LANDMARK · 7 MIN · JUDGE
Conflicting sources
After this landmark, you can detect when an AI answer has flattened a genuine disagreement, and recover the conflict it hid.
Ask a model a contested question and you will usually get a single, calm, well-organised answer. That fluency is a design property, not a finding. When the underlying material disagrees — two studies with opposite results, a rule that changed last year, a figure reported differently by two agencies — the model still has to produce one continuous piece of text, so it tends to blend, pick, or average. Sometimes it picks the most common phrasing rather than the most current or most reliable one, because frequency in training data is not authority. The tell is an answer that feels settled on a question you know to be live: no hedging, no “estimates vary”, no date attached to a number that changes. Your job is to ask the question that forces the disagreement back into view, and then to judge it yourself rather than delegating the judgment.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
When something sounds too settled, ask directly: “do sources disagree on this, and what are the strongest versions of each side?” Then check the dates. A confident answer about a rule, price, or recommendation that has changed recently is the most common way to be quietly wrong.
MAKE A DECISION
You ask an assistant about a regulatory deadline. It gives one date, fluently and without qualification. You know the rule was amended recently. What is the right next move?
CARRY THISTake a question in your field where experts genuinely disagree. Ask an assistant, then note whether it presented the disagreement or resolved it for you without saying so.