Playing D&D with ChatGPT: A Field Report

We ran six sessions across GPT-4o and GPT-4.1 in custom GPTs with uploaded homebrew notes — no dedicated RPG platform. Verdict: excellent scene partner, unreliable referee. If your bar is "write like a novelist," ChatGPT clears it. If your bar is "run weekly 5e with fair mechanics," it fractures by session four.

What worked

Tavern dialogue, environmental description, branching mysteries when we kept scope small. Custom GPT instructions to "always ask for rolls" helped for two sessions, then the model drifted back to narrating outcomes. One-shots framed as "three-act story" completed satisfactorily in ~90 minutes of typing.

What broke

• Inventory: items appear and vanish; selling the same sword twice
• Initiative: turns merge; enemies act out of order without correction
• Spell slots: forgotten after long rests unless re-pasted
• Multi-NPC combat: stat blocks invented mid-fight
• Player agency: model sometimes narrates player emotions and decisions

Mitigation tactics

External spreadsheet for HP and loot. Paste character sheet at start of every session. Explicit "stop — I roll" when the model skips dice. These work but feel like fighting the tool — because you are doing the DM's bookkeeping while also playing.

Multiplayer

Not supported. Screen-sharing a ChatGPT thread is awkward; parallel threads diverge instantly. For couples or roommates, pass-the-keyboard solo is tolerable; for four friends online, use a platform with session rooms.

When ChatGPT is enough

Solo narrative experiments, writer's block, NPC practice, zero budget. When you want lexicon, PartyServer, maps, and brief persistence, migrate to dungeonmaster.website or comparable AI GM stacks — the time saved on bookkeeping pays back session one.

ChatGPT is a co-writer who occasionally grabs your dice and rolls whatever feels cinematic.

Fair comparison requires matching use case: prose solo vs rules-aware multiplayer are different products wearing similar marketing.