On January 12, 2021, Adobe started blocking Flash content from running in its own player, less than two weeks after official support ended. A decade of browser games turned into files most computers refused to open.
Some of them made it out. Scroll through today’s catalog of online games by Plarium, and you’ll still find Stormfall: Age of War and Total Domination, two strategy MMOs that started as Flash browser games in the early 2010s. They survived for a boring reason: the company kept the original code and shipped a player to run it.
That boring reason matters more now that AI has joined the preservation debate. Google researchers have shown a neural network can imitate DOOM well enough that viewers barely tell the difference on short clips. The catch is what the network can’t do. It can redraw a game’s screen. It can’t recover the state underneath: the score, the inventory, the castle someone spent two years building on a server that’s now switched off.
What Actually Broke When Flash Shut Off
The plugin was only the first thing to go. Plenty of Flash games leaned on machines their players never saw.
Leaderboards sat on servers. Sponsor code phoned home to portals. Ruffle’s developers point out that any game using the Armor Games API depended on a protocol called Flash Remoting. Multiplayer worlds like Habbo, Club Penguin and Gaia talked to their backends over network sockets. Pull those servers and a perfectly archived SWF wakes up in a house with half its furniture missing.
Archivists repeat one point for a reason: a game is software, so saving it means saving something that can still run it. A folder of SWF files with no player is sheet music in an empty concert hall.
How Plarium Kept Its Flash Worlds Running
Plarium’s fix arrived before the deadline. In a December 2020 FAQ on its forums, the company explained that its desktop launcher, Plarium Play, carried its own embedded Flash plugin and used it only to run Plarium’s games. Players still worried the launcher would die on January 12. A community manager told them it would keep working.
Nothing about that approach is clever, and that’s exactly why it held up. The original game code kept running. Player progress lived where it always had, on Plarium’s servers. No model had to guess anything.
Ruffle Rebuilds the Player, One Function at a Time
For games with no publisher left to care, volunteers built a new player. Ruffle is an open-source Flash emulator written in Rust. It runs in modern browsers through WebAssembly, minus the security holes that sank the original plugin.
The work is slow and very literal. Looking back on 2023, the Ruffle team tallied the progress: 852 merged pull requests from 43 people, ActionScript 3 language support up from 60% to 75%, and API support up from 60% to 68%. The same post covered working TCP and XML sockets, something the team wasn’t sure browsers would ever allow. Multiplayer Flash worlds can’t function without them.
Each filter effect, codec quirk and half-documented function gets rebuilt by hand, then tested against real games that used it in odd ways. That’s what makes emulation trustworthy. It replays the rules the developer actually wrote.
Can a Neural Network Run a Game It Never Executed?
Google researchers took the opposite route with GameNGen, presented at ICLR 2025. They skipped DOOM’s engine entirely. Instead, they trained a diffusion model on recordings of an AI agent playing the 1993 shooter and asked it to predict each next frame from the player’s inputs. It runs at about 20 frames per second on a single TPU.
The headline result is striking. In the paper’s human test, raters picked the real game over the simulation only 58% of the time on 1.6-second clips and 60% on 3.2-second clips. That’s barely better than a coin flip.
Read the limitations and the picture changes. The model sees a little over three seconds of history. Much of what it seems to remember survives only because health and ammo sit on screen. The authors admit they could often spot the real game after a few seconds of actual play. They also describe a telling error: fire repeatedly, and GameNGen may spawn an enemy, since in its training data shooting usually meant something was there.
The save-file problem
Now imagine the same trick aimed at a lost Flash MMO. The scenario is made up, but the mechanics come straight from the paper. Feed a model thousands of hours of Stormfall footage and it might render a convincing keep, menus and battle screens. Ask it how many troops your alliance had on a given server in 2016 and it has nothing to work from. That number lived in a database, never fully on screen. The model would produce one anyway, because plausible output is its whole job.
Here’s the dangerous part: nothing crashes. The screen looks right, the numbers look right, and they’re invented. Engineers already know this pattern from silent AI agent failures, where a confident system drifts without throwing an error. In preservation, a silent failure means future players inherit a counterfeit and never find out.
The Atari story makes the contrast sharp. When ChatGPT lost to Atari at chess, the 1979 Video Chess cartridge was running inside the Stella emulator. The chatbot kept losing track of its own pieces. The emulated program never did, because the board was the one thing it stored exactly.
Frame generation on modern graphics cards offers a useful middle case. A neural network invents images between rendered frames, yet a real game keeps running underneath and corrects it every time it draws a true frame. A neural simulation of a dead game has no such anchor.
| Approach | Runs original code? | Keeps server-side state? | Main risk |
| Publisher launcher (Plarium Play) | Yes | Yes, while servers stay up | Depends on one company’s commitment |
| Emulation (Ruffle) | Yes | Only if a backend survives or gets rebuilt | Slow, function-by-function work |
| AI frame generation | Yes | Yes | Only fills gaps between real frames |
| Neural simulation (GameNGen) | No | No | Plausible guesses passed off as the game |
Where AI Actually Earns Its Keep
AI still has real work to do in preservation. It just belongs in a supporting role.
Sorting may save the most hours. Fan archives hold thousands of loose SWF files with mangled names, duplicate uploads and missing credits, because the portals hosting them vanished. A model that matches title screens, reads embedded metadata and flags likely duplicates hands an archivist a shortlist instead of a haystack.
Asset cleanup comes with a rule attached. Upscaling sprites or denoising audio for a remaster works fine when the original file stays archived beside the improved one, so nobody mistakes the guess for the source.
Server reconstruction is the big prize. Many browser games lost their backend, and a backend is mostly a contract between client and server: this request goes in, that response comes out. Captures of old network traffic, plus whatever documentation survived, can give developers enough to build a stand-in server that emulated clients will talk to again. AI speeds that up by grouping message types and drafting handlers for humans to check against the recorded responses. The game logic stays authentic. Only the missing plumbing gets rebuilt.
The Next Flash Is Already on the Calendar
Today’s games face their own version of 2021. Always-online requirements, delisted storefronts and quiet server shutdowns can strand a game without any plugin being retired. Each case repeats the Flash pattern: the client survives, and the world it talked to disappears.
The lesson for studios is fairly plain. Keep the source code and the server code, plus captures of what the two said to each other. With those in a vault, AI becomes a strong assistant for whoever reopens it in 2045. With only gameplay videos, it becomes a very good forger.
FAQs
Q.1 Can you still play Flash games after 2021?
Yes. Emulators like Ruffle run many Flash games in modern browsers, and some publishers ship launchers with their own built-in player. Compatibility still varies, especially for ActionScript 3 titles.
Q.2 Can AI recreate a lost video game?
It can imitate how a game looked from recordings, as GameNGen did with DOOM. It can’t recover server data, hidden rules or areas missing from its training footage.
So who’s keeping the server code for the games we’re playing right now?
Related: Can AI Tell What Your Retro Game Collection Is Worth?
