Artificial intelligence in primary care: innovation at a crossroads

Summary

Despite being a cornerstone of health-care delivery, primary care is increasingly under strain. The latest advancements in artificial intelligence (AI) offer new opportunities to transform primary care. However, the rapid deployment of AI ahead of robust real-world evaluation or regulation raises concerns about unintended consequences on the quality of care. We review applications of AI in primary care, covering AI to support primary care providers and people with their health. This Review considers the impact of AI applications on different domains of health-care quality–effectiveness, safety, timeliness, efficiency, patient-centred care, health-care provider experience, equity, and planetary health—and on the primary care-specific attributes of accessibility, comprehensiveness, coordination, and continuity. Implementation of AI in primary care benefits from careful consideration of these quality domains, a focus on universal design principles, digital determinants of health, and AI health literacy, and alignment with patient experiences and values, to support the transformation towards sustainable and high-quality AI-enabled primary care.

Introduction

Primary care is the backbone of health care globally, providing accessible, continuous, comprehensive, and coordinated person-centred care in the community, close to where people live and work. High-quality primary care is known to improve health outcomes, equity, and cost-efficiency in health care.¹ The importance of primary care as a global priority was first established in the 1978 Alma-Ata Declaration and reaffirmed in the 2018 Astana Declaration. The 2023 Global Sustainable Development Report highlights primary care as a key pathway to meeting the UN Sustainable Development Goals.

Search strategy and selection criteria.

We searched MEDLINE (PubMed) and Google Scholar in February, 2025, using a combination of free-text terms and the following controlled vocabulary terms for artificial intelligence (AI; eg, machine learning, deep learning, large language models, and natural language processing) and primary care (eg, primary health care, general practice, family medicine, community health), without date or language restrictions: “Generative Artificial Intelligence” [MeSH], “Artificial Intelligence” [MeSH], “Large Language Models” [MeSH], “Machine Learning” [MeSH], “Machine Learning Algorithms” [MeSH], “Reinforcement Machine Learning” [MeSH], “Unsupervised Machine Learning” [MeSH], “Supervised Machine Learning” [MeSH], “Deep Learning” [MeSH], “Natural Language Processing” [MeSH], “Neural Networks, Computer” [MeSH], “Support Vector Machine” [MeSH], “Speech Recognition Software” [MeSH], “Physicians, Primary Care” [MeSH], “Primary Care Nursing” [MeSH], “Primary Health Care” [MeSH], “General Practitioners” [MeSH], and “General Practice” [MeSH]. Machine translations were planned for any non-English papers that were selected. We also searched the reference lists of relevant articles and highly cited seminal papers known to the authors. One author (LL) screened all the search results while discussing with coauthors as needed to determine relevance and fit. We purposively selected articles that illustrated the breadth and maturity of AI in primary care, prioritising high-quality study designs (including meta-analysis, systematic reviews, and randomised controlled trials, wherever available), recency, and applicability to primary care settings. We included highly cited viewpoint articles, reports, and non-systematic reviews wherever evidence gaps and diverse considerations on the focused topics were identified.

Despite recognition as a global priority, primary care is experiencing substantial challenges, such as workforce shortages and burnout, inadequate funding, poor integration with other levels of care, and increasing demands and complexity of care, hampering the quality of care.2, 3, 4 The COVID-19 pandemic exacerbated these challenges while also accelerating the adoption of digital technologies and virtual delivery of primary care.^3,5 The latest breakthroughs in artificial intelligence (AI) are now uncovering new opportunities in health care (panel 1) and rapidly permeating primary care.¹² In 2024, one in five general practitioners in the UK reported using generative AI in their clinical practice.¹³ Promoting the safe and responsible adoption of digital health innovations and AI in primary care is a key area of focus of The Lancet Commission on Transforming Primary Health Care announced in 2025.⁵

Panel 1. Definitions and health-care applications of artificial intelligence (AI).

AI refers to the ability of computer systems to perform tasks that are usually associated with human intelligence, such as learning, reasoning, problem solving, language understanding, and pattern recognition.⁶ Rule-based AI systems have been around since the 1950s and operate using defined rules and logic programmed by humans. Conversely, machine learning is a subset of AI that involves the capability of algorithms to identify patterns in data (ie, learn from data) and perform automated tasks without explicit programming of every step by a human.⁶ In health care, machine learning is commonly used in risk prediction and clinical decision support by analysing structured data from electronic health records. Deep learning is a subfield of machine learning that uses artificial neural networks with many layers to identify patterns in data.⁶ Deep learning is increasingly being used in medical imaging analysis and natural language processing apps to extract meaningful insights from the text in clinical notes. Natural language processing is the field of AI that focuses on enabling computers to understand, interpret, and produce human language (text or speech) and is commonly applied in chatbots and other conversational AI systems that can engage in human-like conversations.⁷

More recently, the emergence of generative AI has created a technological leap that is disrupting many fields, including health⁸ and health-care delivery. Generative AI is a type of deep learning that can create new content, including text (in the case of large language models [LLMs]), images, audio, video, and data that resemble human-created output. LLMs are trained to understand and generate human language by analysing massive amounts of text data and learning patterns in how words and sentences are used.

Interactions with an LLM typically occur in plain natural language via a conversational AI interface, and the system keeps track of the context and iterations in a chat, resembling human conversation.⁷ Current LLMs can be sensitive to how the prompt (ie, the query or input) is framed and variations in the words, phrases, or form of the prompt can affect the quality and accuracy of LLM-generated responses.^7,9 Prompt engineering is the process of designing and refining prompts to improve LLM performance for a specific task or outcome.

In addition to the general-purpose LLMs commonly available to the public, domain-specific and task-specific solutions focusing on clinical care also exist. LLMs with a focus on health care typically rely on more complex approaches, including retrieval-augmented generation and medical domain fine-tuning. Retrieval-augmented generation is an approach that addresses some of the limitations of LLMs by combining the LLM’s generative abilities with real-time information retrieval from curated data sources, such as clinical guidelines, medical literature, or electronic health record data.¹⁰ Fine-tuning involves retraining of a general LLM using domain-specific datasets (eg, medical question-answering datasets) to optimise performance in that domain, with studies showing that these models can now achieve excellent scores in medical licensing exams.¹¹

In this Review, we assess current and potential applications of AI in primary care, including AI to support primary care providers and AI to support people with their health. We use the term primary care provider to encompass all primary care front-line providers, in addition to the physician (also termed general practitioner), reflecting the multidisciplinary nature of primary care. We will consider the impact of these AI applications on different domains of health-care quality (adapted from the Institute of Medicine’s aims for quality improvement and the Quintuple Aim¹⁴) and on the primary care-specific attributes of accessibility, comprehensiveness, coordination, and continuity¹ (figure). Health-care quality domains considered throughout the Review include effectiveness, safety, timeliness, efficiency, patient-centred care, health-care provider experience, equity, and planetary health. We end the Review by discussing implications for research, policy, and practice.

Figure.

AI in primary health care

AI=artificial intelligence. EHR=electronic health record. PHC=primary health care.

AI to support primary care providers

Several AI tools have emerged in the last decade to support primary care providers with different tasks, including clinical knowledge queries, administrative tasks, clinical care, and leveraging patient-reported outcomes (PROs) and experience measures.

AI for clinical knowledge queries

Large language models (LLMs) are offering new ways for primary care providers to keep up with the evidence and engage with clinical knowledge sources, such as clinical guidelines and scientific publications. In contrast to keyword-based queries typical of search engines, LLMs allow for more precise queries, usually referred to as prompts. Different prompting strategies^7,9 have been proposed to improve the quality of responses from LLMs, which can be leveraged by primary care providers using these models (panel 2). Prompts can also include documents, such as spreadsheets, research papers or guidelines, asking the LLM to answer a specific question based on the input document or requesting a summary, translation to a different language,¹⁶ or a lay language version. Of note, the ability of current LLMs to handle prompts in languages other than English is variable, particularly for clinical knowledge queries. Multilingual LLMs can support some languages but can show low performance for languages and topics less represented in their training corpora.¹⁷ The development of multilingual medical LLMs is a growing area of research.

Panel 2. Strategies and considerations for primary care providers using large language models (LLMs) for knowledge queries.

Different prompting strategies have been shown to improve the responses from existing models. In general, good prompts are direct, include examples, and provide context. Examples of prompting strategies include few-shot prompting (ie, providing examples describing the task), chain-of-thought prompting (ie, including a breakdown of the reasoning for each example), or persona (ie, assigning a role to the LLM, eg, “imagine you are a doctor…”).^9,11 However, prompt engineering is a rapidly evolving field and the performance of different prompting strategies can change each time the models are updated. New models seem better able to handle a broad variety of prompts and can ask for additional information when needed to better tailor their responses.

Considerations for primary care providers using LLMs:

• •The generative nature of LLMs can sometimes lead to factually incorrect but convincing responses, named hallucinations or confabulations.
• •LLMs are not sensitive to the strength of the evidence and can generate responses based on weak evidence or unreputable sources. LLMs rely on publicly available data.
• •LLMs are not updated in real time and might not have access to the latest evidence.
• •LLMs might not be able to correctly report the sources for their responses or the reasoning behind their responses (known as the black box problem). Even when reasoning is provided, the response is subject to the same limitations as any LLM output (eg, hallucinations) and should be interpreted with caution. The inherent limitations of explainable artificial intelligence¹⁵ are even more prominent with LLMs, given that the explanations can provide a strong illusion of a clear box and have a higher potential to elicit trust and lead to biases in decision making, such as confirmation bias and automation bias.
• •LLMs might not communicate uncertainty effectively.
• •LLMs can show sycophantic behaviour and mirror or endorse the user’s thinking, even when the user’s reasoning is flawed or factually incorrect.
• •LLMs can use prompts and input data to train the models, and thus, no sensitive information or patient-identifiable data should be provided to LLMs.⁹
• •The performance of LLMs is variable for non-English languages.

AI for administrative tasks

AI scribes and ambient listening technologies are promising to alleviate some of the documentation and administrative burden borne by clinicians by automatically capturing key content from a consultation and summarising the content into the electronic health record (EHR), drafting notes in real time, which the provider can then review, edit, and approve. Initially working as standalone products, such apps are increasingly being integrated into EHRs and commonly involve LLMs.¹⁸ Studies of AI scribes have shown mixed results for the time spent on documentation and related EHR activities,^19,20 in addition to an increase in the documentation length, but indicate decreased cognitive load and improved satisfaction at work for primary care doctors.¹⁹ However, concerns are being raised about the potential quality degradation of the medical record, effect on provider reasoning, and automation bias leading providers to overtrust AI and failing to detect problems in the generated notes.^21,22

The impact of AI scribes on patient-centred communication and the doctor–patient relationship is also important to consider. Early reports seem to indicate that care recipients appreciate having the provider’s attention focused on them, rather than on the computer (panel 3). AI scribes might foster a return to the doctor–patient dyad, replacing the doctor–computer–patient triad that has characterised primary care consultations since the early 2000s. AI scribes can also be leveraged to improve the doctor–patient relationship and communication with care recipients, for example, by providing feedback on doctors’ communication skills, suggesting ways to improve shared decision making, or generating patient-friendly summaries (panel 3). Conversely, in their efficiency-by-design, AI scribes are also omitting non-clinical social conversations from patient notes, which are key to relationship building in good longitudinal primary care.¹⁸ If AI scribes also leave out relevant biographical information, or patient preferences, values, and beliefs, then the quality of the medical record and care delivery might slowly deteriorate.²¹ Finally, there are worries that time saved from documentation might lead to an increase in the number of care recipients that the primary care providers are expected to see, which could, in turn, worsen provider burnout and patient-centred care.¹⁸

Panel 3. Perspectives on artificial intelligence (AI) scribing from a care recipient and a general practitioner.

Care recipient perspective

David (aged 84 years) is a retired veterinarian living in the USA

“Last week, I visited my health clinic for a check-up on my lungs and had an interesting encounter! First, my doctor asked for permission to record the office visit. Then, she conducted the usual encounter, checking my symptoms, and so on, and I never thought back to all this being recorded. However, I had her whole attention, which struck me as unusual. On my previous visit, the doctor was standing in front of a computer and spent most of the time on the keyboard. At the end of the encounter, she reminded me about the recording and said, ‘Look at this. We are trying out a new system.’ Then, she showed me a transcript of our session, along with a summary of the medical notes focusing on the pertinent information. It was so accurate and free of spelling errors that it was almost spooky! I found it such an improvement in doctor–patient interaction in the office, and the one-to-one was much more like the way it used to be in the old days. Anyway, I am sure this is early days, and there will be improvements and also criticisms or problems to resolve. But this was my first impression from a relatively simple office visit.”

General practitioner perspective

Elizabeth is a general practitioner working in the UK

“I've always considered myself on the slower but more thorough end of the consultation spectrum, and one of the things I wish the National Health Service had considered was some form of universally available dictation software. As my experience grew as a general practitioner, I realised that the most time-consuming aspect was the thorough documentation of notes.”

“I started using [an AI scribe] 4 months ago to document consultations with patient consent. [It] converts the transcript into a consultation note using a preset template I have designed, accomplishing this in seconds. The [summary] can then be easily copied and pasted into the relevant clinical systems—compatible with systems in the UK where I practice. [The AI tool] can also generate a summary of the consultation for patients, if so desired by the patient. Additionally, it can provide feedback on my consulting skills, which can be useful for personal development and appraisal purposes.”

“Key for me is keeping patient details out of the AI system, and I appreciate that [the AI scribe] maintains the dictation function as a free service. While I have found these tools save significant time, I sometimes notice they can insert inaccurate information, necessitating careful review. I would certainly not return to traditional typing methods, as these programmes have significantly expedited my consultation process. I have received positive feedback from colleagues unaware of my use of these programmes, testifying to the quality of notes produced. Looking forward, eliminating the need to write notes represents a substantial advancement in time-constrained clinical scenarios.”

AI is also being proposed as a solution to support primary care providers with other laborious administrative tasks that are key for care coordination and continuity, such as responding to patient electronic messages^23,24 and summarising clinical information to write medical reports and referral letters.²⁵ LLMs can facilitate the process of translating and simplifying medical content and help providers to quickly generate plain language summaries of clinical information for care recipients, such as discharge summary notes and action plans at the end of a consultation, to aid patient comprehension and self-management.²⁶ In addition, in several health-care organisations across the USA, LLMs are being used in EHRs and to automatically draft responses to patient messages. Initial evaluations of these deployments found no evidence of savings in terms of time with LLM-drafted replies^23,24 despite the primary care providers’ perceptions of increased efficiency and reduced cognitive load.²⁷ For care recipients, a study has shown preference for AI-drafted messages and a slight decrease in satisfaction when these messages were accompanied by a statement of the involvement of AI in message drafting, raising ethical questions about the need for AI disclosure in care.²⁸ Importantly, studies assessing AI-generated patient summaries and responses have indicated mixed results for content quality, some including omissions or hallucinations that could lead to severe patient harm.^24,27,29,30 Real-world application of these tools to deliver clinical information and advice to care recipients requires careful provider oversight to ensure patient safety.

AI for clinical care

To date, primary care applications of AI for clinical care that have been implemented or evaluated in randomised trials are mostly pre-generative AI non-LLM solutions comprising deep learning image analysis (eg, diabetic retinopathy screening) and EHR-based machine learning algorithms. Diabetic retinopathy screening was one of the first successful applications of AI in clinical care, enhancing the availability of specialist-level testing in the community and primary care settings, particularly in low-income and middle-income countries, with impact shown across several trials.^31,32 In contrast, robust trial evaluation of EHR-based machine learning algorithms remains scarce despite the plethora of in silico studies reporting promising performance results.^33,34 Few trials in primary care have shown mixed results: a trial evaluating an EHR-based machine learning algorithm to detect low ejection fraction in electrocardiograms showed a significant increase in diagnoses,³⁵ whereas two other trials did not show significant results on atrial fibrillation diagnosis³⁶ and treatment optimisation.³⁷ These trials highlight the challenges with translating algorithmic performance into real-world impact, wherein numerous factors can influence implementation success (eg, heterogeneity in clinician–AI collaboration).³⁸ Other promising AI algorithmic interventions, such as automated proactive outreach and digital biomarkers derived from remote patient monitoring data, currently do not have robust trial evaluation and face the same implementation challenges before they can be integrated into clinical decision making by primary care providers.³⁴

Assessment of an algorithm’s suitability for implementation in the primary care context is key, as factors related to different phases of the algorithm’s lifecycle can influence real-world translation. For instance, machine learning models trained with datasets from secondary-care and tertiary-care settings might not work well in primary care. Training with data non-representative of the patient population and setting in which the algorithm is intended to be implemented can lead to biases such as the spectrum effect (ie, performance variation of a diagnostic or predictive test across different patient populations based on differences in the severity, stage, or prevalence of disease).³⁹ However, a systematic review of AI algorithms aimed for early detection of skin cancer in primary care showed that only two of 272 included studies used training data from clinical settings with a low prevalence of skin cancers, akin to a primary care setting.⁴⁰ Additional issues when evaluating an algorithm’s suitability for primary care include unavailable key data, such as information on development and validation (eg, setting and sample characteristics). A systematic review of predictive algorithms for primary care showed scarce publicly available evidence on key quality criteria, particularly for commercial algorithms with regulatory approval.³³ Transparency in all stages of an algorithm’s lifecycle is important to enable adequate assessment and implementation success.

Generative AI is also showing promise in supporting clinical care tasks and showing expert-level performance in medical question answering¹¹ and clinical management reasoning tasks.⁴¹ However, to date, evidence of good performance in responses to clinical questions has mostly been from studies in simulated conditions, whereas performance in the real world can vary substantially depending on the format of the question, user characteristics, clinician–AI collaboration, and other contextual factors.41, 42, 43, 44 Although AI can excel on well defined structured tasks, primary care operates within a fundamentally different reality—one shaped by the complex interplay of multimorbidity, social determinants of health, evolving patient preferences and values, and the persistent challenge of navigating clinical uncertainty. In addition, the complexity of integrating multiple types of data into clinical practice (eg, imaging, patient records, and doctor–patient dialogues) has been difficult to handle by AI to date and is a challenge that multimodal AI is promising to address.⁴³ For now, defining guardrails for safe clinical use of generative AI, in addition to mitigation strategies that address issues such as automation bias, is important.⁴²

AI for leveraging PROs and patient experience

AI could transform the way primary care services and providers collect and use PROs and patient experience to improve care quality, both for individuals and whole populations. At an individual level, incorporation of PROs into clinical decision-support systems can enable decisions and care to be tailored on the basis of outcomes that matter for care recipients, improving patient-centred care and equity for groups that have historically encountered bias and discrimination in the care they receive.⁴⁵ However, the use of PROs as an input variable in AI interventions is rare.⁴⁶ Improving the use of PROs and patient-reported experience measures (PREMs) in clinical care and AI interventions could be facilitated by streamlining collection of these data through conversational AI interfaces. Conversational AI can interact with people in their preferred language, using text or voice, and gather their responses to PROs and PREMs questionnaires at different times in their care journey.

At the population level, systematic collection, analysis, and use of PROs and PREMs can highlight gaps in care quality, identify changing population needs, facilitate cross-comparison and learning between different locations, and inform quality-improvement programmes and system-wide change towards high-quality primary care.⁴⁷ For instance, AI could also be leveraged to analyse patient feedback that is collected in free-text formats. Patient feedback is especially important in primary care, given that primary care is often the first point of contact for individuals with the health-care system. Feedback helps to identify gaps in communication, care coordination, continuity, accessibility, patient centredness, and overall quality of care that could negatively affect patient trust, adherence, and long-term engagement with care. By capturing patient voices, primary care practices can refine processes, tailor services to community needs, address social determinants of health, and enhance relationships between individuals and providers. AI can facilitate the analysis and classification of patient feedback into different topics, which can then be used to implement quality-improvement programmes and improve patient experience (panel 4).

Panel 4. Example of artificial intelligence analysis of patient feedback for quality improvement (fictionalised).

Anglesea Road surgery is a large general practice in Dublin, with 50 000 registered care recipients. The care recipients are encouraged to leave reviews of the practice on the website. Once a month, the practice manager uploads the last month’s worth of reviews onto ChatGPT and asks the app to identify themes. This process has allowed the practice to identify that parking has been difficult for older individuals and to increase the number of reserved accessible parking bays in the car park. The practice also identified from the reviews that care recipients were struggling to get through by telephone in the mornings and, therefore, has reconfigured the receptionist rotas to meet this need.

AI to support people with their health

Patient-facing AI tools are becoming more prevalent, both within the context of health-care delivery and also as direct-to-consumer apps. AI offers great potential in democratising health information and empowering people to play a more active role in their health and health care. Conversational AI might be able to better engage care recipients and support lifestyle behaviour change, chronic disease self-management, and mental wellbeing—an area in which trial evidence is still emerging.^48,49 Increasingly, AI can be leveraged to personalise health information on the basis of patient needs, language preferences, culture, setting, and level of health literacy.^16,50,51 Early studies suggest that AI can display good levels of perceived empathy, personalisation, and language simplification when used for patient communication and information.^30,51 AI can also be used to better equip individuals for shared decision making and personalise patient decision aids with individual risk–benefit information and integration of patient values and preferences.^52,53 In the near future, patient-friendly summaries of medical reports, discharge summaries,^26,30 and information in EHRs (eg, consultation notes) could become standard of care in AI-enabled patient-centred health-care systems.

With increasing shortages of primary care providers worldwide, health-care organisations and governments are starting to turn to AI to address the rising care demands. For instance, digital symptom checkers (involving AI to varying extents) are being used to enable individuals to input their symptoms and other health data to obtain a list of potential diagnoses and triage advice.⁵⁴ Different countries are now including symptom checkers as part of their virtual front-door approaches, to serve as the first point of contact to individuals in the health-care system, without provider oversight. However, studies evaluating the diagnostic and triage accuracy of the available symptom checkers have shown variably low to moderate performance, leading to potential over-triage and under-triage.⁵⁴ Although over-triage can cause unnecessary use of health-care services and overburden the system, under-triage can have catastrophic consequences in the case of life-threatening conditions.

The rise in direct-to-consumer AI health apps⁵⁵ and system-wide deployment of patient-facing AI tools in the current landscape of insufficient evaluation and regulatory oversight raise important concerns about patient safety.⁵⁶ These risks are particularly high with LLMs, as their human-like conversational abilities and anthropomorphic features can lead people to overtrust and over-rely on incorrect or incomplete AI-generated health advice.^57,58 Collaborative partnerships between industry, academia, health care, and the public are needed to ensure patient safety and support the rigorous and transparent evaluation of these systems in a timely manner, in addition to real-world testing and post-deployment monitoring. Primary care providers play an important role in educating care recipients and the community about safe use of AI tools for health purposes, particularly those that do not have rigorous evaluation.

Equity

AI and other digital technologies can affect health equity in both positive and negative ways by narrowing or widening disparities among socioeconomic and demographic groups in terms of health-care access, health status, and health outcomes.59, 60, 61, 62 Ensuring inclusiveness and equity is a core principle of WHO for governing AI. Bias considerations and mitigation strategies are crucial for preventing AI-related inequities that can arise from algorithmic bias and from real-world implementation of AI systems.^60,62

Despite its futuristic allure, AI is trained on data from the past, embedding biases that threaten to perpetuate inequities in health care. The datasets used to train and test an AI algorithm can be a source of bias when they are unrepresentative of the intended population in which the AI system will be used. A large proportion of the medical datasets used for AI development historically over-represent individuals of European ancestry, which can lead to worse outcomes for the under-represented populations and could perpetuate systemic discrimination.⁶³ For example, a systematic review of AI for the early detection of skin cancer in primary care found that only a small proportion of the studies included skin lesions from darker skin colours.⁶³ A related issue is the variation in the collection, classification, and stringency of race, ethnicity, and other protected attribute data across different geographical regions, which complicates their inclusion as covariates.⁴⁰ As a result, a WHO publication emphasises the importance of AI products being trained on the intended population for which the AI system will be used and retrained when intended to be applied to a different population.

Bias in AI is not just a data problem but can also be present in the assumptions made during algorithm development and evaluation (ie, algorithmic bias). For example, when providers are used as the reference standard for the training dataset, their potential bias in diagnosis can persist as bias in the AI system (eg, women with autism being misdiagnosed with other mental health conditions).⁶⁰ In addition, AI has been shown to be able to predict self-reported race from medical images, often when clinical experts cannot, raising concerns about deployment of models without appropriate strategies to address bias and ensure fairness.⁶⁴ A potential approach to mitigate this issue involves better aligning algorithms with patient experiences and outcomes, rather than the way individuals are treated by the health-care system.⁴⁵ Indeed, an algorithm trained to predict the knee pain reported by the individual, rather than the x-ray interpretation of the doctor, doubled the proportion of Black individuals who were eligible for knee replacement.^45,65 Importantly, robustly evaluating how algorithms perform in various populations through data stratification, subgroup and sensitivity analyses, and by assessing the effects of pre-processing, in-processing, and post-processing techniques in mitigating bias (eg, weighting) is also key.

Aside from biased datasets and algorithms, potential latent biases can also arise when AI is implemented in complex sociotechnical systems, such as health care. Numerous factors, including contextual, provider-level, and patient-level factors (eg, trust in AI), will influence how AI tools are adopted and used, and whether biases emerge at the implementation stage.⁶² At the context level, lower-resource settings might face constraints in the adoption and implementation of AI, leading to inequities in care provision, as compared with those observed in higher-resource settings. Providers might also perpetuate biases when using AI, for example, by applying AI decision-support recommendations unequally between different groups. At the individual level, inequities could be exacerbated by the digital divide and whenever individuals’ capabilities, needs, characteristics, and structural and lived realities deviate from the assumptions embedded in the design of the technology. Development and implementation of AI without careful consideration of latent biases and the social and economic determinants of health might unintendedly increase health inequities and exacerbate the inverse care law, whereby the availability of good medical or social care can vary inversely with the needs of the population served.⁶⁶

Planetary health

Human health and planetary health are deeply intertwined. The greatest gains to human health have come from addressing environmental challenges through simple measures such as the provision of clean air, clean water, and better housing. However, AI’s high consumption of energy and water is putting some of these gains at risk. Estimates of AI’s contribution to greenhouse gas emissions are controversial,⁶⁷ but existing evidence points to concerning trends in water consumption and energy use.⁶⁸ Some estimates indicate that the carbon emissions of developing GPT-3 corresponded to 188 air trips from New York City, NY, USA to San Francisco, CA, USA and that an interaction with an LLM can consume ten times more energy than a standard search engine query.⁶⁹ Data centres are currently using around 1% of global electricity but wide geographical variation exists, with data centres in Ireland consuming more than 20% of the national electricity supply.⁶⁹

Paradoxically, AI also has the potential to support environmental sustainability efforts,⁷⁰ including in health care. AI-enhanced primary care might be particularly well positioned to promote environmentally sustainable care delivery⁷¹ through enhanced health promotion and prevention efforts at the level of the individual, family, and community, improving the health of populations and reducing the demand for hospital care. Primary care providers are also experts in quaternary prevention (ie, prevention of overmedicalisation and overtreatment), reducing unnecessary or inappropriate care, and therefore, avoiding emissions associated with the provision of these services. Finally, primary care’s proximity to where people live and involvement of multiprofessional teams enables delivery of multidisciplinary care with low travel requirements for individuals, as do virtual and AI-enabled options for primary care delivery, which have become more common of late.

Implications for research, policy, and practice

AI deployment in primary care is moving ahead of evaluation and regulation. This rapid pace offers opportunities for faster translation of innovation into real-world benefits for people and care quality, bypassing traditional challenges in digital health implementation. Meanwhile, short stepping evaluation and regulation might have unintended consequences for patient safety, equity, and overall care quality. New regulatory paradigms will most likely be needed as existing pathways are not well equipped to address the fast-evolving nature of AI.^33,72,73 Robust evaluation processes also need to be streamlined, with continuous monitoring during and after deployment of clinical AI tools,^8,42,74 and assessment of different strategies to address automation bias (eg, clinician training and user interface design), considering limitations in explainable AI.¹⁶ Importantly, evaluation studies should consider applicability of the AI system to the primary care context and report their results accordingly, for example, by utilising the Consensus Reporting Items for Studies in Primary Care (CRISP) Statement. Ultimately, multisectoral partnerships and collaborations between primary care providers and professional organisations, care recipients and the general population, technology experts, researchers, industry, government, and other stakeholders will be key in advancing responsible implementation of AI in primary care.

An urgent need exists now to equip providers and the general population with the knowledge and skills required to navigate the AI shifts in primary care. Primary care providers need adequate training and support to appropriately integrate AI tools into clinical practice, ensuring that these technologies do not put additional burden on an already overstretched workforce.³⁹ Education and training programmes need to quickly adapt to equip current and future clinicians with the necessary skills to use AI in a safe, ethical, and effective manner as part of patient-centred care delivery.⁴² In settings with restricted access to care, such as in low-income and middle-income countries, AI might be increasingly seen as an alternative for some primary care needs.⁷⁵ As people turn to AI when facing gaps in care access, improving digital AI health literacy at the population level becomes more important than ever. AI-empowered people will be better able to manage their health and actively seek, and advocate for, high-quality care that aligns with their preferences and values. Moving forward, ensuring free and easy access to health care-curated LLMs that provide evidence-based medical and health information with adequate guardrails and safeguards might be a safer alternative to the growing use of general-purpose LLMs by clinicians and the general population.

Conclusions

The global need for primary care transformation and the latest technological leap in AI are heralding an opportunity to reimagine care delivery. Primary care providers are progressively adopting AI to streamline administrative processes and support their clinical reasoning and knowledge needs. During the consultation, AI is facilitating a return to the doctor–patient dyad, reducing interference from the computer and improving patient-centred care. People in the general population are also turning to AI for support with their unmet care needs, as access to primary care worsens worldwide. Improving AI health literacy has become a global necessity to help people to navigate the benefits and risks of AI in health care. Ensuring equity in an AI-enabled primary care system will require consideration of universal design principles, digital determinants of health, safety nets, and careful alignment with patient experiences and values across the full spectrum of human diversity. Importantly, the current focus on AI applications at the point of care should not distract from the larger-scale interventions that are needed to improve patient outcomes and overall care quality in primary care. Amid the rising global burden of non-communicable diseases, leveraging AI to enhance equitable and scalable preventive health-care approaches in the community could be a pathway to accelerate the transformation towards sustainable and high-quality primary care.

Declaration of interests

We declare no competing interests.