# They Cloned Photoshop with AI. Is This the End of Software as We Know It? > In just a few days, open-source clones of Photoshop, Illustrator and Premiere written with AI showed up on GitHub, along with browser versions of Halo and GTA derived from the original games. AI-powered reverse engineering is scary, but I think the problem is overrated: software has always been copied, and what makes the difference is still excellence, customer relationships and the law. - URL: https://mautoblog.com/en/posts/ai-software-cloning-adobe-reverse-engineering-2026/ - Language: en - Published: 2026-10-14 - Author: Matteo - Categories: Software Development, Artificial Intelligence - Tags: AI-Assisted Development, Open Source, Claude, Career - Translation (it): https://mautoblog.com/posts/ia-clonazione-software-adobe-reverse-engineering-2026/ --- Over the last few days, two kinds of headlines have filled my feed. The first: *"Developer clones Photoshop with AI"*. The second: *"Halo, GTA Vice City and The Simpsons run in your browser thanks to AI"*. Underneath both, the same anxious question: **is software over?** I read a fair number of articles, looked at the repositories and tried to put things in order. My short answer is no. The long answer is this post. --- ## What actually happened ### The Adobe suite clones Brandon Thomas, a developer from Atlanta, published **ArtCraft** on GitHub: a family of seven open-source applications that recreate Adobe's main products, including **PhotoCraft** (Photoshop), **VectorCraft** (Illustrator), **FilmCraft** (Premiere) and **LightCraft** (Lightroom). They're written in Rust with help from Claude Opus 5.5 and presented as *clean-room* implementations, meaning without a single line of Adobe code ([Vice](https://www.vice.com/en/article/everything-we-know-about-the-free-adobe-clones-built-with-ai-do-they-actually-work/), [HotHardware](https://hothardware.com/news/dev-uses-ai-to-build-clones-of-photoshop-and-entire-adobe-suite-for-free)). His motivation, he says, was frustration with Adobe subscriptions and cancellation fees, the same practices the US Department of Justice sued the company over in 2024. According to figures reported over the past few days, the PhotoCraft repository passed **50,000 stars** in no time. But it also has **more than 800 open issues**: PSD file problems, crashes, rudimentary color management, excessive memory usage on large images. One very clear-eyed technical analysis compared it to a Ferrari replica built on a VW Beetle: the body is right, the engine isn't ([AkitaOnRails](https://akitaonrails.com/en/2026/10/13/ai-reverse-engineering-is-adobe-dead/)). ### Games in the browser At almost the same time, [Kotaku](https://kotaku.com/we-might-be-cooked-as-these-vibe-coded-web-browser-ports-of-halo-the-simpsons-hit-and-run-and-gta-vice-city-seem-to-work-perfectly-2000743300) covered a wave of browser versions of games such as **Halo: Combat Evolved**, **GTA Vice City**, **The Simpsons: Hit & Run** and **Call of Duty: Black Ops**, obtained by decompiling the original games with AI assistance. A GTA 5 version went viral and disappeared shortly after. Publishers are responding with the usual cease-and-desist letters and takedown requests, one at a time. Two facts worth keeping in mind, because they'll come in handy: - which models produced each port **is not verified**: the attribution to Claude Opus 5.5 comes from the articles, not the authors; - many of these ports use **maps, models and animations from the original games**. --- ## Software has always been copied The first thing I want to say is: welcome to the history of computing. - The windowed interface was born at **Xerox PARC** in the '70s. Apple brought it to the Lisa and the Macintosh, Microsoft to Windows. Apple sued Microsoft and **lost** (1994): a generic similarity of ideas isn't enough. - **VisiCalc** invented the spreadsheet in 1979. Then came **Lotus 1-2-3**, then **Excel**, then **Google Sheets**. Each one copied the previous idea and improved on it. - Lotus sued **Borland** because its spreadsheet replicated the 1-2-3 menu structure. In 1995 the appeals court sided with Borland: the command hierarchy is a "method of operation", not a protected work. - **SQL** was born at IBM, and within a few years Oracle, Microsoft, MySQL and PostgreSQL implemented it. - More recently, the US Supreme Court ruled that Google could reimplement the Java APIs in Android (**Google v. Oracle**, 2021). Copyright protects **code** (the expression), not **ideas**. For people selling software that's a problem, but for the discipline it's a huge advantage: computing moves at a speed that's unthinkable in more locked-down industries, such as pharmaceuticals. > ⚠️ **An important clarification:** it isn't entirely true that software "can't be patented". In the United States software patents exist, even if they've been harder to obtain since the *Alice* ruling (2014). In Europe programs "as such" are excluded, but software-implemented inventions with a technical effect are patentable. And the clean-room technique protects you against copyright claims, **not against patents**. In this picture, language models don't invent anything. They **democratize** a process that used to take months of work by highly specialized people. A confession: while writing this article I came across a project called **loom** on an MCP app store, a local memory for AI agents built on SQLite. For a moment I thought "they cloned me!" (my own [Loom](/en/posts/loom-local-memory-for-llms-without-docker-rag-2026/) does more or less the same thing). Then I checked the dates: their first commit is five days *before* mine. Different language, different architecture, same name, same idea. If anyone copied, going strictly by the calendar, it would be me. I didn't: it's just that in 2026 the idea was in the air, and when an idea is in the air, lots of people write it at the same time. --- ## Copying, reimplementing, decompiling: not the same thing This is where the biggest misunderstanding in the debate lies. **Decompiling and recompiling is not reimplementing.** If you take a program's binary, turn it into source code (with or without AI) and recompile it, you still have someone else's program. Running the code through a model doesn't make it yours. In Europe decompilation is only allowed in narrow cases, for interoperability (Directive 2009/24/EC). In the United States, rulings such as *Sega v. Accolade* (1992) treated it as fair use for understanding how a system works, not for distributing a copy. For games the problem is even more obvious: a game isn't just code, it's **artwork, music, levels, characters**. A browser port that ships Halo's maps and models is distributing protected material, period. It's no coincidence that serious projects of this kind, such as OpenMW for Morrowind, rewrite the engine and ask you to **supply the files from the game you bought**. **A clean room is something else.** It's an old technique (Phoenix used it in 1984 to recreate the IBM PC BIOS) that separates two groups: 1. the first studies the original product and writes a **specification** of its behavior; 2. the second, which has never seen the original, **implements from the specification alone**. With language models it becomes: one instance analyzes the code, documentation and behavior and produces the specification; another instance, with no access to the first one's logs, writes the new code. This way you could rebuild, for example, a Redis-compatible database under a different license, starting from its documented behavior. Does it work? Probably yes, with two caveats worth knowing: - **the second model isn't really "clean"**: if the original is open source, it's very likely in its training data. That's the objection raised in March over *chardet*, a Python library that moved from LGPL to MIT after an AI rewrite ([Simon Willison](https://simonw.substack.com/p/can-coding-agents-relicense-open)); - **no court has ruled yet** on whether two sessions of the same model count as two separate teams. --- ## What's still protected (and what isn't) To keep it simple: | Can be replicated | Risky or forbidden | |---|---| | Features (a command that draws a circle through two points) | The code, even if decompiled and recompiled | | Well-known, documented algorithms | An identical interface: icons, layout, overall look | | Conventions and shortcuts users expect | A game's artwork, music, levels and characters | | File formats, for interoperability | Trademarks and product names | | Languages and APIs (Google v. Oracle) | Patented technologies | On formats, there's a European ruling every developer should know: *SAS Institute v. World Programming* (EU Court of Justice, 2012). Functionality, programming language and file formats **are not protected by copyright**. For me, the most interesting part of the table is the third row on the left. You can replicate conventions so you don't throw off people coming from Photoshop, and then change everything else: make it simpler, faster, better suited to a certain kind of user. Copying exactly, on the other hand, isn't just risky: it's almost always pointless too. --- ## The moat isn't the code Here's something that's almost always missing from the articles about these clones. Photoshop's core algorithms aren't secret: layers, masks, blend modes and convolution filters have been documented for decades, and language models know them very well. Today a pair made of **an expert designer and a programmer** really can rebuild, and rethink, a complex product. But Adobe's value isn't there. It lies in much less glamorous things: - **file compatibility**: correctly opening a ten-year-old PSD with a hundred layers and the right effects. Not by chance, that's where PhotoCraft has around a hundred open issues; - **color management** and professional print standards (ICC profiles, CMYK, libraries like Pantone); - **proprietary models and data** for generative features, which you can't derive from code; - **ecosystem**: plugins, fonts, shared libraries, training, certifications; - **maintenance**: writing version 1 costs very little today, maintaining it for ten years costs as much as ever. PhotoCraft's 800 issues are exactly that. In highly specialized software, such as professional video editing with neural models and complex algorithms, a clone might reach 70%. 70% is enough for the vast majority of people, but professional studios will keep buying the original. Because the real protection there is the **business-to-business relationship**: no company puts a project found on GitHub into production unless it becomes very popular and reliable. --- ## Big companies have stopped surprising us There's a contextual factor that, in my opinion, explains half the enthusiasm for these clones. There was a time when every new version of Photoshop, Office or an operating system was an event: people waited for it, talked about it, and it brought new ideas. Today it's mostly a subscription update, with a few AI features added and a price that keeps going up. If a user is willing to try a bug-ridden alpha just to get away from Adobe, the problem isn't reverse engineering: it's the customer relationship. Small teams, on the other hand, can still surprise. With AI they can do it starting from a foundation that used to take years of work. --- ## So is this the end of software? For people who make a living from software the problems are real, no point beating around the bush: - if reimplementing is cheap, **fewer companies will manage to make money** selling licenses; - fewer companies making money means **fewer developer jobs**. On the other side there's a potentially huge effect: much broader access to powerful, flexible tools that can be adapted to your own needs. The photographer who only wants RAW development without paying a subscription every month, the studio that wants an editor tailored to its workflow, the school that can't pay licenses for a hundred workstations. Overall I think the balance can be positive. And we've already seen it happen: - **Linux** wiped out much of the commercial Unix market and contributed to the decline of Sun, which ended up inside Oracle; - **MySQL, PostgreSQL and Redis** eroded the market for commercial databases and caches; - **PHP** and open-source CMSs reduced the market for paid website tools to almost nothing. A lot of commercial value was destroyed, and yet the industry came out bigger than before. Those who survived, like **IBM**, did so by staying tied to business processes and large contracts, not to code. There's no guarantee it'll go the same way again: this time the blast could be bigger and faster. But people doing truly excellent work will find room. --- ## A possible return to open source There's also a scenario I find fascinating. In the '90s there weren't many jobs in computing, and lots of people spent their free time writing free software. Paradoxically, with the boom of well-paid programmers, "garage" open source declined: people with a good salary had less time and fewer reasons to do it. If AI reduces the number of jobs, we might see the opposite movement: developers starting **small companies built on free software**, with licenses that still allow them to earn through business contracts, support, donations and sponsorships. The risk is fragmentation: a thousand Photoshop clones, none of them really finished. But the history of Linux, Blender and Krita shows that once a project crosses a certain threshold, the community tends to concentrate there. --- ## The real risk: more laws, not fewer Today the only real protections are legal. On this point the American and European systems are fairly similar: features can be copied, an identical interface can't. But enforcing your rights costs money: you need lawsuits and a budget. Big companies can afford it; the small developer whose software gets cloned, much less so. And this is where I see the biggest risk. When the mechanics of a market change, those who lose out ask for stricter rules: more patents, more things covered by copyright, maybe specific bans on AI-powered reverse engineering. That would be a step backwards for an industry that became what it is precisely because ideas circulated freely. --- ## Conclusion AI-powered reverse engineering isn't an invention of language models: it's a practice as old as computing, and now it's within reach of many more people. Safety filters or not, there are open models and a very competent community: there's no going back. But copying code is still illegal, and copying ideas is still as legal as it has always been. What changes is the cost of the first version. And when the first version is cheap, the things that mattered at the beginning start to matter again: - **excellence** and the ability to surprise; - **knowing your customers**: a clear target, a reasonable price instead of punitive subscriptions, regular updates; - **listening to feedback** and turning it into useful features, fast. If Adobe is afraid of an alpha with 800 open issues, the problem isn't AI. What do you think? Would you use an open-source clone of Photoshop or Premiere for work, or just out of curiosity? Let me know in the comments. --- *Sources: [Vice](https://www.vice.com/en/article/everything-we-know-about-the-free-adobe-clones-built-with-ai-do-they-actually-work/), [HotHardware](https://hothardware.com/news/dev-uses-ai-to-build-clones-of-photoshop-and-entire-adobe-suite-for-free), [AkitaOnRails](https://akitaonrails.com/en/2026/10/13/ai-reverse-engineering-is-adobe-dead/), [Kotaku](https://kotaku.com/we-might-be-cooked-as-these-vibe-coded-web-browser-ports-of-halo-the-simpsons-hit-and-run-and-gta-vice-city-seem-to-work-perfectly-2000743300), [Simon Willison on the chardet case](https://simonw.substack.com/p/can-coding-agents-relicense-open).*