Sense and decide
Make the Model Argue Against
Point AI at the case against, so it stops manufacturing support for a decision already made.
Developed by Superadditive in our own review of the research. It draws on red teaming and the older practice of assigning a devil's advocate.
Ongoing habit · Minutes to start · Anyone on the team can start it
Try this first
Give it one instruction. Make the strongest case this is wrong.
Use it when
The AI is being used to review or check a proposal, and it keeps agreeing with it.
Skip it when
Use it on consequential proposals. And treat the counter-case as a prompt for judgment, not a verdict. A fluent argument against deserves no more trust than a fluent argument for.
How to introduce it
When you bring AI into a review, point it at the target. Ask it for the strongest case against the proposal, not a justification of the answer you already picked. A model can produce a convincing argument for either side, so tell it explicitly to make the strongest case against the proposal.
How to show up
Set the task as opposition, explicitly. Then hold the team to engaging the counter-case instead of waving it off. Watch for people re-aiming the model at what they already want.
How long it takes
No setup. It's a change in how the AI is prompted in review steps.
What makes it hard
The instinct is to ask AI to help build your case, which can make the existing case feel stronger than it is. Turning it around takes intent. The counter-case is as easy to dismiss as it was to generate, so hold the team to answering it.
What it looks like when it's working
AI gets used to attack proposals and turns up weaknesses the team then fixes. Mostly using it to bolster existing views means the team is not using the practice.
How long until it sticks
Expect a few uses before aiming AI at the counter-case becomes the default in reviews.
How you know it stuck
The team asks the model to argue against a proposal as part of testing it.
The idea behind it
Ask a model to support an idea and it can produce a persuasive case for it. Asking for the strongest counter-case gives the team a better chance to see what it is missing.
Where it comes from
our review. A fluent tool can produce persuasive support for the position it is asked to defend, even when the evidence is weak. When an AI is in a dissent or review step, task it to build the strongest case against the proposal and to disprove it, never to support the answer already on the table.
Evidence
No outside source. This move came out of our own review of the research, so treat it as reasoned judgment rather than tested practice.