Every gamer harbours a personal ideal. Perhaps you dream of a title where your favourite voice actor takes the lead, or a game that perfectly aligns with your preferred genre and vibe. Maybe you just want the gacha gods to bless you with a pick-up character in a free ten-pull. Personally, I have a soft spot for apocalyptic settings fused with deck-building, roguelike mechanics, and loot-driven RPG elements. After searching for years and finding nothing that matched my specific itch, I eventually found myself turning to tabletop role-playing games.

When I used Gemini to master a TRPG session, I found a level of unpredictability and fresh excitement that far surpassed my solo note-taking and dice-rolling attempts. However, as is often the case with these tools, there were some fundamental limitations.

The issues with Gemini were twofold: it tended to prioritize player wish fulfilment over the established rules, and it had a habit of suddenly breaking character to trigger extension scripts. To see if the grass was greener elsewhere, I decided to put the heavyweight champion, ChatGPT, through its paces.

Having already identified the pitfalls of generative AI, I arrived at this test with a toolbox full of safeguards. I kept the rules consistent with my previous attempt, but I bumped up the zombies' health to inject some actual tension into the combat. I also added a script allowing players to freely 'throw' held weapons to distract foes or deal damage.

Furthermore, to prevent the AI from generating 'errors'—like the time it tried to trigger a script involving public institutions, which I had to shut down for legal reasons—I explicitly framed the entire session as a 'fictional game'. By repeatedly drilling into the AI that this was a 'virtual world' for the purpose of 'gameplay', I managed to keep the scripts focused on the experience itself.

When I first fed it the rules, ChatGPT used Python-based 'vibe coding'—creating code based on design intent—to ensure it stayed within the bounds of the core ruleset. It wasn't quite up to the level of GitHub Copilot or Cursor, but given the simplicity of the rules, it worked a treat. Unlike Gemini, it ignored my subjective desires and stuck strictly to the dice rolls to determine outcomes.

I also moved away from simple, top-heavy text blocks. By adding emojis before each response, I improved readability significantly. This was a lifesaver during intense combat or complex scene descriptions where wall-to-wall text usually leads to confusion. With an emoji representing each item, backtracking through the script to track my inventory became a doddle.

However, it wasn't all plain sailing. While it handled the core RPG rules well and used its coding prowess to manage essential items, 'hallucinations' remained an insurmountable wall.

For instance, when I asked to search an area for context-appropriate supplies or weapons, it would occasionally throw a curveball. Finding a steel pipe in a ruined restaurant or discovering a manhole cover and a utility pole inside a subway station were prime examples of the AI losing the plot.

Then there was the issue of the steel pipe not disappearing from my inventory after being thrown. While repeated script tweaks reduced the frequency of these errors, they would inevitably creep back in after long sessions or when starting with a new character. Dealing with these persistent counting errors was, frankly, a bit of a headache.

Comparing these two AI powerhouses reveals distinct pros and cons. If you want a GM that builds rich, sprawling content from a basic premise, Gemini is your companion. But if you’re a player who prefers a streamlined experience governed by strict rules, ChatGPT is the superior choice.

ChatGPT is highly recommended for strategic players because of its flexibility with rule additions. It can dynamically adjust success probabilities based on environmental conditions, forcing you to make the most of your character’s stats and inventory. For example, it once factored in rainy weather, making it harder for enemies to spot me, but causing my character to gain fatigue faster.

Conversely, if you're in it for the narrative, I strongly recommend Gemini. It’s adept at weaving complex subplots—like discovering the gear of a fallen character during a new run, or describing mysterious laboratories and destroyed platoons using nothing but atmosphere and subtle clues. It excels at world-building that you might not even think to ask for.