Is making a game with AI easy? What still takes work

Is making a game with AI easy?

It is not easy. It became possible for people who do not code, which is a different thing. AI cleared away the part that stopped most people, writing the code, and left in plain sight the parts that were always hard in any game.

Our take, based on the tests and real cases on this page.

If you do not yet know what working with AI is like, start with can you make a game without knowing how to code?. This page goes deeper into the hard part.

What did the 2026 test show about AI making games?

It showed that the best AI, working alone, solved just over half the tasks. The test is called GameDevBench, and it was published in the proceedings of ICML 2026, an academic machine learning conference¹.

The researchers took 333 tasks from game tutorials, in text and video, all in Godot, a free program for making games¹². The best combination, Google's Gemini 3 Pro with access to screenshots and videos of the game running, solved 53.8% of them².

The test also showed where AI stumbles most. The average success rate dropped from 51.4% on gameplay tasks (what the player does, and the rules) to 33.0% on 2D graphics tasks, such as art, animation, and visual effects. And the solutions required more than three times as many lines of code and files changed as the programming tests that came before it¹.

One finding helps explain the role of whoever tests the game: when the AI could see the screen and the game running, it improved. OpenAI's GPT-5.4 went from 41.1% to 52.0% with that visual feedback¹. In other words, AI makes fewer mistakes when someone, or something, shows it what came out (our take).

Two other 2026 tests, released as preprints on arXiv, a repository of scientific papers, point the same way:

These numbers are a snapshot of 2026. The AIs tested already have newer versions, and the results are likely to change. The tests measure AI alone, in one attempt, with nobody steering, which is exactly what does not happen in a real game.

What does AI do well in a game?

AI is good at building the foundation: the first version that runs, with the rules you described. GameXpert-Bench says AIs are more reliable at producing a playable base and carrying out explicit requests⁴. GameCraft-Bench says they usually get the recognizable parts of each type of game right³.

The cases of people who did not code confirm this. Japanese artist Kazuki Takakura says all the code of his game, a 2D action game, was written by AI, using OpenAI's Codex and Anthropic's Claude Code. He says he still does not understand programming⁸.

Jonathan Bjerk, an accountant from Norway, says a problem that took him a month to solve in June 2025 now takes him two days, thanks to what he learned and, above all, to newer tools⁷.

Where does AI still fail?

It fails where you need to look at the game running: finding the bug, checking that everything works, and not wrecking what was already done. GameXpert-Bench points to exactly these three weaknesses⁴.

GameCraft-Bench adds polish. AIs struggle to deliver games with enough content, with visual feedback that works (the player noticing they hit, missed, or won), and with a coherent presentation³.

The Claude Code documentation itself, for Anthropic's coding AI, warns about two behaviors⁵:

• It forgets in long conversations. When the conversation fills up, the AI may start forgetting earlier instructions or making more mistakes.
• It stops when the work looks done. Without a way to check, "looks done" is the only signal it has, and you become the one who checks everything. The common failure is work that looks right but does not handle the unusual cases.

In practice, this becomes a loop. The AI builds, you play, you point out what went wrong, and it redoes it. The loop works, but every round is your work, and this is where most of the time goes (our take).

What is still your job?

Four things stay with whoever makes the game: deciding, testing, publishing, and marketing. AI helps with all of them, but does none of them for you.

Deciding. Kazuki Takakura lists what, for him, are human decisions: whether the game is fun, the art direction, how tough the bosses are, the feel of the controls, and the satisfaction of clearing a level. According to him, when AI decides these things, the game tends to come out mediocre⁸.

Testing. Bjerk spent hours playing and adjusting values like the turning speed and the thickness of the trail left by the snake in his game. For him, AI cannot tell whether the game is fun; other people can⁷.

Publishing. Stores have their own rules, and AI does not skip them. On Google Play, for example, a personal account created after November 13, 2023, must run a closed test with at least 12 people for 14 days in a row before launching⁹. Bjerk says the store page, the promotional images, the translation, and the servers took a lot of time⁷.

Marketing. Bjerk says marketing was harder than making the game. He reached out to more than a thousand video creators on TikTok for a tournament of his game⁷. The competition is fierce: Steam, the largest store for computer games, received about 20,000 new games in 2025, according to the trackers cited by Chris Zukowski, author of the site How To Market A Game¹⁰.

How did it go for people who tried?

Three real cases show the same obstacles along the way.

The game that worked and turned out bad. In April 2026, the team of the podcast What Next: TBD, from the American magazine Slate, decided to test on the show whether AI coding tools work or are just hype. The host, Lizzie O'Leary, and producer Patrick Fort had never programmed; producer Evan Campbell had a little experience⁶.

With Claude Code, they made a game described as a version of the classic The Oregon Trail about making a podcast. The game worked. But Slate's vice president of technology did not want to put it on the magazine's website, over security concerns, and it was published on the producer's personal site. In the story's headline, they call the game bad themselves⁶.

The accountant who published. Jonathan Bjerk, a 38-year-old Norwegian accountant, knew only HTML and CSS, self-taught to build simple websites. He started Curve Clash, a multiplayer snake game, on May 28, 2025, using ChatGPT, then Cursor, and finally Claude. He launched a test version in August and the full version in mid-December, without quitting his job⁷.

The road was not smooth. Sync bugs in the online mode stalled the work in September and October 2025. According to him, more than 1,500 people played. Summing up his own role, he said AI was his engineering team, not his creative director⁷.

The artist who made his own game. Kazuki Takakura, a Japanese artist, did all the art and sound for a side-scrolling action game himself and left the code to AI. The game was finished on May 16, 2026, and runs in the browser⁸.

All three give similar advice. Bjerk recommends starting with a single-player game, cutting without mercy anything that is not essential, and getting people to test early⁷. Takakura recommends a small, one-level project, as simple as NES games, and avoiding rhythm and fighting games, which require precise timing⁸.

Do game professionals think AI made it easier?

They use it a lot and distrust it at the same time. The two numbers come from different surveys and should not be compared as if they were the same measure.

In the GDC survey, programmers are among the most skeptical: 59% see AI as negative¹². People who already knew how to do the work use the tool and, at the same time, see its risks. For someone who has never made a game, the same tool is what makes starting possible, but the work of deciding and checking that these professionals know so well is still there (our take).

So is it worth trying?

It is, if you go in knowing where the work is. AI removed the biggest barrier, which was programming, and left the rest: deciding, testing, publishing, and marketing (our take).

Three choices shrink the hard part, and all of them show up in the cases above: a small, single-player game with no online mode; real people playing from the first version; and an idea written down before you open any program. If you are still at the beginning, see how to take the first step before opening any program and what vibe coding is, the name this way of building software by talking to AI has come to be known by.

The other side

Making a game with AI is not easy, and this page promises no timeline and no result. The 2026 tests measure AI alone and have already been overtaken by newer versions, for better or worse. The real cases are few, their numbers are self-reported, and none shows there is money in it: the Slate game turned out bad, and Bjerk still works as an accountant. In the GDC survey, more than half of professionals see AI as negative, and some players are wary too. What changed is that not knowing how to code is no longer the end of the conversation.

Your action plan

Check off what you've done. It's saved on this device, so you can pick up where you left off.

1. Write down on a sheet of paper what the player does, how they win, and what stays out of the first version.
2. Pick a small, single-player game with no online mode as your first project.
3. Set aside hours each week to play and test, on top of the time spent asking the AI for things.
4. Line up two or three people who will play the first version and tell you what they thought.
5. Decide which store the game will launch in and read what it requires before you start.
6. Save a copy of the game every time a version is working.