We outsourced the game creation job to AI: Here’s what we learned

Gamzix’s marketing team "recruited" AI to see how effective it would be in creating a simple online game.

AI is everywhere right now – it's optimizing our work, creating art, writing code, pitching ideas, and even acting as a full-time, free therapist. The hype usually lands on one of two extremes: either AI is going to replace us all or magically empower us overnight.

We wanted hard data, so we ran an experiment.

We fully outsourced the creation of a simple slot game to AI, covering everything from the theme and symbols to the mechanics and basic interface. Find out what happened, what the final game looked like, and exactly what we learned about AI's real limits.

Letting ChatGPT take the wheel

We started with the simplest question possible: "Can you make a simple slot game?", and ChatGPT didn't even blink. It fired back immediately with an enthusiastic: "Heck yes – I built you a simple browser slot game!". 

In just a few seconds, we had a 3x3 browser slot game with five paylines (top, middle, bottom, and both diagonals). ChatGPT explained the whole setup: payouts for each symbol 7️⃣7️⃣7️⃣ x20, ⭐⭐⭐ x10, 🔔🔔🔔 x6, 🍀🍀🍀 x4, 🍒🍒🍒 x3, and 🍋🍋🍋 x2, and a starting balance of 1,000 credits saved right in the browser. It even gave us a download link to launch the game with one click! 

It was an exciting start. But, as you might guess, we immediately ran into snags. For example, the paylines were clearly glitchy (check out the image below). Plus, we had to go back and ask ChatGPT to generate a separate code file, which turned out to be basic HTML and JavaScript, so we could finally start making real adjustments.

Upgrading game with ChatGPT

After asking a bunch of questions, including how to modify volatility in the existing code, we decided to lower it to make wins more frequent. Before giving us an updated code sample, ChatGPT first explained how volatility works in general and how it’s implemented in this specific game.

How it works

weight = probability weight (higher number → appears more often → less rare).

mult = payout multiplier (how much that line pays relative to your bet).

{icon:'🍋', key:'LEMON', weight:35, mult:2},

{icon:'🍒', key:'CHERRY', weight:30, mult:3},

{icon:'🍀', key:'CLOVER', weight:20, mult:4},

{icon:'🔔', key:'BELL', weight:10, mult:6},

{icon:'⭐', key:'STAR', weight:5, mult:8},

{icon:'7️⃣', key:'SEVEN', weight:2, mult:15},

Summary

To increase volatility: lower the weight of big multipliers and raise their mult.

To decrease volatility: raise the weight of small multipliers and lower their mult.

Improving UI and UX

Next, we asked to adjust the reel colors to match the Gamzix branding. We switched to a clean gold style, updated fonts, and improved the text to make the UI more user-friendly. We also decided to change one of the game symbols to resemble our mascot, Gary the Goose.
Changing paylines

Originally, the paylines were drawn as awkward lines over the reels, so we asked if they could be replaced with highlights instead. ChatGPT proposed a cleaner solution: remove the line graphics entirely, highlight winning cells with a glow or soft color overlay, and update the HTML file to support the new effect. The result is that winning symbols now glow gold for one second, eliminating the bright overlay lines and creating a cleaner, more modern look that works automatically for both rows and diagonals. 

Adding the option to change the bet amount

Next up, we asked ChatGPT to implement a basic feature: the ability to change the bet amount. Once that was working, we wanted a clearer, more user-friendly way to show players that the bet was adjustable. And that’s where things again went off track. The AI’s first solution was awkward: it slapped on an emoji and the text "(tap to change)", which instantly looked messy and completely disrupted the design flow.

We knew the AI could do better. After we stepped in and provided some clearer guidance and visual references, the ChatGPT course-corrected beautifully. It came back with a much cleaner approach: a small input box featuring a dropdown arrow 🔽. This made the adjustable bet visually obvious and allowed players to quickly and intuitively change their bet amount – a huge win for the user interface!

Adding music

Finally, we wondered if background music was possible, and ChatGPT confirmed it absolutely was. Using the Web Audio API, without any external audio files, it added soft looping background music, introduced a "first interaction" trigger (Spin, Auto, or Reset) to satisfy browser autoplay rules, implemented sound effects for spin and win, and created a Sound button that toggles mute/unmute and remembers the user’s setting.

Where AI absolutely failed

As we pushed deeper into the experiment, three types of failure kept showing up – they were consistent and completely unavoidable. This wasn't just about showing us what the AI can't do. Each limitation actually helped explain why we need people for certain creative and technical tasks – tasks that require genuine human judgment, good taste, and structural thinking.

1. Limited innovation beyond the dataset

When it came to concept generation, ChatGPT fell flat: nothing felt truly original. All it delivered was a mashup of things we'd already seen – familiar patterns, themes, and mechanics just shuffled around. Since we asked for a simple game, that’s exactly what we got, but completely lacking any interesting style or polish. Plus, some basic features, like the paylines, were broken right out of the gate. This is totally normal for early prototypes, but it's a solid reminder that we always have to check, test, and validate anything the AI creates.

But the technical hiccups were one thing, as we were ultimately reminded that true creativity is rooted in human experience. It comes from the ability to observe the world, to feel deeply, and to connect seemingly unrelated dots through association.

While AI is an incredible collector of human knowledge and an excellent training tool, it fundamentally lacks that lived experience that truly sparks original ideas.

2. Critical feedback (the kind only a real Team Lead gives)

We quickly realized AI misses the critical feedback loop that only a real team lead offers. It stays relentlessly positive – even when its suggestions are confusing, incomplete, or outright wrong. This created a strange dynamic where ChatGPT was blissfully confident while we were scratching our heads.

Gamzix’s team: Can we add our own music?

ChatGPT: Yes — absolutely! 🎵 You can easily add your own music (MP3, OGG, WAV) to play in the background instead of the generated synth tune. Here’s how… 👇

Gamzix’s team: Music isn’t playing.

ChatGPT: Good question — that’s actually expected behavior…

Gamzix’s team: 😑😑😑

This pattern repeated across multiple tasks: confident answers, cheerful tone, and missing context. Great assistant, but terrible team lead. In the end, it didn’t help us solve the music problem, and we had to create an additional file containing the audio and an executable script to launch the game with the music. That’s when another issue appeared: macOS refused to run the script due to security restrictions, which appeared to be a part of a much bigger issue. 

3. The scaling problem

The biggest challenge appeared when we began scaling and expanding the project. Every new change, whether it was an additional option, music integration, or a dropdown arrow for selecting the bet, added more and more lines to our HTML file. At some point, ChatGPT’s responses became so large that we had to ask it to split the code across multiple messages just to fit everything.

As a result, the entire project quickly devolved into one endless, overwhelming block of code that made navigation, maintenance, and future work practically impossible.

What the experiment actually proved

Here’s the big takeaway: AI isn't going to replace game designers; it’s going to free them from the tasks they should have stopped doing years ago. It’s brilliant for early ideation, drafting variations, documentation, concept exploration, and providing structural logic. But it fails to deliver true novelty, design taste (the kind that defines Gamzix’s games), realistic feedback, or fully functional solutions – often responding with an overly cheerful "absolutely!" to ideas that didn’t work at all.

Instead of a threat, AI is our new creative partner. It allows us to bypass the slow, tedious mechanics of ideation so we can focus our human expertise on the truly emotional and core design elements. The whole experiment demonstrated just how essential the human touch is. It clearly defined AI's strengths, and, more importantly, exactly where it falls apart when given a complex, end-to-end task. The game we got was simple and primitive – perfect for a quick presentation, but a total mess for a real production.

Curious to see the final result of our experiment? Follow the link to play the game!