How AI Symptom Checking Works

Health Copilot uses a structured approach to symptom analysis that mirrors the clinical reasoning process doctors use, adapted for a conversational format. Understanding this process helps you provide better information and interpret the results appropriately.

Natural language understanding: The first layer of the system translates your everyday language into clinical concepts. When you say 'my stomach hurts after eating,' the AI maps this to several clinical parameters: symptom (pain), location (abdominal — with follow-up needed to determine which quadrant), timing (postprandial), and trigger (food intake). This mapping happens automatically, so you never need to learn medical terminology. The AI handles colloquialisms, regional expressions, and imprecise descriptions — 'I feel like my chest is tight' is understood as a description that could indicate respiratory, cardiac, or musculoskeletal issues, triggering appropriate follow-up questions for each possibility.

Differential diagnosis tree: Based on your initial description, the AI constructs a tree of possible conditions. Each follow-up question is designed to prune branches of this tree. If you report that your abdominal pain is in the upper right quadrant, conditions affecting the lower left quadrant become less likely and are moved down the list. If you report the pain worsens after fatty meals specifically, gallbladder-related conditions move up. The tree typically starts with 15-30 possible conditions for a common symptom and narrows to 3-7 after the questioning phase.

Bayesian probability weighting: The AI does not just match symptoms to diseases — it uses probabilistic reasoning that accounts for base rates. Chest pain in a 28-year-old non-smoking woman has very different probabilistic implications than chest pain in a 62-year-old man with hypertension and diabetes. The AI factors in demographic information and any medical history you share to adjust its probability estimates. This is why it asks about age, sex, and relevant medical history — not to be intrusive, but because these factors genuinely change what conditions are more or less likely.

Safety-first design: The system is explicitly designed to prioritize sensitivity (not missing serious conditions) over specificity (correctly excluding non-serious conditions). This means the AI will sometimes include a serious condition on its possibility list even when it is statistically unlikely, if your symptoms have any overlap with that condition's presentation. Some users find this overly cautious. We consider it the right trade-off — the cost of false reassurance (missing something serious) is much higher than the cost of an unnecessary doctor visit.

Being transparent about limitations is especially important in health. Here is what the AI fundamentally cannot do, no matter how sophisticated it becomes, and why these limitations mean you still need a human healthcare provider.

It cannot examine you physically. A doctor's physical examination provides information that no description can fully convey. Pressing on your abdomen to check for tenderness, guarding, and rebound pain. Listening to your heart and lungs with a stethoscope. Looking at your throat, ears, or skin. Checking your range of motion. Measuring your blood pressure, heart rate, and oxygen saturation. These physical findings often differentiate between conditions that present with identical self-reported symptoms. An AI working from text descriptions alone is fundamentally limited in a way that will not be solved by better algorithms.

It cannot order or interpret tests. Many diagnoses require laboratory work (blood tests, urine tests, cultures), imaging (X-rays, CT scans, MRIs, ultrasounds), or specialized tests (ECGs, endoscopies, biopsies). The AI can explain which tests might be relevant to your situation and why — helping you understand what your doctor orders — but it cannot order those tests, receive the results, or provide a definitive interpretation. Lab Results Copilot can help you understand test results once you have them, but it supplements rather than replaces your doctor's interpretation.

It cannot prescribe treatment. The AI can explain general treatment approaches for various conditions (how acid reflux is typically treated, what medications are commonly used for hypertension, what physical therapy might look like for back pain), but it cannot prescribe specific medications, determine dosages for your situation, or account for drug interactions with your current medications. Treatment decisions require a licensed healthcare provider who can weigh benefits, risks, and your complete medical history.

It cannot provide emergency care. In an emergency, every second counts. The AI is designed to recognize emergency symptoms and immediately direct you to call 911 or visit an ER. But it cannot provide the interventions an emergency requires — CPR guidance, wound management, or stabilization. If you believe you are experiencing a medical emergency, call 911 first. The AI is a preparation and education tool, not an emergency response system.

It cannot replace the doctor-patient relationship. Medicine involves trust, continuity of care, knowledge of your personal and family history over time, and the human judgment that comes from years of clinical experience. The AI provides a point-in-time analysis based on what you share in a single conversation. Your doctor sees the full picture — and that full picture matters for your health.

Health is the domain where AI safety matters most. A false negative — telling someone their symptoms are not concerning when they actually indicate a serious condition — could be dangerous. Here is how every layer of the system is designed to prevent that scenario.

Emergency detection is always-on. Before any analysis begins, every user message is screened for emergency indicators. Descriptions of chest pain with shortness of breath, sudden severe headache, signs of stroke (facial drooping, arm weakness, speech difficulty), severe allergic reactions, thoughts of self-harm, or significant traumatic injuries trigger an immediate emergency response that interrupts any ongoing analysis. This screening runs continuously throughout the conversation, not just on the first message — because a user might mention a critical symptom casually mid-conversation.

Escalation triggers are conservative. The thresholds for recommending medical attention are deliberately set lower than a clinician might set them. A doctor seeing a healthy 25-year-old with 3 days of mild headaches might say 'take some Tylenol and call me if it gets worse.' The AI, unable to perform the physical examination that informed that clinical judgment, takes the more cautious approach of recommending a doctor visit if symptoms persist beyond a specified window. We accept higher false-positive rates (recommending doctor visits that turn out unnecessary) as an acceptable trade-off for lower false-negative rates (missing something that needed attention).

No definitive diagnoses, ever. The AI never says 'you have condition X.' It says 'your symptoms are most consistent with condition X, but this should be confirmed by a healthcare provider.' This is not just a legal disclaimer — it reflects a genuine limitation. Without physical examination and diagnostic testing, a definitive diagnosis based on self-reported symptoms alone would be irresponsible. The AI presents possibilities ranked by likelihood, not conclusions.

Medication safety boundaries. The AI will provide general information about common treatments but will not recommend specific dosages, will not suggest stopping prescribed medications, and will flag potential interactions when users mention multiple medications. For detailed medication questions, Medication Copilot provides more specialized analysis, always with the recommendation to verify with your pharmacist or prescribing physician.

Mental health sensitivity. If the conversation reveals symptoms of depression, anxiety, suicidal ideation, or other mental health concerns — whether the user asks about them directly or they emerge during symptom discussion — the AI responds with appropriate resources including crisis helplines (988 Suicide and Crisis Lifeline) and encouragement to seek professional mental health support. Mental Health Copilot provides more specialized support for ongoing mental health concerns.

We report that our AI includes the correct condition in its top-3 list 89% of the time and in its top-5 list 94% of the time. Here is what those numbers mean in practice, how they were measured, and what they do not tell you.

What 89% top-3 accuracy means: In our validation study of 1,000 symptom presentations with confirmed diagnoses, the AI's top-3 ranked conditions included the actual diagnosis 89% of the time. This means that if you read the AI's top 3 possibilities, there is roughly a 9-in-10 chance the correct condition is listed. For the remaining 11%, the correct diagnosis was often in positions 4-7 on the list, and in a small number of cases (approximately 3%), the correct diagnosis was not listed at all — typically because the condition was rare or presented atypically.

What the accuracy does NOT mean: It does not mean the AI correctly identifies the single correct condition 89% of the time. Symptom-based assessment is inherently probabilistic — the same set of symptoms can be caused by multiple different conditions, and without testing, narrowing to a single diagnosis is often impossible even for an experienced physician. The AI's job is to help you understand the range of possibilities so you and your doctor can investigate efficiently.

Where accuracy is highest: Common conditions with distinctive symptom patterns (urinary tract infections, acid reflux, seasonal allergies, strep throat, pink eye) have top-3 accuracy above 95%. The AI performs well when symptoms are specific and the symptom-disease association is strong.

Where accuracy is lowest: Conditions with vague or overlapping symptoms (chronic fatigue, non-specific pain, autoimmune conditions in early stages) have lower accuracy because multiple conditions produce similar presentations. The AI handles this honestly by presenting a broader list of possibilities and recommending professional evaluation to differentiate between them, rather than forcing a false sense of certainty.

The 99% emergency detection rate: This is the number we are most focused on getting as close to 100% as possible. Of 200 emergency cases in our validation set (confirmed heart attacks, strokes, appendicitis, severe allergic reactions, meningitis), the AI correctly flagged 198 as requiring immediate emergency attention. The 2 missed cases involved atypical cardiac presentations — and this is an active area of improvement. We continuously expand the emergency detection system to cover more atypical presentations.

Health information is among the most sensitive data that exists. Here is exactly what happens with the symptoms and health details you share with the AI, written plainly without corporate hedging.

During your conversation: The symptoms, medical history, and health details you share are processed in server memory to generate your assessment. This data exists in RAM only during your active session. It is not written to a database, not stored on disk, and not indexed or searchable.

After your session ends: All health data from your conversation is purged from memory within 60 seconds of session end (or 60 minutes of inactivity). There is no archive of your symptoms, no record of what conditions were discussed, and no way for anyone — including Copilotly employees — to retrieve your health conversation after the session ends.

We do not use your health data for AI training. Your symptom descriptions, medical history, and conversation content are never used to train or fine-tune AI models. The AI's medical knowledge comes from published medical literature and clinical datasets — not from user conversations.

No data sharing with any third party. Your health information is not shared with insurance companies, employers, pharmaceutical companies, data brokers, advertisers, or any other entity. Period. We have no data-sharing agreements that involve user health information. Our revenue comes from subscriptions, not data monetization.

HIPAA considerations: Copilotly is an educational health information tool, not a covered entity under HIPAA in the way that a hospital or insurance company is. However, we voluntarily implement security practices that meet or exceed HIPAA security standards, including encryption in transit (TLS 1.3), encryption at rest for any temporary data, access controls, and audit logging. For enterprise healthcare organizations evaluating Copilotly, we can execute a Business Associate Agreement (BAA) and provide our security documentation.

Your control: You decide what to share. The AI works with whatever information you provide — you are never required to share your real name, age, medical history, or any identifying information. The assessment may be more accurate with more context, but the choice of what to disclose is entirely yours.

AI symptom checking is a powerful tool when used appropriately and a poor substitute when used in the wrong situations. Here is our honest guidance on when Health Copilot adds value and when you should skip it and go straight to a medical professional.

Good use cases (where AI adds clear value):

• Understanding new mild-to-moderate symptoms: You have a rash that appeared 3 days ago, a cough that has lasted a week, or occasional dizziness. You want to understand what might be going on before deciding whether to see a doctor. The AI helps you assess urgency and prepare for a potential appointment.
• Preparing for a doctor visit: You already have an appointment and want to organize your thoughts, understand what questions to ask, and make the most of your limited appointment time. The AI's doctor-visit preparation summary is valuable regardless of the condition.
• Understanding a condition after diagnosis: Your doctor diagnosed you with something and you want more context — what the condition involves, what treatment options exist, what to expect. The AI provides patient-friendly explanations that supplement (not replace) your doctor's guidance.
• Middle-of-the-night symptom anxiety: It is 2 AM, your child has a fever, and you are not sure if it warrants an ER visit or can wait until morning. The AI helps you assess urgency based on specific symptom details so you can make a more informed decision.

Poor use cases (go straight to a professional):

• Obvious emergencies: Chest pain, difficulty breathing, severe bleeding, loss of consciousness, sudden weakness on one side of the body, severe allergic reactions. Call 911. Do not stop to consult an AI.
• Worsening known conditions: If you have a diagnosed chronic condition that is suddenly getting worse, contact your treating physician directly. They know your history and current treatment plan.
• Mental health crises: If you are experiencing suicidal thoughts, self-harm urges, or a psychiatric emergency, call 988 (Suicide and Crisis Lifeline) or go to your nearest ER. Mental Health Copilot is designed for ongoing support, not crisis intervention.
• Post-surgical complications: Contact your surgeon's office. They have specific post-operative protocols and need to know about complications directly.
• Seeking a second opinion on a serious diagnosis: For a genuine second opinion, see another physician. For understanding what a diagnosis means, the AI can help — but it is not qualified to confirm or refute another doctor's diagnosis. Second Opinion Copilot can help you formulate questions for an in-person second opinion consultation.