AI reverse engineering and copyright: the ArtCraft case

This article reflects the landscape as of October 13, 2026, and the subject moves almost daily, both in what AI can do and in how companies and governments respond, so parts of what follows may change quickly.

Between September 30 and October 7, 2026, the team behind ArtCraft published on GitHub free, open-source versions of seven Adobe applications (Photoshop, Lightroom, Premiere, Illustrator, After Effects, InDesign and Acrobat), along with equivalents of Word, Excel, PowerPoint, AutoCAD and Pro Tools. The code was written in Rust with Claude Opus 5.5, Anthropic’s AI model, which is listed as co-author on most of the changes, and the Photoshop clone alone passed 50,000 GitHub stars.

Contrary to what spread on social media, this is not leaked Adobe code, because the projects present themselves as “clean-room” reimplementations built from public specifications and observed behavior, and all of them are still at an early stage.

How we got here

AI coding tools can now write and test large volumes of code, and some file formats are publicly documented, such as Adobe’s own PSD specification and the Office Open XML standard behind Word and Excel files. A developer describes the expected behavior, compares the output with the original program and asks for adjustments until the two converge, which in theory requires no access to the competitor’s code.

In parallel, REA (“Reverse Engineer Anything”), an open-source tool that connects AI agents such as Claude Code to classic reverse engineering software such as Ghidra, passed 100,000 GitHub stars in October. Game publishers have already pushed back, since an AI-built “Fallout” browser game was abandoned after a notice attributed to Bethesda and Activision had a video of an AI-decompiled Modern Warfare 2 taken down. ArtCraft’s stated trigger was Adobe’s early cancellation fee, the subject of a US$150 million settlement with the U.S. Department of Justice in March 2026.

The Sony precedent and the Macintosh

The debate is not new, and the case that weighs most in it involves a PlayStation emulator. In the late 1990s, Connectix released the Virtual Game Station, which let Sony’s games run on Macintosh computers, and to build it the company studied the console’s BIOS (the basic program that makes it work), making intermediate copies during reverse engineering, although the final product contained no Sony code.

In February 2000, the U.S. Court of Appeals for the Ninth Circuit held that those intermediate copies were fair use, following Sega v. Accolade (1992), and noted that Sony sought control over the market for devices that play its games, a monopoly that copyright law does not confer (Sony Computer Entertainment v. Connectix, 203 F.3d 596). The win did not save the product, because in March 2001 Sony acquired the emulator’s assets and it left the market that year.

That case and its successors, such as Google v. Oracle, decided by the U.S. Supreme Court in 2021, consolidated the idea behind today’s projects: copyright protects code as text, not the function it performs, so one may study a program and write another that does the same, as long as the final product does not copy the original’s expression.

What changes when the clean room is an AI

The clean room began as an evidentiary technique, in which one team studies the program and describes what it does while another, which never saw the code, writes the new one, so that the separation shows there was no copying. When a single developer directs several AI agents trained on vast amounts of public code, that proof becomes harder, and Andres Guadamuz, writing on TechnoLlama, considers that ArtCraft may be on the right side of U.S. law if the process was genuinely clean, while pointing to licenses, patents and trademarks as the weaker spots.

Licenses are the most sensitive point, since Adobe’s terms of use prohibit reverse engineering, including by monitoring a system’s inputs and outputs (section 17). U.S. courts tend to enforce such clauses (Bowers v. Baystate, 2003), while in the European Union Directive 2009/24/EC lets lawful users observe and test how a program works and voids contrary terms (Articles 5(3) and 8), and the Court of Justice has held that functionality and data file formats are not protected (SAS Institute, C-406/10). How closely the interface resembles the original, which reviewers found striking in the Photoshop clone, is another open front.

How Brazilian law sees it

In Brazil, software is protected as a literary work under the Software Law (Law No. 9,609/1998, Article 2), but the Copyright Law excludes ideas, systems and methods as such (Law No. 9,610/1998, Article 8, item I), and the Software Law states that similarity between programs caused by functional characteristics, technical standards or the lack of an alternative form of expression is not an infringement (Article 6, item III). Unlike EU law, Brazilian statutes have no express decompilation exception, and the use of software in Brazil is governed by a license agreement (Article 9), which makes the contract the main battleground. Reproducing code for commercial purposes without authorization remains a crime (Article 12, paragraph 1), and that is the line between independent reimplementation and piracy.

Where things stand, and what to watch

As of this writing, no public statement from Adobe on the projects was known, but commentators expect litigation, and a few precautions already make sense for technology companies doing business in Brazil:

• Software vendors should review the reverse engineering clauses in their Brazilian license agreements and protect what copyright does not cover, through trade secrets, trademarks and patents.
• Companies considering these clones should note that open-source licenses such as Apache and MIT carry no indemnity against third-party claims, no support and no security certifications.
• Teams using AI to build compatible software should document the process, keep the final product free of copied code and honor the licenses they accepted.

If the landscape shifts, through court decisions, government or industry positions, or further advances in AI, this article may become outdated, which is why the date at the top matters.