I grew up in a house of secrets.
It took decades to understand what they were: my father's many affairs, the circumstances of my adoption (yeah I know, there's a potential Netflix series in it).
But it did teach me something useful for a future strategist: people are spectacularly good at presenting confident lies when the truth feels too uncertain to admit.
Currently there is a heated debate about the merits, or lack of, with AI-generated consumer panels and the resultant insights.
But fascinating new research (Link in comments) from PyMC Labs and Colgate-Palmolive suggests we're looking at it entirely wrong.
LINK HERE = Research = https://arxiv.org/abs/2510.08338
Instead of forcing AI into the prison of 1-10 rating scales, the researchers let it think in natural language first. Explain its reasoning. Express uncertainty. Only then map that language to a distribution of possibilities.
The results? 90% correlation with actual human judgment. More realistic responses. And crucially, less prone to the positivity bias that plagues traditional research.
As a dyslexic, this resonates deeply. My brain has never trusted neat numerical gradients. When you ask me to rate something 1-10, I'm already translating complex, contradictory feelings into an arbitrary system that feels fundamentally dishonest.
The number is a lie we agree to tell because it's easier than expressing nuance.
The AI, when freed from that constraint, was more honest. More critical. More useful.
Think about that. The synthetic consumer, when allowed to be uncertain, gave better insight than humans trained to be agreeable in focus groups or rushing through an online survey.
This isn't a research shortcut. It's potentially an upgrade.
Faster, cheaper, AND more honest insight. The entire "synthetic data is a cheat" argument collapses when the synthetic version is more truthful than the original.
Death of positivity bias. Imagine pre-testing creative and getting feedback that isn't softened by British politeness or American enthusiasm. That's not just different data—it's better data that drives better insights.
True qualitative depth at quantitative scale. We've always had to choose. This method suggests we might not have to anymore.
I love the irony. While the world panics about AI-driven misinformation, our industry has discovered that the same technology, used differently, might actually get us closer to truth.
After 20+ years in strategy, most of it learning from spectacular failures (Google "Hicklin Slade Sharon Bridgewater" if you fancy a laugh), I've learned this: comfort with uncertainty beats confident ignorance.
Every. Single. Time.
Question for you: What other “AI cheats" in our industry are actually just better methods we're too nervous to trust?