AI Could Make Games Last Forever. But What Else Could It Learn to Rewrite?
For years, one of the possibilities I have found most exciting about artificial intelligence in gaming has had surprisingly little to do with better graphics or computer-controlled characters becoming harder to beat. It is the possibility that games might eventually stop running out of things for us to do.
Imagine finishing Baldur's Gate 3, only for the game to create another properly structured adventure based upon the characters who survived, the decisions you made and the state in which you left its world. Perhaps Space Marine 2 could receive two entirely new Operations every month, generated around existing locations, enemies and storylines without a development team having to manually build every encounter. An open-world RPG could continually create new characters, conflicts and objectives that actually remembered what happened previously.
Instead of buying a game with 80 hours of content, we could eventually buy a world.
That possibility has always seemed fascinating. What I had not expected was how quickly some of the technology around it would begin looking plausible.
AI is beginning to understand the game, not just play it
Some of the most eye-catching experiments appearing recently have involved AI-assisted tools analysing existing video games and then helping developers alter them in ways that would normally require considerable specialist knowledge.
One recent example is SkyCraft, an extraordinary project that effectively connects Minecraft and The Elder Scrolls V: Skyrim. Minecraft runs alongside Skyrim in the background, with the two games exchanging information constantly. The result allows Minecraft-style movement, blocks, weapons and inventory systems to operate inside Skyrim's world while Skyrim continues handling its quests, characters and environments.
It is important not to exaggerate what happened here. An AI did not simply look at Skyrim, understand everything Bethesda had built and spontaneously transform it into Minecraft. Human developers, existing modding tools and years of community knowledge remain enormously important.
What is changing is how much of the difficult work AI systems can increasingly help perform. Projects such as Universal Modder are explicitly designed to let coding agents identify which engine a game uses, inspect its underlying systems, determine an appropriate modification route, build code and test the result inside the running game.
That might sound like a niche development for people who enjoy breaking Skyrim in increasingly inventive ways, but something much bigger is happening underneath.
AI is getting better at working out how complicated software behaves.
The speed is the slightly frightening part
Perhaps the most striking recent example comes from a project rebuilding the 1990 game Midwinter. The team behind the remaster says it gave a newer AI model the game's original executable and asked it to analyse how the software worked.
According to the developers, the model mapped 602 reachable functions overnight and reconstructed major systems covering terrain generation, enemy behaviour, vehicles, sound, file formats and the rules governing how the game was won or lost. The team stresses that this was not a magical recovery of the game's original source code, and some findings still required verification, but they also say similar work using earlier AI tools had previously taken around six months of piecemeal effort.
Midwinter is obviously not a modern AAA game. A title released in 1990 is dramatically smaller and simpler than something such as Cyberpunk 2077, Baldur's Gate 3 or Grand Theft Auto V.
Yet the interesting part is not whether AI can completely understand the largest game ever created today. It is the direction in which the capability is moving and, particularly, how quickly the gap appears to be closing.
That is where my excitement starts acquiring a slightly darker edge.
The dream of the game that never ends
There is already serious research into games where generative AI operates while we are playing rather than merely helping developers create them beforehand.
A 2026 survey of so-called AI-native games found systems already experimenting with generated dialogue, quests, characters and worlds at runtime. Researchers describe one of the central challenges as keeping those generated experiences tied to stable rules, goals, game state and player choices. In other words, generating something is becoming easier. Generating something that still feels like a coherent game is the difficult part.
Microsoft and Google are approaching the same general problem from another direction. Microsoft's Muse research is attempting to learn the behaviour of interactive game worlds from gameplay itself, while Google's Genie 3 can generate interactive environments that respond to a player's actions in real time. Both remain research technologies with significant limitations, particularly when it comes to maintaining consistency for very long periods, but the principle is extraordinary: the AI is learning not simply what a game looks like, but something about how an interactive world behaves.
Combine those ideas with conventional game engines and it is not difficult to imagine where things could eventually go.
A future RPG might know which characters are alive, which factions you have angered, what equipment you carry and which towns you have already visited. An AI system could potentially use that information to construct another quest while still obeying the rules, history and geography of the original game.
For players who become attached to a world, that sounds brilliant.
Then comes the uncomfortable question.
What if the software isn't a game?
A modern video game is one of the most complicated pieces of consumer software most of us will ever interact with. Huge games contain millions of lines of code, physics systems, artificial intelligence, networking, databases, scripting languages, graphics engines and decades of accumulated development knowledge.
If AI becomes increasingly capable of examining software like that and working out how its pieces fit together, there is no particular reason the underlying ability should stop at video games.
The same basic capability could be enormously useful elsewhere. Businesses still rely on old software whose original developers retired years ago. Government departments, manufacturers and financial organisations operate systems built over decades, sometimes containing code that few current employees fully understand. An AI capable of analysing those systems could help explain old software, migrate it onto newer platforms, identify failures and preserve technology that would otherwise become increasingly difficult to maintain.
That could be transformative.
It could also create obvious risks.
A system capable of rapidly understanding how complicated software works may also become better at discovering where that software behaves unexpectedly or where weaknesses exist. The exact same ability that helps a developer repair twenty-year-old code could potentially help somebody understand proprietary systems they were never supposed to understand.
The capability itself is neither particularly heroic nor sinister. What matters is who is using it and what they are asking it to do.
Games may simply be where we notice it first
Perhaps this is why these gaming experiments are worth paying attention to even if you have absolutely no interest in turning Skyrim into Minecraft.
Games provide a comparatively safe and entertaining environment in which these capabilities can become visible. When an AI helps somebody make Steve wander through Whiterun, the result is funny. When it reconstructs the mechanics of a thirty-six-year-old game overnight, it is impressive.
Those demonstrations also make something abstract suddenly understandable.
We spend a lot of time discussing whether AI can write an email, generate an image or answer an exam question. Understanding a complicated piece of software is a different level of capability because software increasingly controls the world around us.
Our cars contain it. Businesses depend upon it. Factories run on it. Hospitals, councils and financial institutions all operate enormous collections of software that have accumulated over decades.
The idea that AI might eventually be able to examine those systems and understand them rapidly could be one of the technology's greatest strengths.
It is also exactly why it deserves our attention.
I still want the endless RPG. I would happily return to a favourite game months later and discover that the world had continued without me, with new characters, new missions and new stories waiting to be explored. I think that could represent one of the biggest changes gaming has seen since worlds first moved online.
But watching how quickly AI is learning to work with the systems underneath those games has changed the question slightly.
It is no longer simply, “Imagine what AI could do to a video game.”
It is becoming, “If it can do that to a video game, what else can it do?”
Tags:
