Editorial
Retro gaming has always been a fight against time.
Cartridges fail. Disc drives die. Online stores close. Original hardware becomes expensive, temperamental or simply difficult to connect to a modern television. For decades, the answer has largely been emulation, official re-releases and the extraordinary patience of preservation communities willing to reverse-engineer old games by hand.
Artificial intelligence is beginning to disrupt that equation.
AI coding agents are now being used to help turn games built for decades-old consoles into native applications for modern PCs. The results range from genuinely impressive technical experiments to projects that veteran retro developers dismiss as poorly understood “slopcomps.”
And a fight that erupted around an AI-assisted Banjo-Tooie recompilation this week has exposed the question the retro community will increasingly have to answer:
If AI can bring an old game back faster, does that automatically mean it has preserved it better?
The Banjo-Tooie fight is bigger than one N64 game
The immediate controversy surrounds an unofficial Banjo-Tooie recompilation whose developers disclosed that AI tools generated substantial portions of its code.
Darío, the developer behind the acclaimed Banjo: Recompiled and Zelda 64: Recompiled projects, publicly distanced his work from the new project. He argued that high-quality recompilation requires care, maintainable code and an understanding of the human research underpinning the tools being used.
His criticism was unusually sharp: the rush toward AI-generated recompilations, he argued, risks becoming a “race to the bottom.”
That reaction matters because this is not simply another argument about whether generative AI belongs in games.
Nobody is asking AI to invent Banjo-Tooie.
The game already exists.
The question is whether AI should be trusted to help reconstruct the technological bridge that keeps it playable for another generation.
Recompilation is changing what a retro port can be
Emulation and recompilation solve similar problems in very different ways.
An emulator recreates enough of an old console’s hardware environment in software for the original game to run. Your modern PC is, in effect, pretending to be the machine the game expected.
Recompilation can instead translate or rebuild game code so that it runs directly on modern hardware. Decompilation-based source ports go further by reconstructing human-readable source code and compiling that for contemporary platforms.
Once the game is no longer tightly bound to its original machine, interesting things become possible.
The Retro Porting Toolkit currently catalogues dozens of native-port projects across multiple platforms. Its projects demonstrate features such as widescreen rendering, modern controls, higher frame rates, translations, bug fixes, save states, rewind and modding.
That’s a fundamentally different vision of preservation.
Instead of merely recreating an N64 accurately enough to play an N64 game, the goal can become preserving the game’s logic while allowing the experience around it to evolve.
GoldenEye is an extraordinary AI experiment
One of the clearest examples arrived in September.
Developer JK Dansereau released an AI-assisted native PC port of the original 1997 GoldenEye 007. Version 0.4.0, released September 28, runs the full campaign at 60fps on Windows and Linux, including Steam Deck, although the developer stresses that it remains a pre-1.0 project.
What’s particularly interesting is how openly the project documents its AI workflow.
Dansereau describes using two coding agents: a locally hosted open-weight model running on a single RTX 5090 and Claude as a hosted frontier model. The agents handed work back and forth through shared written notes while one human directed the project.
This was not a simple “make GoldenEye run on PC” prompt.
The agents had to deal with roughly 230 translation units of reconstructed Nintendo 64 C code, assumptions about the N64’s big-endian 32-bit MIPS architecture, modern little-endian 64-bit PCs, graphics hardware behaviour and bugs that could manifest far away from their actual cause.
The developer describes the port as much as an experiment in agentic software development as a GoldenEye project.
That distinction is important. AI didn’t magically discover GoldenEye’s source code. The port builds on an existing human-created decompilation and years of accumulated understanding of the original hardware and game.
AI accelerated a bridge that human preservation work had already made possible.
This is where AI could transform preservation
Now imagine that workflow applied beyond famous Nintendo and Rare games.
The most important beneficiaries of faster porting may not be Mario, Zelda or GoldenEye. Those games have huge communities and strong commercial value. Somebody is likely to keep them playable.
The real opportunity is the forgotten middle of gaming history.
Thousands of games are trapped on discontinued consoles with no modern PC version, no current storefront release and little financial incentive for their original publisher to fund a remaster.
Historically, rescuing one of those games could require a small group of specialists to spend years reverse-engineering hardware behaviour and code.
If AI agents eventually reduce portions of that work from years to months, or months to weeks, the economics of preservation change.
A niche game no longer needs an enormous fan community to justify the labour.
A handful of technically capable people, assisted by increasingly competent coding agents, may be enough.
But “it boots” is not the same as preservation
This is where the Banjo-Tooie dispute becomes important.
AI coding tools are exceptionally good at producing code that looks plausible. Retro ports require something harder: code that behaves correctly in thousands of edge cases and remains understandable enough for somebody to maintain years later.
A game can launch and still be wrong.
Physics can behave differently above the original frame rate. Audio can drift out of sync. Collision can fail in one obscure room. A save file can corrupt after dozens of hours. A renderer modification can break when the framework beneath it receives an update.
For a normal software prototype, “mostly works” can be progress.
For preservation, the original game is the reference.
If an AI-assisted port changes its behaviour in ways nobody understands, we’ve preserved something that merely resembles the original.
The real issue isn’t AI versus humans
The easiest version of this debate is also the least useful.
One side says AI can make ports dramatically faster, so resisting it is nostalgia for old development methods. The other says AI-generated code is inherently low quality and has no place in preservation.
Neither position captures what’s actually happening.
The GoldenEye project shows that AI agents can contribute to difficult low-level engineering when a knowledgeable human directs them, documents the process and tests the result.
The Banjo-Tooie controversy shows what established developers fear when speed becomes the primary measure of success: unreadable code, fragile dependencies, poor attribution and projects that become impossible to maintain once their original creator moves on.
The future is therefore unlikely to be “AI replaces retro developers.”
It is more likely to be retro developers who understand the hardware gaining dramatically more leverage from AI.
AI also inherits decades of human work
There is another uncomfortable part of the conversation.
Modern AI-assisted recompilation does not begin from zero.
It stands on decompilation projects, emulator research, renderer frameworks, documentation, reverse-engineered hardware behaviour and countless fixes contributed by people who sometimes spent years solving problems nobody else wanted to solve.
When an AI agent can suddenly produce a port in a fraction of the time, it can create the illusion that the machine solved the entire problem.
Usually it didn’t.
It arrived after humans had already mapped much of the territory.
That makes attribution especially important. AI can reduce the visible labour involved in a project without reducing the amount of historical labour the project depends on.
Copyright doesn’t disappear because AI wrote the code
AI also doesn’t solve the legal complexity surrounding unofficial ports.
Many recompilation and source-port projects deliberately avoid distributing copyrighted game assets and require users to provide their own legally obtained game files or ROMs. GoldenEye’s current AI-assisted project, for example, distributes neither a GoldenEye ROM nor the game’s assets.
That design can reduce some obvious infringement risks, but the legal position of any particular reverse-engineering project depends on jurisdiction, implementation and what exactly is distributed. AI-generated code is not a magic exemption from copyright law.
If anything, dramatically lowering the barrier to creating unofficial ports could make those tensions more visible as more projects appear.
Retro gaming could become less about emulating hardware
For players, the long-term consequence could be profound.
Imagine buying or dumping a legally owned copy of an old game and feeding its data into an open native runtime.
The game launches instantly on Windows, Linux or a handheld PC. It understands modern controllers. It supports ultrawide displays and high refresh rates. Community patches fix decades-old bugs. Accessibility options can be added. Translation projects become easier. Mods no longer have to fight against the limitations of a machine designed in the 1990s.
At that point, retro gaming becomes less about keeping an old console alive and more about keeping the software alive.
That may be recompilation’s biggest philosophical change.
Preservation should be measured in decades, not GitHub stars
The excitement around AI game ports is justified.
Watching a single developer use modern coding agents to help move GoldenEye’s reconstructed N64 code onto contemporary PCs would have sounded implausible only a few years ago.
But preservation has a different timescale from most technology trends.
The important question isn’t whether an AI-assisted port gets thousands of downloads this week.
It’s whether somebody can understand, repair and compile it ten or twenty years from now.
That is why the current fight over “vibe-coded” ports matters. The retro community isn’t merely arguing about whether programmers should use AI. It is negotiating what standards should apply when AI makes it possible to produce preservation projects faster than ever before.
Used carefully, AI could become one of the most important preservation tools gaming has ever had. It could give obscure games a second life, help tiny teams tackle impossible codebases and allow classic software to outlive the machines it was designed for.
Used carelessly, it could produce an avalanche of ports that work just well enough for a viral video and then become abandoned technical debt.
The technology will probably deliver both.
And that means the defining question of the AI retro-gaming era won’t be “Can AI port this game?”
Increasingly, the answer will be yes.
The question that matters is:
Will that port still deserve to be called preservation?
Sources: GoldenEye 007 PC Port project and agentic-development documentation; Retro Porting Toolkit; reporting from Video Games Chronicle and FRVR on the Banjo-Tooie recompilation dispute. This article is editorial analysis. Featured image: Lorenzo Herrera / Unsplash.
