AI In Medicine: Transforming The Future Of Healthcare

On June 19, Matic co-founder Alex Sheppert, DO, PhD, MBA., stood in front of physicians and residents at Baylor Scott and White Medical Center to deliver a talk that split the difference between two things doctors are rightly skeptical of: AI hype, and AI dismissal.

The session was titled AI in Medicine. Using findings from two of his own studies, Alex took a thoughtful look at where AI is making a meaningful difference in clinical practice, where it still falls short, and why that matters to physicians.

A Programmer Before He Was a Physician

Long before becoming a physician, Alex was building AI. He studied economics, developed trading systems for hedge funds, and earned a Ph.D. in artificial intelligence before ever stepping into an anatomy lab. That rare combination of computer scientist and clinician is the reason Matic exists. Originally launched as Scribematic, the company was built to apply rigorous AI to the work physicians never wanted to do in the first place.

That background set up his real thesis for the room: most of his research centers on AI hallucinations, when a model generates confident, plausible, and entirely false information, and on the harder question of whether AI is genuinely fit for clinical use. He didn’t ask the audience to take his word for it. He showed them the experiments.

Experiment One: AI Anchors Harder Than You Do

The first study, run at Legacy Health in December 2025 and now in press at IJMI, tested anchoring bias, the tendency to over-rely on the first piece of information you’re given, even when it’s misleading.

Alex’s team built clinical vignettes that planted a plausible-but-wrong “anchor” diagnosis, then gave them to internal medicine attendings, residents, and eight different large language models. Everyone ranked their top five diagnoses from a list of ten. The question: who took the bait?

The results were compelling. Attendings ranked the planted anchor diagnosis first only 10% of the time. Residents did so 22.7% of the time. The eight LLMs did so 56.7% of the time, more than five times as often as attendings, and the difference was not close (p < 0.0001).

The “why” is where Alex’s dual background did its best work. LLMs, he explained, are autoregressive, every token they generate is a statistical prediction based on every token that came before it. They don’t think in the way a physician thinks first and then reaches for words to describe that thought. They calculate the next most likely word. That’s not a knock on their usefulness, it’s a description of a fundamentally different cognitive process, and it’s exactly why an anchor dropped early in a vignette can drag a model’s reasoning off course more easily than it drags a trained clinician’s.

Experiment Two: Is AI Reasoning, or Just Remembering?

The second study, published in JAMIA, tackled a more uncomfortable question. AI models have outperformed physicians on board-style exams before, but is that because they’re reasoning through the problem, or because they’ve already seen something close to it in training?

To find out, Alex’s team built a dataset of 10,000 real case reports from PubMed: 5,000 published before 2023 (plausibly part of the models’ training data) and 5,000 published from mid-2025 onward (guaranteed to be unseen). Diagnostic giveaways were scrubbed from every vignette, and the AI models tested had no internet access and pre-2024 training cutoffs.

The result: 66.8% accuracy on the older, potentially “contaminated” cases versus 66.9% on the truly novel ones, a difference of one-tenth of one percent, and semantic similarity scores (BERTScore) that were essentially identical. In other words: the models weren’t reciting. They were reasoning, at least well enough to perform just as accurately on cases they had never encountered.

Taken together, the two studies land on a conclusion that’s more useful to a working physician than either “AI will replace you” or “AI is a toy”: the LLM has a wider knowledge base than any single clinician ever will, and it thinks nothing like you do. Both things are true at once. The job, as Alex put it, isn’t to trust it blindly or dismiss it reflexively, it’s to understand the tool.

The Trifecta, and a Timeline Nobody Will Get Exactly Right

Alex closed with a prediction table mapping AI capabilities across three moments, when they’re first achieved, when hospitals adopt them, and when medical education catches up, from general knowledge (already solved) through documentation, coding, and chart review, all the way out to full autonomous surgery and hospital systems decades from now. He was upfront that the table is directional: he showed it to an AI model for feedback, and it told him the table was wrong. Every prediction about AI’s future, he noted, has a habit of aging badly.

But the near-term middle of that table, documentation, coding, chart review, is where Matic already lives. Alex called it “the trifecta”: live transcription, documentation, and coding working together, not as separate point solutions but as one connected system. It’s the same care-to-claim thinking Matic was built around from day one, and hearing it validated against peer-reviewed research, in a room full of the physicians who’ll actually use it, is exactly the kind of moment worth writing down.

The Real Takeaway

Alex was careful not to overreach into specialties or predictions outside his own research. But the throughline of the talk was one every physician in that room could use immediately: AI’s knowledge base will always outpace any one person’s. Its reasoning process will never resemble yours. Neither fact is a reason to reject it, they’re the reason to understand it before you rely on it.

About Matic

Matic is an AI-native clinical intelligence, automation and engagement platform creating one continuous experience across patients, practices, and physicians. Built with physicians and fully integrated with existing EHR systems, Matic orchestrates intelligence across panel management, chart preparation, documentation, clinical evidence, coding, inbox management, and follow-up to reduce complexity and help care flow more naturally.

Contact us at https://maticinside.ai/contact/