AI Cloned Photoshop. What's Left of the Software Moat?
Open-source AI clones of Photoshop show software can now be rebuilt by watching it run. Here's what that means for software moats.
Can AI really clone commercial software like Photoshop?
Yes, and it already has. Someone pointed an AI agent at a local copy of Photoshop, had it click through every tool and panel, record what each action did, and rebuild the same behavior from scratch. The result is an open-source clone called PhotoCraft. No one read Adobe’s source code. The AI just watched the software behave and reverse engineered the outcome. The same approach is being used on games, utility apps, and other paid tools, and it’s forcing a hard question for anyone building a software business: if the interface and features can be observed and rebuilt, what’s actually defensible?
TL;DR
- PhotoCraft, an open-source Photoshop clone, was built by having AI observe a running copy of the software and recreate its behavior rather than reading any original code.
- The same technique is being applied to video game decompiling, where projects that once took human teams years (like Super Smash Bros. Melee) are now being replicated in weeks or days.
- A GitHub tool called Reverse Engineer Anything (REA) wires AI agents into established decompiling tools like Ghidra, Hopper, and IDA, letting an agent inspect and rebuild apps such as Apple’s Preview or TV remote software.
- Clones built this way are not line-for-line copies of the original code, since compiling strips comments, renames variables, and restructures logic, but they can look, feel, and behave identically to users.
- This works because software has a verifiable goal: you know what the output should look like, so an AI agent can keep guessing, checking, and iterating until it matches.
- One video creator’s side-by-side test found a vanilla coding agent without any specialized reverse-engineering library produced a comparable clone of a Mac app, suggesting the underlying agent, not the tool, is doing most of the work.
- The practical consequence is that software-only moats are weakening, and value is shifting toward services, data, trust, and ongoing support that can’t be copied by just watching an app run.
How does AI decompiling actually work?
Software starts as source code that a human can read, then gets compiled into machine code that computers run efficiently. Reversing that process, called decompiling, means starting from machine code and working backward to recover something resembling the original logic. Historically this took specialists years of manual guesswork: write a theory about what a chunk of code does, test it, revise, repeat.
AI turns that into an automatable loop. An agent proposes code, checks whether it produces the correct behavior, and keeps iterating until it matches. This is the same loop structure that shows up across agentic coding tools: give the model a verifiable end state and let it run until it gets there. Decompiling fits that pattern almost perfectly, because you can always check whether the reconstructed software behaves like the original.
Speed comes from parallelism and raw throughput rather than cleverness alone. Chunks of the problem get split across many agents running simultaneously, and AI writes and tests code far faster than a human ever could. One widely cited example: a decompiling project for the game Snowboard Kids that originally took close to two years was redone with AI assistance in a fraction of that time, and that timeline keeps shrinking as models improve.
Why doesn’t a decompiled clone match the original code exactly?
Because compiling throws information away. Comments disappear. Variable and function names vanish. The compiler simplifies and rearranges logic to make it efficient for a machine to execute, not for a human to read later. A function that literally says “return 2 + 3” might get compiled down to just “return 5,” since the two are mathematically identical. When an AI decompiles that machine code, it has no way of knowing the original was written as an equation rather than a flat value, so it reconstructs the simpler version.
This matters for the moat question because it means these clones aren’t piracy in the traditional sense of stolen code. They’re independent reconstructions that happen to produce the same user experience. The code underneath can differ completely while the software “feels” identical to someone using it, which is really the only thing that matters commercially.
What is Reverse Engineer Anything (REA) and why did it go viral?
REA is a tool that connects AI coding agents to established reverse-engineering software: Ghidra (released by the NSA around 2019), Hopper (from 2011), and IDA (dating back to the 1990s). Instead of a human manually operating these decompiling tools, an agent can query them directly, pull out pseudocode and behavioral notes, and use that information to iteratively rebuild an app’s functionality.
It spread quickly because of eye-catching demos: people rebuilding an Apple TV remote app in under an hour, or attempting to reproduce entire creative software suites. But a closer look suggests the tool itself isn’t the breakthrough. The underlying decompilers are old, sometimes decades old. What’s new is that an AI agent can operate them fluently and tirelessly.
One direct comparison tested this: building a clone of a simple Mac photo viewer app using REA’s full toolchain against building the same clone with a general coding agent and no specialized reverse-engineering library at all. Both approaches produced apps that looked and functioned almost identically to the original, with the REA version finishing with some extra fidelity but taking considerably longer. The implication is that the heavy lifting comes from the general intelligence of the coding agent itself, and the specialized tooling provides a speed and accuracy boost rather than a capability enabled only by that tool.
Is the software moat actually gone?
Not entirely, but the part of the moat that used to be “nobody else can build this” is eroding fast. If an AI agent can watch any application run, infer its features, and rebuild comparable functionality without touching the original source, then the software itself stops being the scarce asset. This is already playing out with “wrapper” apps, products built as thin interfaces on top of models like Claude or GPT, which get swallowed as the underlying models add the same features natively.
What doesn’t get cloned by watching an app run: a company’s accumulated customer data, enterprise trust and compliance history, support relationships, integrations into existing workflows, brand recognition, and ongoing updates tied to a roadmap only the original team controls. Those become the actual differentiators once the interface and feature set are commoditized. A business whose entire pitch is “we built this software” is more exposed than one whose pitch includes service, maintenance, data, or community built around the software.
There are also real limits worth naming. Intellectual property law doesn’t go away because an AI reconstructed something independently; selling a cloned commercial app is still legally risky. And several of the loudest viral demos involve small, well-scoped utility apps (a photo viewer, a remote control app) rather than sprawling enterprise software with years of edge cases, integrations, and regulatory requirements baked in. Cloning the surface of an app is not the same as cloning everything a mature product has accumulated over a decade of real-world use.
What should builders take away from this?
Treat the user interface and feature list as commodity, not asset. If a competitor (or an AI agent run by anyone with a laptop) can observe your product’s behavior and approximate it in days, the thing worth protecting is whatever sits outside the visible software: proprietary data, domain expertise baked into workflows, trust from regulated customers, or a support relationship a model can’t replace. Software still has value. It just isn’t a moat on its own anymore.
Frequently Asked Questions
Did AI actually read Adobe’s Photoshop source code to clone it?
No. The clone, PhotoCraft, was built by having AI observe a running copy of Photoshop and infer behavior from its outputs, not by accessing Adobe’s proprietary source code.
Is a decompiled AI clone identical to the original software’s code?
No. Compiling strips out comments, variable names, and structural detail, so an AI reconstruction typically produces different underlying code that behaves the same way to a user.
What is Reverse Engineer Anything (REA)?
It’s a tool that connects AI coding agents to established decompiling software like Ghidra, Hopper, and IDA, letting an agent inspect an application’s internals and iteratively rebuild similar functionality.
Do you need specialized tools like REA to reverse engineer an app with AI?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Not necessarily. Tests comparing REA against a general-purpose coding agent with no added tooling found both could produce comparable clones of a simple app, suggesting much of the capability comes from the underlying AI agent itself.
Does this mean software companies have no moat at all?
Not entirely. The interface and feature set are becoming easier to replicate, but things like proprietary data, enterprise trust, support, and deep integrations remain hard to clone just by watching software run.