treats intention as generally irrelevant to use. In either jurisdiction, whether distributing code infringes directly or indirectly ultimately requires a claim-specific analysis.
Patents therefore present a potentially significant legal risk for those seeking to replicate software functionality. Patent protection is not, however, without its own limitations and risks for software developers and patent holders. To enforce a patent, the holder must identify an applicable claim, establish infringement and successfully defend against any challenge to the patent’s validity, a challenge that may ultimately result in its invalidation and the loss of the associated exclusive rights. Moreover, although software is not categorically excluded from patent protection, not every functional aspect of a software product necessarily constitutes patentable subject matter.
iii.Contract May Be the Most Immediate Constraint
Software licences may regulate conduct that copyright law would not otherwise prohibit. Adobe’s General Terms of Use define prohibited reverse engineering to include monitoring inputs and outputs for the purpose of recreating Adobe’s products, and Microsoft’s volume-licensing Product Terms likewise restrict use of its products to replicate their functionality. Such provisions are particularly relevant where a developer uses the licensed application as an automated benchmark for a competing product.
The ArtCraft facts illustrate the importance of contractual evidence. Reports indicate that Thomas had used Adobe products and that the project operated Photoshop through scripts. Whether a particular person or entity assented to a particular version of Adobe’s terms, whether the relevant licence was then in force, and what conduct it covered remain fact-specific questions.
Contractual enforcement is, however, subject to several important limits:
- Privity and assent. Licence terms generally bind only those who accepted them. A downstream developer, or a person relying solely on public information or open-source materials, may never have accepted the software’s terms of use. Notice, assent and the applicable version of the contract must each be established.
- Survival and interpretation. Restrictions do not necessarily survive termination. A survival provision that omits the relevant covenant may create an interpretive issue, depending on the agreement as a whole.
- Territorial differences. Contract and intellectual property laws vary considerably across jurisdictions, including the United States, Canada, the European Union and other international markets. The same software development practice may therefore give rise to materially different legal rights, obligations and liabilities depending on the applicable legal framework.
- Dispute resolution. Arbitration and forum-selection clauses may affect the practical availability of relief.
- Agentic AI liability. It remains unsettled whether the use of an autonomous AI agent displaces the responsibility of the person deploying it. It is equally uncertain whether an agent that accepts the software’s terms of use on its own authority while exploring the program binds its user to that agreement.
III.Implications for Licensing and Transactions
Legislators have yet to address AI-enabled software replication comprehensively, and it is uncertain whether they will do so in the near term. Until they do, software developers and their counsel must take their own steps to protect their products and intellectual property now that replication has become inexpensive and readily accessible. In response, software vendors are likely to rely more heavily on contract, server-side services, proprietary data, integration ecosystems and technical controls. Licence drafting will increasingly focus on the purpose of access, automated observation of inputs and outputs, competitive development, the use of AI agents, and restrictions on training or building substitute systems. Enforceability will depend on assent, statutory rights, territorial limits, and the proportionality and scope of the restrictions.
In practice, a well-structured agreement should address whether prohibited activity includes black-box testing, which automated uses are authorized, who is responsible for agents acting through customer credentials, whether restrictions survive termination, and how amendments are communicated and accepted. Drafting alone is insufficient, however, where the person creating a substitute is not bound by the contract.
M&A and financing diligence must also evolve. A representation that software was “developed in-house” says little about the materials an AI agent accessed or the contractual conditions attached to its tools. Acquirers and investors should therefore seek disclosure of AI development systems, accepted third-party terms, code provenance, clean-room controls, open-source compliance, patent clearance and pending disputes. Where warranted, these inquiries should inform specific representations, warranties, covenants and indemnities.
Finally, patent prosecution must also adapt. Because an AI agent can reimplement a feature in entirely new code, claims tied to a particular implementation are readily designed around, and practitioners should instead consider claiming the functional result at several levels of generality, through method, system and computer-readable-medium claims. Such claims must nonetheless satisfy subject-matter eligibility, be supported by the disclosure and be distinguishable from the prior art. Patent lawyers should therefore bear in mind that a claim drafted more broadly than it can be defended invites invalidation when it is eventually asserted.