Hi!

Last week, Google released the new Gemini 3. The model can read text, video, audio, and images together.

It can follow longer instructions and has deeper reasoning capabilities.

It can handle bigger files without losing track. And break down complex tasks into clear steps.

Can be integrated with Google tools — Gmail, Google Docs, Sheets — for workflow automation.

For us in medicine, Gemini can solve complex problems that have multiple layers. Help review complex information and give clearer explanations.

Have you tried it Gemini 3?

Now, let’s dive into today’s issue.

🤖 AIBytes

Are Ambient AI Scribes Reducing Clinician Burnout?

Researchers tested an ambient AI scribe in a real health system to see how it changed documentation time and burnout for outpatient clinicians.

🔬 Methods

Design: Randomized trial in real-world clinics.

Participants: 66 outpatient clinicians (physicians and advanced practice providers).

Technology: Abridge ambient AI, integrated with Epic, used to record visits and draft notes.

The system turned clinician–patient conversations into draft notes that clinicians review and sign.

Duration: 24-week rollout in three 6-week waves (August 2024 to March 2025).

Assessments:

Work exhaustion/interpersonal disengagement and professional fulfillment scales.
Time spent on documentation
Work outside work
Billing accuracy
Note quality

📊 Results

71,487 notes were written; 27,092 (38%) were generated with ambient AI.
Work exhaustion/interpersonal disengagement improved significantly.
Professional fulfillment increased.
Time spent on notes dropped by 0.36 hours per day.
Work outside work decreased by 0.50 hours per day.
Diagnostic billing codes improved significantly with ambient AI.
Note quality stayed high.

🔑 Key Takeaways

Ambient AI scribes reduced documentation time and work exhaustion for outpatient clinicians.
Note quality stayed high.
Small improvements in professional fulfillment were seen.
Even with AI-generated notes, clinicians kept control of documentation content.

🔗 Afshar Y, Larochelle MR, Montague J, et al. NEJM AI. 2025;2(12). doi.org/10.1056/AIoa2500945

AI Exercise Coaches: Helpful Tool or Not Ready yet?

This scoping review mapped how researchers are testing LLM-based exercise and health coaches, and how rigorous those evaluations are.

🔬 Methods

Design: Scoping review.

Inclusion: Original studies that evaluated LLM-based exercise or health coaching systems.

Model types: Systems that are either Large Language Models or show LLM-like abilities.

Quality scoring: Custom 5-point Evaluation Rigor Score (ERS) to rate methodological depth.

📊 Results

20 studies met the criteria for inclusion.
75% used LLMs (such as ChatGPT).
80% were rated by humans, either experts or users. 40% used automated metrics.
Median Evaluation Rigor Score 2.5/5, with 55% of studies rated low rigor.
Only 40% used real-world data from actual users.

🔑 Key Takeaways

LLM exercise and health coaches are being studied, but most evaluations have low methodological rigor.
Most studies use small, short-term tests that do not mirror real clinical practice.
Safety, long-term adherence, and clinical outcomes were not always measured.
Clinicians should treat current LLM exercise coaches as experimental tools, not ready-made clinical interventions.

🔗 Lai L, et al. J Med Internet Res. 2025;27:e79217. doi.org/10.2196/79217

🦾TechTool

Notion AI — Organize all your projects in one space (Link)

Built into Notion, this AI tool helps you organize all your documents in one workspace.
You can pull key points from your reading list, projects, and meeting notes.
Helps teams stay organized with projects and tasks in one shared workspace, so everyone stays aligned.

Eleven Labs — Text to audio (Link)

Turns written content into natural, clear human-sounding audio.
Creates consistent voices in multiple languages, including English and Spanish.
Great for generating clean voiceovers for video, presentations, or tutorials.

Scholarcy — Research helper (Link)

An AI-powered research platform to analyze and organize your research.
You can summarize videos, papers, and documents and turn them into flashcards.
Can be integrated with other apps like Excel and Notion.

Retrieval-Augmented Generation (RAG)

A method that helps an LLM use real information before answering by pulling documents, papers, or data that match the question.

When to use it:

When you want answers grounded in evidence.
When accuracy matters (research summaries, guidelines, documentation).
When you need the model to consider external sources before generating a response.

Prompt example

I’m attaching this patient’s past 3 clinic notes and recent shoulder MRI report.
Use Retrieval-Augmented Generation to review the documents first.
Question: Based on the retrieved information, what is the differential diagnosis?

RAG improves accuracy, but human oversight remains essential.

🧬AIMedily Snaps

AI Prescribing without human oversight? The Healthy Technology Act of 2025 (a bill being considered by Congress), might allow AI systems to prescribe medications (Link).
A short Video where experts talk about where AI is headed (Link).
Leaked executive order draft lays out Trump's plan to squash state AI policy (Link).
Will Clinical AI become a medical specialty? (Link).