What AI can do, where it can go wrong, and the safeguards needed to protect privacy, judgement and human relationships.
By Isabel Peace | September 2026
Artificial intelligence is already part of healthcare. It helps analyse medical images, organise records, produce clinical notes, support research and assist with selected decisions and procedures.
In Australia, the Therapeutic Goods Administration maintains a register of medical devices incorporating AI or machine learning. The Australian Digital Health Agency has also established an expert advisory group focused on the clinically safe implementation of AI-enabled care.
This movement will continue. The useful question is no longer whether healthcare will use AI. We need to ask where it can genuinely improve care, where it may cause harm, and what responsibility remains with the practitioner.
I have been thinking deeply about this while completing postgraduate study in large language models and agentic AI. A recent paper sent to me by my scientist friend Igor Nazarov sharpened the question. His central concern is that healthcare does not suffer only from a lack of information. It also faces a reasoning, ethical and relational challenge.
A machine can process an enormous volume of data, but somebody must still judge what matters, recognise uncertainty, take responsibility and care for the person affected by the decision.
I agree with that concern. I also believe AI can offer more than administration. With careful boundaries, it can support pattern recognition, clinical preparation, education, follow-up and continuity. The important distinction is between assistance and authority.
Different forms of AI perform different jobs.
Machine learning can classify information or predict an outcome from patterns in data. Deep learning is especially useful with complex images and signals, which is why it is being explored in radiology, pathology, retinal screening and cardiac monitoring.
Generative AI creates new content. A large language model can draft a summary, turn technical material into clearer language, organise a complex history or suggest questions for a practitioner to consider. Its fluency can be misleading.
A response may sound authoritative while containing an error, overlooking important context or inventing a detail.
Agentic AI can complete a sequence of permitted actions rather than answering only one question. An agent might retrieve authorised information, compare it with defined rules, prepare a draft and check that required elements are present.
This could be very useful in a busy clinic. It also raises the stakes because a system that can act needs firmer limits, narrower access and better monitoring than a simple writing assistant.
One of the most valuable lessons from my studies has less to do with computer coding than with human responsibility.
AI can produce a convincing piece of code, document or analysis remarkably quickly. However, the person using it must still understand the original problem, verify the logic and test what happens when the situation is unusual, incomplete or unexpected.
Something can look polished and technically correct while solving the wrong problem.
Healthcare carries the same risk. An AI system can only respond to the information it receives. It may not know that a symptom has suddenly changed, that a medication was omitted from the history, that an image was distorted by lighting or that a person is too vulnerable for an automated interaction.
This becomes even more important as AI moves from assisting with information to taking action.
A writing tool prepares something for a human to review. An AI agent may be able to search records, communicate with other systems, update information or complete a sequence of tasks. Each additional ability, connection and permission increases both its usefulness and the possible consequences of an error.
The question is therefore no longer only, “Is this answer accurate?” We must also ask, “What is this system allowed to do before a human checks it?”
Human oversight cannot be added as a final proofreading step. It needs to be designed into the system from the beginning.
Some of the most sensible early uses reduce repetitive work. AI-assisted documentation can prepare notes and letters for practitioner review. It can help organise long histories, retrieve relevant information and create patient education in accessible language. Used well, this may allow clinicians to spend less time typing and more time listening.
Other applications include analysing images and physiological signals, identifying trends across repeated measurements, supporting drug discovery, monitoring data from wearable devices and assisting surgical planning or navigation.
These applications vary greatly in maturity and risk. A tool that formats a letter is not equivalent to software that suggests a diagnosis or treatment.
Australian regulation recognises this distinction. The TGA notes that a digital scribe may begin as a documentation tool, but if it starts suggesting diagnoses or treatments, its intended purpose may cause it to become a regulated medical device. Capability, purpose and risk all matter.
Pattern recognition is already central to my work. I often need to consider relationships between nutrition, pathology, digestion, sleep, hormones, stress, pain, emotional history and changes over time.
Chinese medicine adds another layer through tongue and pulse findings, constitutional tendencies and patterns that may move as treatment progresses.
Research is exploring AI-supported tongue-image analysis, extraction of knowledge from traditional texts, prediction of acupuncture responses and analysis of point combinations.
A 2025 review in Frontiers in Medicine described promising applications while also identifying small or inconsistent datasets, variable labelling, limited clinical validation and difficulty explaining how some models reach their conclusions.
Those limits are crucial. A tongue photograph changes with lighting, camera settings, food, hydration and timing.
The same symptom can mean different things for two people. A traditional pattern label may be useful, but it doesn’t capture the person’s full medical, emotional, or social context.
In complementary and integrative healthcare, we must also separate established evidence, emerging research, clinical observation, traditional theory and speculation. AI can blend these categories into one smooth explanation unless the practitioner deliberately keeps them distinct.
AI can compare patterns. It cannot assume that the pattern is the person.
AI may help with psychoeducation, journalling prompts, skills practice, appointment preparation and structured reflection between consultations.
Some people may find it easier to begin putting a difficult experience into words with this support.
Psychotherapy, however, is not the production of comforting sentences. It involves relationship, timing, body language, nervous-system cues, rupture and repair, and the practitioner’s capacity to recognise when the planned direction is no longer safe or helpful.
A system may generate empathic language without experiencing empathy or accepting responsibility for the relationship.
My interest is therefore in AI that supports psychotherapy around the edges, not technology that impersonates a therapist. It must never become a substitute for crisis support, qualified assessment or the human therapeutic relationship.
Health information is sensitive. In Australia, the Privacy Act and Australian Privacy Principles apply when AI systems handle personal information.
The Office of the Australian Information Commissioner recommends that organisations avoid placing personal, particularly sensitive, information into publicly available generative AI tools.
It also advises organisations to assess whether AI is necessary, minimise the data used, conduct proper due diligence, explain material uses clearly and review systems throughout their lifecycle.
For practitioners, this means convenience is not enough. Before a tool touches clinical information, we need to understand what data it receives, where that information goes, who may access it, how long it is retained and whether it may be used for training.
We also need a plan for incorrect output, unauthorised access, service failure and changing software behaviour.
De-identification helps, but it is not magic. A detailed combination of age, occupation, location, family circumstances and uncommon symptoms may still identify someone. Data minimisation means using only what is necessary for a defined task.
My study is helping me move from simply using AI to understanding how these systems are built, connected, tested and governed. I am developing the following standard for any future use within Flow in Nature:
These standards also mean being honest about what is still under development. I will not present an experimental tool as a diagnostic authority, and I will not publish a capability before its privacy, security and clinical boundaries have been properly examined.
Once carefully developed, AI support could make care feel more organised and continuous. With appropriate consent and safeguards, patients may receive clearer preparation before a consultation, better organised histories, personalised education reviewed by their practitioner and easier summaries of the plan discussed in the room.
An agent might help track agreed symptoms or goals between appointments, prepare questions for the next consultation, identify missing information for the practitioner to review or select relevant resources from a controlled Flow in Nature knowledge base. It could also reduce repetitive administrative work, leaving more of the consultation for listening, assessment and treatment.
There are firm boundaries. An agent should not independently diagnose, prescribe herbs or supplements, make psychotherapy decisions, interpret a red flag without escalation or contact a patient without permission. Its role and limits should be visible rather than hidden behind a human-sounding interface.
Igor’s paper argues that the human clinician must remain the moral and relational guardian of care. I share that conclusion, although I see a broader support role for AI.
It can help us hold information, notice relationships and maintain continuity. It may even help practitioners examine their reasoning by presenting alternatives and exposing gaps.
None of this removes the need for a skilled person who understands the clinical domain, the limits of the evidence and the individual sitting in front of them.
The future I want is technologically capable and recognisably human. If AI saves time, that time should return to attention. If it finds a pattern, the practitioner should test its relevance. If it drafts an answer, a responsible human should remain accountable for what happens next.
Responsible engagement with AI does not require blind optimism or constant fear. It requires curiosity, informed caution and a willingness to keep asking who benefits, who may be harmed, who remains accountable and how human control will be preserved as these systems become more capable.
Progress will not be measured by how much healthcare we automate. It will be whether the technology helps us provide safer, clearer and more attentive care without surrendering privacy, judgement or relationship.
If you would like to understand more about how I work and how these new technologies can support your health and wellbeing now and in the future, feel free to reach out by phone or book a call.
