Presenter outlines benefits and risks of AI in health care

A Presenter outlined how artificial intelligence is reshaping health care and research while warning of significant limits and hazards.

Speaking in a slide-driven talk, the Presenter said AI "simulates human intelligence" but behaves differently from human learning, using a video of a child learning what a 'car' is to contrast human generalization with the data-heavy training requirements of machine-learning models. She described the model internals as input, hidden and output layers and noted that scale—she referenced "96 layers" when discussing early large language models—makes full transparency difficult.

The talk emphasized why clinicians and health systems are turning to AI: the volume of electronic health record text is large enough to overwhelm clinicians. Quoting from a book she recommended, the Presenter said the book notes that "one out of five patient records is longer than Moby Dick," and argued that ambient-listening and summarization tools are attractive because they can surface clinically useful information from that bulk.

At the same time, the Presenter raised privacy and reproducibility concerns. She described research showing magnetic-resonance imaging can be reidentified and cautioned that material researchers and clinicians treat as deidentified may not remain so once advanced analytics are applied. "How many unintended consequences of what AI will be capable of?" she asked, urging attention to data protections.

The Presenter also warned about commercialization pressures. Citing the venture-capital model around digital tools, she said developers of open-evidence and clinical-AI products may seek revenue in ways that shift priorities away from long-term development and safety: "eventually they're going to want to make some money ... they're going to start squeezing everybody," she said.

On practical limitations, she recounted a library holiday-card experiment using Copilot in which the generator repeatedly failed to honor requested counts of animals and produced distorted images. "It's just really bad at numbers — it's really bad at math," she said, using the example to make a broader point about generative models' weaknesses for precise, safety‑critical tasks.

The presentation illustrated robotics and 'latent actions'—robots learning to move in real environments—by referencing recent robotic performances at a Chinese New Year gala, and noted both impressive progress and the shift from scripted behaviors to learned adaptability.

She closed by urging a balanced, human-centered approach: preserve clinicians 'in the loop' for oversight, watch for privacy and safety pitfalls, and pursue uses that support autonomy, mastery and purpose for health workers. She said slides and links would be shared after the talk.

The Presenter said she would send the talk materials and invited attendees to review linked resources for deeper reading.