Artificial intelligence (AI) scribes using ambient voice technology (AVT) are rapidly becoming popular tools in primary care consultations. These use AI to transcribe digitally recorded consultations into medical records and documents.
Their proponents extol their ability to save administrative time and cognitive effort. The reduction in ‘cognitive burden’ has been associated with a lower risk of burnout. NHS England and health ministers view AVT as a tool to ‘free up more time’ to improve efficiency and see more patients.
For 5 months in 2025, I used AVT for most of my GP consultations. From this experience, I would like to highlight some of the risks of delegating these key cognitive tasks to AI.
As a poor typist, I was initially a very enthusiastic convert. However, after a few months of use, I became increasingly disconcerted by seeing patients for follow-up and realising I had no recollection of their previous (AVT documented) consultation.
This was markedly different from my usual follow-up experience, in which the combination of records and recall triggers memory. This was both embarrassing and profoundly unsettling. It also felt less safe and much less efficient.
Assuming that this was because it was not recording the consultation in my usual language, I started to amend the notes more actively. When this had no effect, I recognised it was a fundamental behaviour of my brain when using AVT. After discussion with colleagues who were also using AVT, I realised I was not unique.
We are all familiar with learning and committing information to memory. The commonest techniques include active listening and writing. This experience of using AVT to record consultations illustrates its ability to inhibit the normal memory acquisition process. Once we know AVT will do it, at some deep level we stop trying to remember. This observation is supported by evidence showing that using AI can disrupt and harm learning processes.
Thus, an inherent problem when delegating the recording process to this technology is that it prevents us from remembering our consultations and patients properly. There are several risks associated with this. These particularly impact our ability to perform aspects of care where deep, longitudinal knowledge of our patients is particularly important. This is key to providing continuity, holistic care, and managing complexity.
How then do we balance these risks and harms against the current and potential benefits of AI and AVT?
My own response has been to stop using AVT altogether. Although AVT significantly reduced the short-term cognitive stress I felt during consultations, I am more concerned about its impact on my longer-term functioning, particularly in the most complex consultations. Those consultations where I feel anxious from being presented with too much data to remember, process, organise, and document are often the very ones in which memory is most important.
AVT offered a balm to this cognitive stress and reduced my anxiety. Giving it up was uncomfortable. However, I feel better able to tolerate this discomfort in the knowledge that it is my brain generating memory, which I believe is essential for continuity and safe holistic care.
It is calming to remind myself that I can only be as good as my physical limitations — including those of my brain. This is a scenario where technology cannot enhance my output, just as there is no benefit in setting a gym-machine to lift weights on my behalf. Imperfect though we might be, in the service of holistic knowledge and continuity of care our ability to hold another person’s life in our minds cannot be enhanced by outsourcing to machine memory.
My response will not be the same for everyone. There are many scenarios where factors such as continuity and connection appear less important than efficiency and reproducibility. There is good reason for nuance in the debate around AI and AVT use in our consultations.
However, we should consider the questions it asks us about what is important to us as a profession and as a society. Is reducing the cognitive burden of work worth the cost of memory on our consultations? Do we value our unique input, knowledge, and relationships, or are we simply cogs in a mechanism? Is medicine crafted or mass produced?
In economics, the Baumol effect states that costs in sectors reliant on human input such as health and education increase faster than those where technology brings efficiency savings. Much of the attraction of AVT is due to efficiency. However, this narrative points to some of the risks. The best interpretations of the Baumol effect stress the importance of valuing the unique qualities that human work offers, and funding it appropriately.
By choosing to prioritise values where machines outperform us (access, efficiency, and reproducibility) instead of more human values (connection, care, recognition of suffering, and wisdom), do we risk a future where machines replace us?
