Signing in

You will be sent to MillerKnoll sign-in.

AI

Spot Fusion Failures Before You Present

Fusion is when Copilot blends material from multiple inputs into one answer. Useful when sources agree; dangerous when it invents a bridge between conflicting facts.

Lesson 3

Smooth output can hide mixed-up sources.

Fusion failures look fluent: a executive summary that merges two client names, a slide outline that attributes a quote to the wrong meeting, an image caption that describes furniture you never specified.

Defense: separate extraction prompts per source, then a synthesis prompt that must cite which source supported each claim.

Core principles

  1. Symptom: output sounds unified but you cannot trace claims back to a source.
  2. Fix: extract facts per document first; synthesize second with explicit source tags.
  3. Images and slides: verify labels, product names, and quantities against authoritative lists.
  4. Failure test: remove one source and see if the output still claims something only that source contained.
  5. High-stakes decks: Tier 2 or Tier 3 verification on every named fact.

Go deeper: AI Judgment — verification tiers

Check yourself

What is a fusion failure in everyday Copilot work?

Do this in Copilot

Run a two-source synthesis on real work. Mark one claim you cannot trace to a source.

Paste this into Copilot Chat and work through it before moving on.

Extract then synthesize

Step 1 only: From [SOURCE A] and [SOURCE B] separately, list key facts as bullets tagged A or B. Do not merge yet. Wait for my go-ahead before Step 2 (synthesis).
Open Copilot →
  • Chain of thought
  • Source material
  • Step 2 (after Step 1): Synthesize a [FORMAT] for [AUDIENCE]. Every bullet must keep its A or B tag. Flag any claim that required guessing.

Did you run this in Copilot? Mark complete when you have tried it.

Next lesson: Use One Source to Create Another Format →