Everyone's selling AI that agrees with you. I built one to argue with me inst...
Everyone's selling AI that agrees with you. I built one to argue with me instead.
I run several. One's an editor trained on my own writing, built to catch me when I hedge or dress up a weak point (not to make me sound smart). Another takes a recording, a meeting, a webinar, a sales call, and runs it through a second pass whose only job is to find what doesn't hold up.
I ran a recorded business-scaling webinar through it. Three things it caught that a straight summary swallows whole:
An unsourced stat doing the heavy lifting. Stated as fact: "94 to 96% of companies stay stuck or go out of business, only 4 to 6% break through." A scary round number with no citation, quietly bundling three different outcomes (stay stuck, stall out, go under) into one figure to manufacture urgency. A summary repeats it as authority. The adversarial pass catches the conflation and the missing source.
Winner stories sold as a repeatable method. "One client hit 10x in 12 months. Another, 8x in 27 months and nine offers." The missing number is the denominator. Those are the winners, and the method's actual success rate is the one figure never shown. The presenter even hedges mid-sentence ("you can't expect that every time"). Then sells it as repeatable anyway.
The most rigorous-sounding slide is the biggest leap. "A $4M business worth $1.6M becomes worth $16M just by converting to recurring revenue, no additional growth required." The whole 10x rests on the market granting a software multiple to what was a services business. Asserted, never defended. Relabeling the offer doesn't automatically earn the multiple.
A single-pass summary repeats what was said. The adversarial pass finds what wasn't supported. Same input, second lens. That's the whole point.
Mike Masnick at Techdirt just described the same setup: a tool trained on his voice to challenge his facts, and the one thing he won't let it do is write. The features he values most make the work harder, not faster (that's the tell).
Same instinct. The flattering AI feels great and teaches you nothing. Friction does a job a yes-man can't.
If you're evaluating AI for your team, ask one question: does this make me feel good, or does it make the work better? Not the same thing (and the vendors selling you the first one are counting on you not noticing).
Build the one that pushes back. Keep the thinking yours.