Creative Coding

AI Decompilation Arrived and Game Preservation Got Years of Work Done in Days

Flash-era Binding of Isaac now runs natively at 1440p and 165fps. A compatibility layer is porting PS5 titles to PC. The same capability that rescued them makes every licence on every binary negotiable.

Decompiling a game by hand is a years-long act of devotion. OpenGOAL, which reconstructed the Jak and Daxter games, and Ship of Harkinian, which did the same for Ocarina of Time and Majora’s Mask, took sustained community effort measured in years, not months — and those are the famous successes, in a field with many more abandoned attempts.

Over the past months that cost has collapsed. Indie developer Popcar, writing on 14 October, catalogued the result: AI-assisted workflows are decompiling games in a matter of days, the output is messier than hand-written reconstructions, and it compiles and runs anyway.

Popcar is worth quoting on their own position, because it is not an enthusiast’s: “I generally don’t like AI… I hate that the internet is full of slop and bots, I hate that it’s used as justification to fire hundreds of thousands of workers.” And then: “every now and then there’s AI news that becomes impossible to ignore.”

What decompilation actually is

Compilation throws information away. Source code becomes machine code; names, types, comments and structure are discarded because the CPU does not need them. Decompilation is the attempt to reconstruct source from the binary — not to run it in an emulator, but to recover code you can read, change and recompile.

The distinction from emulation is the whole point. An emulator reproduces the original hardware so the original binary can run unmodified; you inherit the original’s frame rate, resolution and bugs. A decompilation gives you source, and source can be changed. That is why decompiled ports can do things emulators structurally cannot: native resolutions, uncapped frame rates, modern input handling, modding hooks.

It has always been possible by hand. It has never been quick.

The projects

IsaacRecomp ports the original 2011 Flash version of The Binding of Isaac to C++ — not the remake, the Flash original, which ran at a locked 30fps in a low, blurry resolution and had become unstable for the obvious reason that running Adobe Flash in 2026 is an act of hope. The port runs at native 1440p, interpolated to 165fps.

Popcar’s framing of that: “it was all made automatically. In a few days. By an AI.”

Others in flight or finished:

  • OpenCMR2 — Colin McRae Rally 2.0, another title that barely ran on modern systems
  • FFTARecomp — Final Fantasy Tactics Advance, now at arbitrary aspect ratios with dynamic zoom
  • MetroidPrimePort — Metroid Prime
  • A Ratchet and Clank port described as near completion

And the most consequential: AnyPS5, a project aiming to automatically port any PS5 game to PC. This is explicitly not an emulator — it is a compatibility layer, closer in spirit to Proton, which lets Windows games run on Linux. Animal Well has already been ported this way.

Popcar counts “hundreds of these projects now, with more coming every day”, and admits to having lost track.

Why this is the strongest preservation argument in years

The preservation case is not sentimental. Console makers have moved decisively to digital distribution, and games whose online components are hosted become entirely unplayable when those servers shut down. That is not a hypothetical; it is the normal end-of-life for a modern title. A purchased game can simply stop existing.

Decompilation changes the terms. A game with available source can be ported to current hardware, kept building against current compilers, and — as several of these projects do — compiled to WebAssembly so it runs in a browser with no installation at all. Popcar reports Skate 3 and Black Ops Zombies running that way, the latter at a bad frame rate, which rather proves the point: it runs at all.

“We could own software we bought regardless of what the company thinks.”

There is a secondary effect for creative practice. Modding has always been gated on someone first working out where a game’s code can be hooked — which is why some games have enormous mod scenes and structurally similar ones have none. If models can locate those seams automatically, previously unmoddable games acquire modding frameworks, and the gate moves.

The early output of that is predictably silly. Popcar’s example of the genre is a mod putting Half-Life’s Gordon Freeman into Jak and Daxter — and notes, carefully, that this one was not made with AI. It runs on OpenGOAL, the hand-made decompilation. The irony is worth keeping: the mashups getting attention right now are running on top of years of unpaid human reverse-engineering.

And the same capability points the other way

Here is where the piece stops being good news.

Everything above is a general software capability that happens to have been pointed at games first because games are where the enthusiast labour is. It applies to any binary.

Popcar states the implication without hedging: “Anybody can decompile other software and use it any way they’d like, and nobody can stop them. Companies can steal from competitors. Your paid game or app can be open-sourced against your will.”

The legal position is awkward in a specific way. Decompilation for interoperability has long-standing legal protection in many jurisdictions, and reverse engineering is not itself unlawful. The new problem is evidentiary: if a model produces functionally equivalent code with different names and structure, the resemblance that copyright infringement is normally proved by may not be there to find. Someone has already built a joke website on the premise — copy code with a model, declare it a clean-room implementation, claim no licensing obligations attach.

The joke is doing serious work. Clean-room reimplementation is a real and legitimate practice, and its legitimacy rests on a procedural claim about what the implementers were allowed to see. That claim is now unfalsifiable from the outside.

Which means the exposure is not symmetrical, and it falls hardest on the people this site writes about. Open-source licences depend on being enforceable. The GPL works because copied code is identifiable, and a project that strips a copyleft notice can be shown to have done so. If equivalent code can be produced that no comparison identifies as derived, the practical enforceability of every copyleft licence weakens — and the losers are individual developers and small studios who licensed their work on the understanding that reciprocity could be required. A large company with litigation budget is far better insulated than a solo developer who released a library under the GPL.

Popcar expects “some major application” to get decompiled and shared within months, and a serious corporate response to follow. The plausible responses are all unpleasant: more invasive DRM, more hostility to running software you own, more pressure to keep anything valuable server-side where no binary ever reaches the user.

What to take from it

The honest position is the one Popcar lands on: the preservation win is real and worth having, and the licensing problem is not a side effect that can be legislated away separately. They are the same capability, and nobody gets to keep only the half they like.

If you make creative software, the practical implications start now. Your binary is readable. A licence is a social and legal instrument, not a technical one, and it was always more fragile than it looked — it is just that the gap between the two has become visible.

If you care about old games, go and look at what has been rescued in the last few months. Some of it is work nobody was ever going to fund.

Sources