The AI That Refuses to Guess

Scroll wellness social media at one in the morning, as Renee Deehan admits she sometimes does, and you’ll find an endless stream of confident advice. Most of it rests on a single study with sixteen participants. “That’s great that that study said MCT oil was going to lower your cholesterol,” she says, “because there’s six others that say it won’t have any effect. But you’re not talking about those.”

Deehan is a molecular and cell biologist who spent six years “squeezing eggs out of frogs” before moving into computational biology and, eventually, to running science and AI at InsideTracker. The company turns your biomarkers, fitness-tracker data, and DNA into personalized, non-prescription lifestyle recommendations. What makes it worth an episode is how deliberately it refuses to guess.

An engine that can’t hallucinate

The core of InsideTracker’s recommendation engine is not a large language model. It’s knowledge representation and reasoning, a branch of symbolic AI that stores knowledge explicitly and reasons over it with deterministic rules.

That design choice buys something a language model can’t offer. “It literally can’t hallucinate,” Deehan says. “It’s not probabilistic, it’s absolutely deterministic. If you put in the same query 1,000 times, you will get the same answer 1,000 times.” And it’s auditable end to end: if you ask why the system recommended red-light therapy, she can hand you the studies that supported it and the exact chain of logic that fired.

When Leon pushed on the technical layer underneath (an old habit of his), Deehan was candid: it’s mostly a relational database and a set of deterministic inference rules, not an ontology or the full semantic-web stack. The point was never the formalism. It was that you can inspect the reasoning.

Evidence convergence, not cherry-picking

The bar for making a recommendation is high. An intervention has to be shown effective in humans across at least three independent clinical studies, and it has to move a biomarker that’s actually out of range for that specific person. “There has to be convergent evidence,” Deehan says, “not just one paper.”

The MCT-oil case shows the engine deciding not to recommend. The team scrapes every relevant study, then weighs the balance. If ten studies split into three positive, two negative, and five showing no effect, that’s a borderline call, and a human committee makes it. The system is comfortable representing “not enough signal yet.” That, in a sentence, is the opposite of the influencer model.

Put the AI on the toil, the human on the call

Where do the language models live, then? On the unglamorous 85%. Deehan’s team mapped how a recommendation actually gets built and found that 85 to 90% of the effort is compiling research papers, reviewing them, and doing a preliminary “mini meta-analysis” of the evidence. Only a sliver is the writing itself.

So that’s where they aimed the automation, with NLP and, more recently, Claude Code, cutting a process that took an hour down to about fifteen minutes. The recommendation writing, the actual judgment, still belongs to humans, with a primary reviewer and often a backup. Deehan draws the line plainly: if you want AI generating your health advice with no human in the loop, as Leon put it on the show, “go use Grok or OpenAI Health.” It doesn’t scale as fast. That’s the trade they’ve chosen, because traceability is the product.

The discipline of not overclaiming

InsideTracker has the data to brag, and mostly doesn’t. Their PLOS Digital Health study followed 20,000 users with at least two blood draws. Cohorts that started out of range improved at a population level, and a subgroup with diabetic-level HbA1c dropped into pre-diabetic range by the fifth draw. When Steve asked whether the work was causal, Deehan didn’t blink: “I will never say that we established causality.” It’s retrospective. Without clean adherence data, the team infers behavior instead: the users who improved their LDL over a year were taking, on average, about 2,000 more steps a day than those who didn’t.

Two halves of the brain

Leon’s read is that InsideTracker is an early, practical sketch of neuro-symbolic AI: a deterministic reasoner and a language model working as complementary cognitive modules, each doing what it’s good at. Deehan’s own elevator advice for any executive trying to build trustworthy health AI points the same direction. “Figure out a way to use a really robust knowledge base,” she says, one built not just from published science but from your institution’s own data, decisions, and mistakes. The hard part isn’t the model. It’s the knowledge, and the humans who vouch for it.

Listen to the full conversation.

You Might Also Enjoy

• S1E27 with Charlie Harp — A deep look at healthcare data quality and the PIQI framework, the backbone any trustworthy reasoning system depends on.
• S1E43 with Zak Kohane — On who controls healthcare data and how to measure the values baked into clinical AI, a natural companion to Deehan’s auditability-first design.
• S1E40 with Jeffery Smith — Transparency and AI regulation from the government’s side of the table, the policy counterpart to “show me your evidence.”