The Two-Model Disagreement Signal
When different AI models reach meaningfully different conclusions on the same brief, the disagreement is diagnostic. Averaging the answers throws that information away.
I ran the same brief through multiple AI models. Same context, same constraints, same ask.
They did not disagree on the wording. They disagreed on what the actual problem was.
One thought the opening needed work. Another thought the structure was weak. A third questioned whether the premise was strong enough to carry the piece.
The obvious response is to combine the strongest parts of each answer. That can improve the draft, but it can also remove the most useful information the exercise produced.
When different models arrive at meaningfully different conclusions, I have started treating the disagreement as a prompt to inspect the brief more closely.
Different diagnoses, one average
It does not automatically mean the problem is complex. Sometimes one model has misunderstood the ask. Sometimes important context is missing. Sometimes the models are optimising for different things: clarity, completeness, originality, or safety. And sometimes the available information genuinely supports more than one direction.
Those are different diagnoses. Averaging the answers makes them look like one.
The useful questions are:
- What did each model assume?
- What did it treat as the main constraint?
- Which part of the brief allowed more than one valid reading?
If the disagreement disappears once the brief is clarified, the problem was ambiguity.
If it remains, you may have found a judgment boundary: the point where the models can explain the available options but cannot decide which trade-off is right for your audience, your business, or this particular moment.
I think of this as the Two-Model Disagreement Signal. The disagreement is not the answer. It tells you where the answer still needs work.
Agreement is useful. Disagreement is diagnostic.