A Sous-Chef, Not a Co-GM: Where AI Belongs in Your 2026 Prep Kit
Photo: Master Sgt. Stan Parker, Public domain, via Wikimedia Commons
If you GM regularly, you probably already use some kind of AI tool somewhere in your prep. Maybe it names taverns. Maybe it turns your scrawled notes into something legible. Maybe it just listens while you talk through an NPC’s motivation at 11 p.m. and you pretend that’s not what’s happening.
That’s fine. What’s changed by 2026 isn’t whether AI shows up in the hobby — it’s that the tools got good enough at prose that the line between “helpful” and “hollow” has gotten blurrier. So here’s a working boundary map, organized by how much risk you’re taking on.
The Green Zone: Work Nobody At the Table Will Ever See
Green-zone tasks are the ones where a mistake costs you thirty seconds and zero trust. Hand them off freely.
Name lists. Generate forty NPC names in a consistent phonetic style before the session, then pick from the list live. This is the single highest-value use of AI at a fantasy table, because the alternative is “you meet a guy named… Gary… again.”
Reskins and flavor passes. You have a stat block. You need it to be a clockwork hound instead of a wolf. That’s a text transformation, and it’s exactly what these tools are good at — provided you read the final numbers yourself.
Recap drafts. Paste in your own session notes and ask for a 200-word recap to open with. You’ll still edit it. That’s the point. Editing is faster than composing, and it stops you from opening a session with “so, last time… what did we do?”
Random tables. Weather, roadside encounters, what’s in the chest, gossip overheard in a market. Ten options, pick the one that fits the table’s mood.
Prop text. Handouts, letters, wanted posters, a merchant’s price list. Check the numbers against your own economy, then print.
Your own notes, reorganized. Feed it a messy page of prep and ask for a clean three-column layout. This is filing, not authoring.
The Yellow Zone: Useful, But Read Everything Twice
These are the tasks where AI gives you a genuinely helpful draft and a genuinely plausible error in the same breath. Use them with your eyes open.
Rules lookup. Models are better at this than they were, and still confidently wrong often enough that you should never quote one at the table. Use it to find where a rule lives, then open the book.
Encounter balancing math. It’ll crunch numbers. It doesn’t know that your party has a paladin with absurd AC and a wizard who only prepares utility spells. You know that. The math is a starting point, not a verdict.
Continuity checks. If you keep a canon document, pasting in a chapter and asking “who is the Baron’s sister again, and did I contradict myself?” is genuinely useful. Keep the canon doc yours.
NPC voice practice. Talk to a simulated version of your villain before the session to find their rhythm. Weird? Slightly. Effective? Very.
Anything with a license attached. Do not paste large chunks of published material into a chatbot and then redistribute the output. Respect the SRD, respect the publisher, respect the creators. This is the least glamorous rule and the one most likely to bite you.
The Red Zone: Hands Off
Some things at the table belong to people, not tools.
Safety calls. Lines and veils, the X-card, the moment someone says “can we not.” These require a human who can read a room, improvise, apologize, and change the plan mid-sentence. Never delegate this. Not to software, not to a flowchart, not to a policy you wrote once and never revisited.
Live rulings. When a player asks “can I do this?” at the table, the answer comes from you, fast, with a ruling you’ll honor. Stopping the action to consult a chatbot is the single fastest way to drain tension out of a scene. If you’re unsure, make a call, note it, look it up after.
Player trust. Don’t scrape your friends’ messages or private writing into a tool without asking. Don’t generate personalized content that reads like surveillance. The table runs on the assumption that everyone is safe being a little weird in front of each other, and that assumption is fragile.
Someone’s character. Backstories, names, and voices belong to the player who made them. Offer help if asked. Never generate a person’s PC for them and present it as a gift.
The dice. Fudging is a human decision with human consequences. Outsourcing it — to a randomizer, a “difficulty engine,” or a tool that decides when to soften a roll — puts a machine between you and the trust your players handed you.
A Twenty-Minute Prep Loop
Here’s what the green and yellow zones look like in practice, on a real Tuesday night:
- Dump your session notes in and ask for a cleaned-up outline. Edit for ten minutes.
- Generate twelve names and five rumors. Keep four names, two rumors.
- Reskin one stat block you already planned to use.
- Draft the opening recap from last week’s notes. Cut it in half.
- Close the laptop. Sit with your NPCs for five minutes and figure out what they want. No tool required.
That last step isn’t a joke. The part of prep that actually shows up at the table is the part where you understand your people — the villain’s petty grudge, the innkeeper’s fear, the guard’s favorite niece. No model has ever run a good session. It just clears the desk so a person can.
Tell Your Table
One last thing, and it’s not a small one: say it out loud. “I use AI for name lists and recap drafts.” Most tables genuinely don’t care. Some players do, and they deserve to know before session one rather than finding out from a typo in a handout.
Disclosure is cheap. It buys you the thing you actually need at the table, which is attention — yours, pointed at the players, with nothing humming in the background.