Should we be worried when Dr. Jones hires the AI version of himself?

This partner never calls in sick, never complains, never asks for a raise, never sleeps and never retires — but never signs the note.

An “actress” generated by artificial intelligence (AI) and named Tilly Norwood has been announced as the lead in a feature film called Misaligned, still in development by the London studio Xicoia. The Screen Actors Guild has stated bluntly that Tilly is not an actor, but an artificial construct made by software. It has been trained on the work of countless performers, without their permission and without their being compensated.

Whether Hollywood embraces Tilly is beside the point. A digital “person” can now have a recognizable face, a familiar voice and a seemingly natural conversation. And the same technology can move from the movie set into the examination room.

Imagine a primary-care physician creating a digital double –– not a generic chatbot with a stock photograph, but an AI assistant that looks like him, speaks in his voice and is programmed around his way of practicing medicine. After the patient is in the room and the door has closed, the AI assistant appears on a secure monitor:

“Good morning, Mrs. Smith. I’m Dr. Jones’s AI-enabled digital assistant. I use Dr. Jones’s voice and practice approach to help gather information before he comes in. I am not Dr. Jones. I cannot diagnose, prescribe or replace a visit with Dr. Jones.”

That introduction is not a formality. It is the foundation of the model. So the question is not whether this is coming — it is. What is gained and what is at risk when Dr. Jones hires the AI version of himself?

The case for hiring your own clone

An AI assistant with the familiar face, voice and manner of the patient’s own physician can make intake feel like a continuation of an established relationship — not an encounter with an unfamiliar, generic bot. It may improve patient comfort, engagement and trust while still making clear that the physician remains responsible for the care and final clinical decisions.

It absorbs work that never required a license. Independent primary care is drowning in predictable labor: interval histories, medication reconciliation, care-gap screening, questionnaires and the same education delivered dozens of times a week. A digital double can run the routine chronic disease and geriatric check-ins — symptom control, adherence, monitoring, overdue screenings — consistently, every time.

It changes what the visit is for. A 20-minute encounter should not be consumed by intake questions, chasing medication discrepancies or learning at minute 14 that the patient was hospitalized two weeks ago. If the physician walks in already knowing the patient’s priorities, what has changed since the last visit and any possible red flags, the visit shifts toward examination, diagnosis, counseling and shared decision-making, the parts that require a physician.

It is consistent in a way humans are not. An AI assistant is never rushed and asks the tenth patient of the day the same questions it asked the first. It does not skip the social history because the schedule slipped.

It can carry the practice’s culture, not a vendor’s script. Most health care chatbots are generic and detached from the physician-patient relationship. A digital double is configured around one physician and defined workflows. Experienced physicians develop habits that reflect judgment, such as asking about affordability before labeling a patient “noncompliant.” Those habits translate into approved phrasing: “Many people miss doses because medicines are expensive, confusing or make them feel poorly. Has that been happening for you?” rather than “Are you adherent to your prescribed medication regimen?” That is not merely friendlier. It invites honesty without blame.

It presents an argument for independence. The tasks most likely to be automated well are those that make small practices uncompetitive with systems that have layers of staff.

The case against — or at least, the costs

There are factors that work against creating a new virtual doctor.

Realism increases liability rather than reducing it. A patient who sees the physician’s face and hears his voice may reasonably believe the doctor is involved, monitoring live or giving advice. The more persuasive the imitation, the more explicit the correction must be. Calling a product an “assistant” does not determine its legal status; what matters is what it does. An AI assistant that collects approved information and routes concerning answers to a clinician may be a sophisticated intake tool. One that interprets symptoms, offers reassurance or decides who needs care becomes a clinical decision maker — with exposure under malpractice principles, state practice acts, privacy and consumer-protection law and, potentially, device regulation.

The most dangerous failure mode may be silence rather than a visible error. The dangerous scenario is a patient reporting chest pressure, focal weakness, fainting, severe shortness of breath, new confusion, suicidal thoughts, major bleeding or a possible severe drug reaction, and getting a smooth, algorithmic nonresponse. The system must halt, acknowledge that a clinician must be involved and alert a named human. “Sends an alert" is not a plan. A plan spells out who gets the alert, how quickly they must act, who covers if they're unavailable, and what happens if no one responds.

Documentation is its own liability. An AI-generated history can be incomplete, misheard or wrong, asserting things the patient never said, omitting uncertainty or hardening a tentative concern into a definitive statement. A physician should never sign a note because it reads well. The record should distinguish patient-reported information, AI-generated summaries and the physician’s own assessment, and AI documentation should remain a draft until a clinician accepts it.

Vendor assurances are not a defense. The Health Insurance Portability and Accountability Act (HIPAA) remains in effect in the age of AI. “We’re HIPAA compliant” and “We’re only an assistant” do not resolve the practice’s risk or substitute for independent review of intended use, contract terms, performance evidence, insurance implications and state law.

The privacy surface is larger than the closed door. These systems can capture audio, video, facial imagery, transcripts, behavioral-health data and, if linked to an electronic health record, a wide swath of protected health information (PHI).

Settle the terms up front:

• What is recorded and retained?
• Who can access recordings and audit logs?
• Do interactions train the vendor’s model?
• What happens to the data and voice model if you switch vendors?
• Does a business associate agreement exist?

Give the assistant only the minimum necessary — selected diagnoses, medications, allergies, recent results and the plan –– not the whole chart. The Health and Human Services Department is clear: Responsibility for safeguarding electronic PHI remains with the covered entity, regardless of the technology.

You may be giving away your face. A physician’s likeness and voice are professional assets. The vendor agreement should say who controls the avatar, voice model, recordings and prompts; bar reuse of the likeness for another customer or for model training; and let the practice deactivate the avatar at termination.

The 24/7 question

When it comes to availability, the appeal is obvious. Patients ask questions at 9 p.m. or on Sunday mornings. A digital double could take a postdischarge check-in, collect symptom detail before Monday’s callback list or confirm a new prescription was actually filled, turning dead hours into structured information waiting when the office opens.

The problem is that the guardrail on the daytime version is a human standing behind it. At 9 p.m., there is no nurse to receive the alert and no one to notice that the patient described crushing chest pressure and then hung up. An unanswered escalation is worse than no interaction at all, because the patient reasonably believes the practice has been told. After hours, the AI assistant should state plainly that no one is reviewing this tonight, when a human will respond and that anything urgent goes to the on-call line or 911 now. Any red-flag answer should end the session with that instruction rather than continue collecting history. Absent a staffed escalation path, there is real liability when using the assistant after normal business hours.

The guardrails that make the upside available

The temptation is to focus on realism. That is the wrong first question. The first question is: What may the AI assistant do, what is it forbidden to do and who is accountable if it fails? It will be essential to have a written governance framework before the first patient interaction. Also needed are the following:

• Plain disclosure at the start of every interaction, repeated whenever confusion is likely.
• Meaningful choice. Patients can decline the AI assistant and ask for a staff member without delay, penalty or reduced service.
• Strict scope limits. It may collect a history, reconcile medications, administer approved questionnaires, provide patient education resources, summarize the findings, and flag concerns for the physician or appropriate care-team member. It must not diagnose, prescribe, change a plan, declare symptoms nonurgent or decide a patient does not need to be seen.
• Real-time escalation rules, with named recipients, response times and backups.
• No invisible delegation. Flag missing information, uncertain answers and statements that are not understood.
• Auditing. Track missed alerts, incorrect summaries, medication errors, complaints, refusals and performance differences across patient populations.

This aligns with professional guidance. The American Medical Association holds that when AI affects patient care, access, decision-making or the medical record, its use should be disclosed and documented. Patients should be able to request review by a licensed clinician, and AI-generated content should not be produced on a physician’s behalf without consent and final review.

Starting small to prove safety

The sensible first version is not an autonomous virtual doctor. It is a limited, transparent, supervised intake tool used only after rooming, with willing patients, and with a narrow scope: interval histories, medication updates, care-gap screening and a short summary for clinician review.

Then test it against reality. Does it understand older voices and accents? Does it handle hearing impairment, limited English, cognitive impairment and distress? Does it recognize uncertainty and escalate correctly? Does it reduce burden or just create another inbox? Measure physician time saved, medication discrepancies found, missed escalations, refusal rates and gaps between AI-collected and clinician-confirmed histories.

Hollywood’s AI actress is a preview, not a clinical model. In medicine, realism is the least important measure of success. The digital double that works will not be the one that convinces a patient it is Dr. Jones. It will be the one that says plainly what it is, stays inside carefully drawn boundaries, guards patient information and steps aside the moment a case needs clinical judgment. It absorbs the predictable paperwork so the practice runs leaner and the physician arrives at each visit fully prepared and free to focus on the patient. When the partnership is done right, a physician and an AI assistant deliver more together than either does alone: faster access, fewer things missed, better documentation and follow-through, and measurably better care. Efficiency earns its place only when it strengthens the bond with the doctor the patient actually came to see.

Robert Resnik, M.D., MBA, is a board-certified internal medicine physician practicing in Cary, North Carolina. He earned his medical degree from Eastern Virginia Medical School and completed his residency at East Carolina University. He also holds an MBA from Duke University.