Example: overly agreeable
Sounds reassuring
Absolutely! AI summaries are a great idea. They will save your team time and make research much easier to use. I would add them to every study.
Interactive example / Human-AI interaction
Make the assistant's defaults visible.
AI assistants can sound helpful while agreeing too readily or hiding uncertainty. Adjust four response preferences and see how the same design question might be answered.
How this works The responses below are written examples, not replies from a live AI model. The controls create custom instructions you can copy. Instructions can influence an assistant's behavior, but cannot guarantee a result.
01 / Experiment
Tune the interaction
Same question in both examples
“Should we add AI summaries to our research repository?”
How readily should the assistant question your idea?
SupportiveChallenging
How visibly should it separate evidence from assumptions?
LightExplicit
How much help should it give you to move forward?
Answer onlyCoach me
How much detail should the response include?
BriefDetailed
With your preferences
Changes as you adjust
Illustrative wording only. A real model may respond differently.
02 / Take it with you
Your custom instructions
Paste these into an assistant's custom-instructions or preferences field. Edit the wording to suit your work.
03 / What this teaches
Control needs a clear mental model.
01
Make behavior legible
Plain-language controls let people express preferences without learning instruction syntax.
02
Show the consequence
A before-and-after example makes each setting easier to understand than a label alone.
03
State the limit
Preferences guide responses. They do not set a universal baseline or guarantee how every model will behave.
Inspired by Nadia Piet's LLM Counter-Defaults experiment . This is an independent Cognitive Engine teaching example.