When Patients Bring AI to the Exam Room

Introduction: The New Reality

Your patient walks in with a printout from ChatGPT suggesting she might have lupus. Another asks why you didn't order the MRI that "the AI" recommended. A parent wants to know if the treatment plan you've outlined matches what Google's Gemini told them about their child's ear infection. These scenarios are no longer hypothetical—they're becoming routine.

According to a 2024 KFF poll, approximately 17% of U.S. adults now use AI chatbots at least monthly for health information and advice, with that figure climbing to 25% among adults under 30. A cross-sectional study published in the Journal of Medical Internet Research found that 21.5% of respondents had used ChatGPT specifically for online health information. The Elsevier Clinician of the Future 2025 survey reveals a striking mismatch: while patients increasingly turn to AI tools, only 16% of clinicians report using AI in direct decision-making, and just 30% feel they've received adequate training.

This topic examines the emerging landscape of patient AI use, how it manifests in clinical encounters, and—most importantly—how clinicians can transform what might feel like a challenge into an opportunity for deeper engagement and better outcomes.

Part 1: Understanding How Patients Use AI

Who Is Using AI for Health Information?

The demographic profile of health AI users challenges some assumptions. Research from ResearchMatch reveals that ChatGPT users for health information tend to be younger (average age 32.8 versus 39.1 for non-users), but—perhaps surprisingly—have lower rates of advanced degree attainment (49.9% versus 67% holding a bachelor's degree or higher). These users also show greater utilization of transient healthcare settings like emergency departments and urgent care.

This pattern suggests that AI health tools may be filling gaps where traditional primary care access falls short. Patients who lack a consistent primary care relationship, face longer wait times, or experience cost barriers may turn to AI as a more accessible—if imperfect—alternative.

What Are Patients Asking AI?

Patient AI queries cluster into several distinct categories:

• Symptom interpretation and self-diagnosis: "I have a headache behind my eyes and a stiff neck—what could this be?"
• Lab result explanation: Uploading MyChart screenshots and asking what their numbers mean
• Medication questions: Side effects, interactions, whether a prescribed medication is "really necessary"
• Treatment validation: Checking whether their doctor's recommendation aligns with what the AI suggests
• Understanding diagnoses: "What is rheumatoid arthritis and what should I expect?"
• Preparing for appointments: Generating questions to ask, understanding what tests might be ordered

A study examining AI chatbot use for perioperative health information found four key themes: patients prefer AI chatbots over traditional search engines for their conversational interface, they appreciate the improved accessibility and perceived quality of information, they sometimes prefer AI interaction over human interaction for sensitive questions, and they recognize that effective prompting skills matter for getting useful responses.

Why Patients Choose AI Over Traditional Sources

The appeal is understandable when you consider the patient's perspective. As one patient told the New York Times: "ChatGPT has all day for me—it never rushes me out of the chat." This reflects a fundamental tension in modern healthcare: the average primary care visit lasts about 18 minutes, with much of that time consumed by documentation and administrative tasks. AI offers unlimited time, immediate availability, and—crucially—no judgment.

Research on patient motivations identifies several drivers:

• Accessibility: Available 24/7, no appointment needed, no insurance required
• Perceived comprehensiveness: AI provides longer, more detailed explanations than time-constrained clinicians often can
• Reduced stigma: For sensitive topics (mental health, sexual health, embarrassing symptoms), AI feels "safer"
• Cost: Free or low-cost compared to a physician visit
• Control: Patients can ask follow-up questions without worrying about taking too much time

Research from NYU found that patients had difficulty distinguishing between AI-generated and physician-generated responses to health queries—correctly identifying the source only about 65% of the time. This suggests that AI responses are reaching a level of sophistication that makes them credible to patients, for better or worse.

Why Patients Trust AI—And What It Teaches Us

Here's an uncomfortable truth that the research reveals: AI chatbots often outperform physicians in perceived empathy. A landmark JAMA Internal Medicine study found that healthcare professionals preferred ChatGPT responses over physician responses 79% of the time. Even more striking: empathetic responses were 9.8 times more common from ChatGPT (45.1%) than from physicians (4.6%).

Why? The differences are instructive:

• Length: ChatGPT responses averaged 168-245 words; physician responses averaged 17-62 words
• Validation: AI consistently acknowledges patient concerns before providing information
• Patience: AI never rushes, never seems annoyed by "obvious" questions, never makes patients feel burdensome
• Availability: AI is there at 2 AM when anxiety peaks, with no wait time or cost

The Limitations Patients Don't See

While patients perceive AI as comprehensive and knowledgeable, several critical limitations remain invisible to non-expert users. First, AI tools like ChatGPT lack real-time medical knowledge—they're trained on data with a cutoff date and don't incorporate the latest research, drug alerts, or emerging treatment protocols. A patient asking about a recently approved medication may receive outdated or simply fabricated information.

Second, AI cannot perform a physical examination. The entire diagnostic process in medicine often hinges on findings that can only be obtained through direct observation and touch: the texture of a skin lesion, the quality of a heart murmur, the tenderness pattern of an abdomen. AI is limited to what patients describe—and patients often don't know what's clinically significant to mention.

Third, AI lacks the ability to integrate the subtle contextual cues that experienced clinicians process automatically: the patient's affect, their living situation, their reliability as a historian, the nonverbal signals that suggest whether symptoms are being minimized or amplified. A study by symptom checker researchers found that even with perfect information, AI-based tools listed the correct diagnosis in the top three options only 51% of the time, compared to 84% for physicians given the same vignettes.

Finally, AI can hallucinate—generating confident-sounding information that is entirely fabricated. Studies have documented AI chatbots inventing medical research, citing papers that don't exist, and recommending treatments with no evidence base. Patients have no way to distinguish hallucinated information from accurate information without independent verification.

Part 2: How AI Health Use Shows Up in Clinical Encounters

The Prepared Patient

Some patients arrive having done their AI homework in constructive ways. They've generated lists of questions, researched their symptoms, and come ready for an informed conversation. An internist at Beth Israel Deaconess described a patient who uploaded his lab results to ChatGPT and arrived with organized questions. "I welcome patients showing me how they use AI," he noted. "Their research creates an opportunity for discussion."

These encounters can be efficient and satisfying for both parties. The patient feels heard and prepared; the clinician can address specific concerns rather than starting from zero.

The Worried Well (Now More Worried)

AI chatbots, when prompted with symptoms, tend to generate comprehensive differential diagnoses—including rare and serious conditions. A patient with a headache might receive a list that includes tension headache, migraine, sinusitis, but also meningitis, brain tumor, and aneurysm. Without clinical context to weight these possibilities, patients may fixate on the most frightening options.

An emergency physician who tested ChatGPT on real patient presentations found it performed reasonably well for classic presentations with complete information, but noted: "Most actual patient cases are not classic. The vast majority of any medical encounter is figuring out the correct patient narrative." The AI missed an ectopic pregnancy in a patient presenting with abdominal pain—a potentially fatal diagnosis that required the nuanced history-taking that revealed she didn't know she was pregnant.

The Second Opinion Seeker

Increasingly, patients use AI to validate—or challenge—their physician's recommendations. This can manifest as straightforward questions ("ChatGPT said I should ask about X—what do you think?") or more confrontational challenges ("The AI says I don't need this medication—why are you prescribing it?").

A Customertimes survey found that 40% of Americans are willing to follow medical advice generated by AI. While most still expect AI to serve as a supportive tool rather than a physician replacement, the willingness to act on AI recommendations without professional input raises safety concerns.

The Vulnerable Patient

Perhaps most concerning: research published in JMIR found that patients failed to distinguish potentially harmful AI advice from safe advice. While physicians lowered their ratings for responses they identified as harmful, patients' assessments of empathy and usefulness remained unchanged. Most harmful responses involved either overtreatment/overdiagnosis or undertreatment/underdiagnosis—errors that require clinical expertise to recognize.

The study noted that "profound knowledge of the specific field is necessary to identify harmful advice"—such as knowing that gallstones greater than 3 cm are associated with increased cancer risk. Patients simply don't have the framework to evaluate clinical accuracy.

Recognizing AI-Influenced Encounters

Clinicians should listen for signals that a patient has consulted AI:

• Technical or clinical terminology that seems memorized rather than understood
• References to "reading" about their condition that sound conversational rather than like standard medical websites
• Questions about specific diagnostic possibilities that seem out of proportion to their symptoms
• Requests for tests or treatments "because the AI said..."
• Printed or screenshotted AI conversations

Consider simply asking: "Have you looked this up online or talked to any AI tools about it?" Normalizing the question removes stigma and opens dialogue.

Part 3: Partnering With AI-Using Patients

Reframe the Dynamic

The instinct to dismiss AI-derived information or feel threatened by it is understandable but counterproductive. Patients who research their health—through any medium—are demonstrating engagement. That's valuable.

Consider the analogy of a patient who arrives having read everything they could find about their diagnosis. We generally view this positively, even when their sources are imperfect. AI use is similar: it represents patients taking initiative in their healthcare. Our job is to help them do it effectively.

The "Curious Colleague" Approach

Rather than positioning yourself as the authority correcting AI mistakes, try approaching the conversation as a collaborative evaluation of information. Some phrases that work:

• "That's interesting—let's look at what the AI told you and see how it applies to your specific situation."
• "AI can be helpful for general information. What it can't do is examine you or know your full history. Let me show you what I'm seeing that changes the picture."
• "You've done some good research. Let me help you figure out which parts apply to you."
• "The AI gave you a good starting list. Here's how I'm thinking about narrowing it down based on your exam."

Addressing Discrepancies

When your recommendation differs from what the AI suggested, transparency is key:

1. Acknowledge the AI's reasoning: "Based on symptoms alone, the AI's suggestion makes sense."
2. Explain your additional data: "But I can see that your exam shows X, your history includes Y, and that changes the probability significantly."
3. Make your reasoning visible: "Here's why I'm less worried about the scary diagnosis and more focused on..."
4. Invite questions: "Does that help explain why I'm thinking differently? What other questions do you have?"

When the AI Was Right (Or Partially Right)

Sometimes patients bring AI-generated insights that are genuinely useful—a medication interaction you might have missed, a question about a relevant clinical trial, or a reasonable concern that warrants investigation. Acknowledge this:

• "Good catch—that's worth checking."
• "The AI actually raised a reasonable point. Let me look into that."
• "I hadn't considered that angle—I'm glad you brought it up."

This builds trust and models the appropriate use of AI as a supplement to—not replacement for—clinical judgment.

Communication Scripts for Common Situations

Managing Anxiety Amplified by AI

For patients whose AI research has increased their anxiety, the clinical approach matters:

1. Validate the concern: "It's understandable that seeing 'brain tumor' on a list would be frightening."
2. Explain probability: "AI lists everything possible, but doesn't tell you what's likely. For your symptoms, in your age group, with your exam, here's what's much more probable..."
3. Provide concrete reassurance: "Here's specifically what I looked for on your exam that tells me we don't need to worry about X."
4. Create a safety net: "If you develop any of these specific symptoms, that would change my thinking and you should call right away."

Part 4: Teaching Patients to Use AI More Effectively

Rather than discouraging AI use—which is unlikely to work and may damage trust—consider equipping patients with skills to use these tools more safely. This represents a form of digital health literacy that's becoming as important as understanding how to read a nutrition label.

Core Principles for Patients

1. AI Is a Starting Point, Not a Destination

Help patients understand that AI can be useful for generating questions, understanding terminology, or exploring possibilities—but it cannot examine them, know their full history, or make clinical decisions. The appropriate endpoint for serious health concerns is always a qualified professional.

2. Prompting Matters

Research shows that the quality of AI responses depends heavily on how questions are framed. Teaching patients basic prompting skills can improve the information they receive:

• Include relevant personal context (age, relevant medical history, current medications)
• Ask the AI to "act as a clinician" or "explain as if to a patient"
• Ask one focused question at a time rather than multiple questions
• Request that the AI explain its reasoning
• Ask for counterarguments or alternative explanations

3. Privacy Isn't Guaranteed

Consumer AI tools are not HIPAA-compliant. Data entered into ChatGPT, Claude, or similar tools goes to companies that may use it for training or other purposes. Patients should avoid including identifying information (name, date of birth, Social Security number) in health queries.

4. AI Can Hallucinate

Large language models can generate confident-sounding but entirely fabricated information. Patients should be especially skeptical of specific claims (drug interactions, dosages, treatment protocols) and verify anything important with a professional or authoritative medical source.

5. Emergency Symptoms Require Emergency Care

AI is not appropriate for evaluating urgent symptoms. Patients should understand that chest pain, severe shortness of breath, neurological changes, or other potentially serious symptoms require immediate professional evaluation—not an AI conversation.

Practical Tips for Patient AI Use

Consider providing patients with these concrete strategies:

The "Verify Big Claims" Rule

As one physician advises: "If it tells you something that you think is really big, verify it with an expert human." This simple heuristic captures much of what patients need to know. AI can be useful for routine questions and background information. For anything significant—a worrisome symptom, a treatment decision, a major diagnosis—professional verification is essential.

Part 5: Creating AI Tools for Your Patients

Rather than hoping patients use general-purpose AI wisely, you can create custom AI tools tailored to your practice and patients. This gives you control over the information patients receive while meeting them where they are. No coding required—just the skills you've already developed in this course.

Why Create Patient-Facing AI Tools?

• Control the information: Your custom GPT can only reference your approved materials, protocols, and guidance
• Reduce after-hours burden: Patients get immediate answers to routine questions without calling the on-call line
• Consistent messaging: Every patient gets the same accurate information, in your voice
• Extend your care: Patients feel supported between visits
• Triage appropriately: Build in guidance about when to seek urgent care vs. wait for an appointment

Ideas for Patient-Facing Custom GPTs

How to Build It

If you've read the Everyday Ways to Use AI topic, you already have the skills. The process is identical:

1. Choose your platform: Custom GPTs (ChatGPT Plus) or Claude Projects work well. Custom GPTs can be shared via link.
2. Write clear instructions: Define the GPT's role, what it should and shouldn't discuss, and when to recommend professional care.
3. Upload your materials: Patient handouts, FAQs, post-procedure instructions, condition guides.
4. Test thoroughly: Try to break it. Ask questions that should trigger safety warnings.
5. Share with patients: Provide the link at checkout, in follow-up emails, or on your patient portal.

Part 6: Special Populations and Considerations

Equity Considerations

The intersection of AI health tools and healthcare disparities deserves attention. AI use both reflects and potentially exacerbates existing inequities.

The finding that AI health tool users have lower educational attainment and rely more on emergency/urgent care settings suggests these tools may serve as a workaround for patients with less access to traditional primary care. According to the American Hospital Association, over 26% of Medicare beneficiaries lack home computer or smartphone access, and more than 22% of rural Americans lack adequate broadband.

AI systems are trained on data that reflects historical patterns—including healthcare disparities. Populations that have been historically marginalized are underrepresented in training datasets, potentially leading to less accurate or relevant information for these patients.

Pediatric and Adolescent Considerations

Parents researching their children's health present unique dynamics:

• Developmental information: Parents frequently query AI about milestones, autism spectrum features, ADHD symptoms. AI may provide generic checklists without context about normal variation.
• Adolescent self-diagnosis: Teenagers increasingly use AI to research mental health conditions and physical symptoms. This can represent healthy self-awareness or lead to premature self-labeling.
• Privacy for adolescents: Teens may use AI for health questions they're uncomfortable asking parents or physicians. Consider asking adolescents directly about AI health use in confidential portions of visits.

Mental Health Specific Concerns

AI use for mental health information deserves particular attention:

• AI tools may underestimate suicide risk, with some studies showing AI failing to appropriately escalate urgent safety concerns
• Self-diagnosis of personality disorders or complex trauma can be harmful without professional assessment
• AI-generated coping strategies may be generic and miss patient-specific contraindications
• Patients in crisis should not be relying on AI chatbots—ensure patients know crisis resources

Chronic Disease Management

Patients with chronic conditions present particular opportunities and risks. These patients have ongoing need for information and may make daily self-management decisions.

Guidance for chronic disease patients using AI:

• Use AI for understanding and tracking, not for treatment changes
• Bring AI-generated questions to your clinical team
• Never adjust insulin, anticoagulation, or other high-risk medications based on AI advice
• Use AI to help communicate symptoms accurately to your healthcare team

Part 6: The Opportunity

Patient AI use isn't going away. The question for clinicians is whether to view it as an obstacle or an opening.

Patients who consult AI are engaged in their health. They're taking initiative, seeking information, and thinking about their conditions between visits. This is exactly what we want from patients—active participation in their healthcare.

AI can surface questions patients wouldn't otherwise ask. When a patient brings an AI-generated list of considerations, they may raise issues that wouldn't have come up in a routine visit.

Teaching AI literacy is a form of patient education. Helping patients use AI tools effectively reinforces critical thinking about health information generally.

Transparency about AI builds trust. When clinicians openly discuss AI's strengths and limitations—rather than dismissing it—they demonstrate honesty and respect for patients' autonomy.

Building AI Literacy Into Your Practice

Consider how you might systematically address AI use:

• Intake forms: Add a question about what resources patients have consulted, including AI tools
• Waiting room materials: Provide guidance on using AI tools effectively and their limitations
• Visit routine: Normalize asking about AI use as part of history-taking
• After-visit materials: Recommend trusted resources as alternatives to unvetted AI queries
• Staff training: Ensure all team members can respond to patient questions consistently

Documentation Considerations

When patient AI use is relevant to clinical decision-making, consider documenting it:

• Note AI-influenced concerns: "Patient expressed concern about [condition] based on AI chatbot research. Discussed differential diagnosis and clinical reasoning."
• Document education provided: "Discussed appropriate use of AI health tools, including limitations."
• Record disagreements: "Patient preferred alternative approach based on AI recommendation. Discussed risks and benefits."

Quick Reference: When Patients Bring AI to the Visit

Key Messages for Patients

1. AI is a starting point, not a diagnosis.
2. Better prompts yield better information.
3. Your data isn't private—avoid identifying information.
4. AI can hallucinate—verify important claims.
5. Emergency symptoms need emergency care, not AI.

Resources for Further Learning

Essential Reading

Podcasts

For Your Patients

Building Patient-Facing AI Tools

Patient-Facing AI Health Tools

Beyond general-purpose chatbots like ChatGPT and Gemini, a growing category of purpose-built AI health tools is emerging specifically for patient use. These tools aim to provide more clinically appropriate responses than general AI, often with built-in safeguards, physician oversight, or connections to care. Patients may mention using these tools, and it's worth understanding what they offer.

Purpose-Built Patient Health AI

What Makes These Different From ChatGPT?

Purpose-built health AI tools typically offer several advantages over general-purpose chatbots:

• Clinical guardrails: Built-in protocols for recognizing emergencies and escalating appropriately
• Structured symptom collection: Systematic approaches that gather relevant clinical information
• Physician oversight: Some tools have physicians reviewing outputs or available for escalation
• Healthcare compliance: Many are HIPAA-compliant, unlike consumer chatbots
• Care connections: Pathways to telehealth, appointments, or specialist consultations

That said, these tools still share core limitations with all AI: they cannot perform physical examinations, they may miss atypical presentations, and they work best as a complement to—not replacement for—professional care.

Talking With Patients About These Tools

When patients mention using purpose-built health AI:

• Ask what they learned: "What did the app tell you? Let's see how that fits with what I'm finding."
• Validate appropriate use: "Using that to organize your symptoms before coming in was helpful."
• Clarify limitations: "These tools are good at generating possibilities, but they can't examine you or know your full history."
• Distinguish from diagnosis: "The app gave you a list of possibilities—let me help narrow that down based on your exam."