The last decade has witnessed major advances in the development and application of artificial intelligence (AI) technologies. Coined by John McCarthy in 1955, AI was originally described as the “science and engineering of making intelligent machines, especially intelligent computer programs” [1]. It is already featured in many aspects of modern life, including the use of smart home devices, search engines, and facial recognition software. Indeed, given its potential to improve efficiency and productivity across a range of industries and public sectors, AI has been regarded as one of the most important new technologies of the century for policy makers, governments, and citizens alike [2].

In the context of health, it has been claimed that AI has the potential to revolutionize health care by offering improved diagnostic accuracy, reducing costs, and enabling a more personalized approach to medicine [3,4]. Currently, there are a limited number of AI applications that have been approved for clinical use in Europe and the United States, with most of these being in the field of medical imaging [5]. In 2021, the first AI-based pathology software was approved by the US Food and Drug Administration to assist with the detection of prostate cancer [6]. This milestone was indicative of the transformational shift that has occurred in pathology in the last few decades, a shift that has seen a move away from the traditional microscopic review of a glass slide toward the use of state-of-the-art digital scanners to generate a high-resolution image that is wholly representative of the original tissue section and of comparable diagnostic quality [7]. This digital approach offers many advantages in terms of improved efficiency and productivity in clinical laboratory workflows, but it is also the availability of complex visual information in these digitally generated images that has provided a rich source of data for the development of novel pathology AI algorithms to potentially aid with diagnostic and analytical procedures [8]. Many studies have since used these digitized slide images to train computer algorithms for aspects of pathology practice, including automated identification of lymph node metastases [9], prediction of prognosis [10], disease grading and classification [11,12], and identification of clinically actionable targets (biomarkers) for cancer treatment [13], with many studies claiming comparable or superior performance of AI to that of human intelligence [9-12].

The potential of these AI tools in terms of improved efficiency, cost-effectiveness, and accuracy of diagnosis cannot be underestimated. However, in addition to regulatory considerations, several challenges must be addressed in the development and implementation of such technologies in diagnostic pathology practice. One of these is public support for regulated access to health care datasets for AI research. The development of machine learning and deep learning approaches in pathology not only requires access to large numbers of digitized whole slide images (WSIs) but also requires the generation and aggregation of vast amounts of anonymized patient data, including but not limited to data from health care records and investigative tests. Therefore, the public must have trust and confidence that their data will be used safely, securely, and appropriately [14]. While safeguards are in place to govern the sharing of data [15], it is well documented that corporate profiteering and data privacy and security are consistently among the public’s top concerns regarding the sharing and use of health data in research [16-18]. High-profile cases such as the one involving the AI application Google DeepMind, which was granted “legally inappropriate” access to National Health Service (NHS) data, have tested public confidence in how the NHS safeguards access to data for secondary use [18]. Furthermore, evidence indicates that public support for the use of health data in AI research is conditional on the potential benefit in return [19]. This is important to note given that there appears to be a general lack of public knowledge and awareness of data science and AI [20], which could result in a failure to see the benefit and value of its application in a health care setting such as diagnostic pathology.

Second, while research in pathology algorithm development uses data in the form of digitized WSIs, the WSI must still be generated from a human tissue sample. In the United Kingdom, the use of human tissue samples for research is governed by the Human Tissue Act of 2004 [21]. This provides a regulatory framework that allows deidentified residual clinical material (eg, formalin-fixed paraffin-embedded tissue blocks residing in diagnostic archives and surplus to further clinical requirement) to be used for research purposes without explicit consent so long as the sample is rendered anonymous to the end user [21]. This position is in line with public consensus supporting the use of tissue samples in research [22-25] and particularly in cancer research [25]. However, we do know that various factors influence public attitudes toward the use of their samples in research. These include sociodemographics, health status, cultural and religious values, awareness, knowledge, past experiences, beliefs about the potential benefits and risks, attitudes toward genetic and medical research, experience with the health care system, and trust in government and scientific institutions [22-27]. These factors have yet to be explored in the context of the use of tissue and linked data in AI research.

Finally, there needs to be public support for the use of AI in their own health care, particularly in disciplines such as pathology, where it is predicted that AI will significantly transform practice [28]. One UK study surveyed men who had undergone a prostate biopsy (N=1276) and reported that respondents were mostly supportive of AI as a diagnostic companion to professional decision-making (1058/1276, 83%) [29]. However, more widely, there have been documented concerns from the public about an overreliance on AI in health care at the expense of human involvement [30,31] and uncertainty about the delegation of clinical responsibilities to machines and systems [32,33]. A report from the Academy of Medical Royal Colleges recommended early and ongoing public engagement in the development and implementation of AI to maximize impact and ensure that the public has their say when examining the acceptability of such technologies [34]. Currently, there is a dearth of evidence on which to draw such conclusions in pathology practice as there has been limited investigation of this to date beyond the male cohort surveyed by Rakovic et al [29]. To this end, there is a need to examine the extent of public awareness of AI and their views on its potential use in diagnostic pathology to gauge the level of support and identify any concerns, ultimately fostering informed public endorsement of the integration of AI applications in pathology and the anticipated benefits in terms of efficiency and accuracy. Therefore, the aim of this study was to investigate public awareness of and attitudes toward the use of AI in pathology research and practice with the following objectives: (1) to assess public awareness and attitudes regarding the development of AI technologies in pathology; (2) to assess public awareness and attitudes regarding the implementation of AI technologies in pathology; (3) to identify matters of concern and sources of support regarding the design and implementation of AI technologies in pathology; and (4) to determine whether sociodemographic characteristics, experiences with cancer, awareness, beliefs, and attitudes influence support for the development and implementation of AI in pathology.

A nationally representative cross-sectional, web-based mixed methods survey was conducted to (1) assess the UK public’s awareness of and views on the use of AI in diagnostic pathology and (2) ascertain levels and predictors of support for the use of AI in diagnostic pathology and the use of deidentified personal health data and images of tissue in the development of AI technologies for cancer diagnosis and management.

The survey was administered on the web (November 30, 2022), with respondents recruited via an online crowdsourcing platform managed by Prolific Academic Ltd [35] to provide a nationally representative sample of approximately 1500 UK respondents stratified across 3 demographics: age, gender, and ethnicity. To be eligible for this study, respondents had to reside in the United Kingdom, be aged >18 years, and have the capacity to express their own opinion.

The survey was created and hosted on Qualtrics (Qualtrics International Inc), a web-based software to assist with survey management. The web-based survey consisted of 30 closed questions, 2 open questions, 2 attention checks, 2 screener validation questions (age and country of residence), and 1 CAPTCHA challenge (Multimedia Appendix 1). The survey was developed by the research team with input from a patient and public involvement panel and pilot-tested with individuals external to the research team (n=5). To ensure that the questions were easy to understand, lay terms such as cancer diagnosis and management were used in place of medical terminology. In the first section of the survey, sociodemographic information and information related to the respondents’ experience with cancer were collected. The remaining sections of the survey captured both quantitative and qualitative data on public views on the use of AI in cancer diagnosis and management and the use of data (deidentified personal health data and images of tissue) in the development of AI technologies for diagnostic pathology. Unless otherwise stated, responses to the closed questions were rated on a 5-point Likert scale measuring agreement (1=strongly disagree; 5=strongly agree).

To assess awareness, respondents were asked to rate their level of awareness of AI being developed for use in cancer diagnosis and treatment planning on a 5-point Likert scale (not at all to extremely). For beliefs about the potential impact of AI, respondents rated their level of agreement with the following 4 statements: “AI will help improve the diagnosis of cancer and treatment planning,” “AI will help reduce the risk of medical error or misdiagnosis,” “AI will result in less harm for patients,” and “AI will improve efficiency and performance in the NHS.” For risks and benefits, respondents could select any or all that applied from a list of potential risks (harm, error or misdiagnosis, lack of oversight or regulation, lack of accountability, loss of privacy, lack of personal interaction, or none) and benefits (more accurate decision-making, more efficiency, or money savings for the NHS; less risk of harm; less risk of error or misdiagnosis; more time freed up for staff to work on other tasks; or none) of using AI in cancer diagnosis and management.

To measure support for the implementation of AI in diagnostic pathology, respondents rated their level of agreement with the following 6 statements: “I support a diagnosis of cancer being made or ruled out by AI only,” “I support a diagnosis of cancer being made or ruled out by a pathologist with the assistance of AI,” “I support a diagnosis of cancer being made or ruled out by a pathologist only with no AI input,” “I support the diagnosis of biological ‘markers’ which can inform cancer treatment pathways by AI only,” “I support the diagnosis of biological ‘markers’ which can inform cancer treatment pathways being made by a pathologist with the assistance of AI,” and “I support the diagnosis of biological ‘markers’ which can inform cancer treatment pathways being made by a pathologist only with no AI input.”

For support for the development of data-driven AI technologies for pathology, respondents rated their level of agreement with the following 4 statements: “I support my deidentified personal health data being used in the development of AI technologies to diagnose and manage cancer by public bodies outside the NHS (eg, universities),” “I support my deidentified personal health data being used in the development of AI technologies to diagnose and manage cancer by private commercial research organizations outside the NHS,” “I support deidentified images of my tissue being used by public bodies outside the NHS (eg, universities),” and “I support deidentified images of my tissue being used by private commercial research organizations outside the NHS.”

To assess trust in data security, respondents rated their level of agreement with the following 6 statements: “I trust that my deidentified personal health data shared with public bodies outside the NHS (eg, universities) will not be used for purposes other than the development of AI technologies to diagnose and manage cancer,” “I trust that my deidentified personal health data shared with private commercial research organizations outside the NHS will not be used for purposes other than the development of AI technologies to diagnose and manage cancer,” “I trust that the NHS have effective security and privacy protocols for the sharing of personal health care data that will protect my anonymity,” “I trust that deidentified images of my tissue shared with public bodies outside the NHS (eg, universities) will not be used for purposes other than the development of AI technologies to diagnose and manage cancer,” “I trust that deidentified images of my tissue shared with private commercial research organizations outside the NHS will not be used for purposes other than the development of AI technologies to diagnose and manage cancer,” and “I trust that the NHS have effective security and privacy protocols for the sharing of image data that will protect my anonymity.”

Qualitative data were obtained from responses to the following open-ended questions: (1) “AI has the potential to improve the diagnosis and management of cancer, however patients and the public may have concerns about the development and application of AI in this context. What do you think would help yourself or others to support the use of AI technologies in cancer diagnosis and management? Please explain”; and (2) “AI has the potential to improve the diagnosis and management of cancer, however patients and the public may have concerns about the development and application of AI in this context. What do you think would help yourself or others to support the use of deidentified images of tissue and personal health data in the development of AI technologies to diagnose and manage cancer? Please explain.”

Ethics approval for this study was granted by the Queen’s University Belfast Faculty of Medicine, Health, and Life Sciences Research Ethics Committee (reference: MHLS 22_107). Relevant information such as background of the study, aims and objectives, and specific instructions about how to take part in the survey and what would be involved (eg, length of the survey and completion time) was clearly described on the front page of the survey website with a link to a detailed participant information sheet. Respondents were asked to indicate their consent to participate by using a checkbox. The survey was anonymous, no identifiable information was collected, and all data were released to the research team using a unique ID allocated by Qualtrics. Participants provided their age in years, and all other sociodemographic information was collected at a broad categorical level. A nominal monetary incentive was provided to respondents via Prolific in line with standardized rates (£2.80 [US $3.59]).

An initial total of 1665 responses were received. One response was removed for not meeting eligibility criteria (ie, not a resident of the United Kingdom). In total, 0.54% (9/1665) of the responses were removed for not providing consent. The attrition rate was <1% (ie, 15/1665, 0.9% of the responses were removed for completing <50% of the survey items). The median completion time for this sample was 9.88 (IQR 5.84) minutes. A total of 3.06% (51/1665) of the responses were removed for having a completion time of less than half the median (ie, <4.93 min). In total, 1.92% (32/1665) of the responses were removed for failing 1 of 2 attention checks, and 1.74% (29/1665) were removed for evidencing straight-lining behavior on 2 matrix questions, indicative of low effort in responding. The final number of respondents in the dataset was 1518.

Age ranged from 18 to 82 years (mean 46.12, SD 15.68 years). A total of 51.32% (779/1518) of the sample were women; 47.63% (723/1518) were men; and 1.05% (16/1518) of the respondents identified as nonbinary, third gender, or other (eg, genderqueer) or preferred not to say. Most respondents were White (1328/1518, 87.48%), resided in England (1291/1518, 85.05%), were educated at the graduate level (992/1518, 65.35%), and had no religion (855/1518, 56.32%) or were Christian (537/1518, 35.38%).

When asked to describe their personal circumstances in relation to cancer, most respondents (910/1518, 59.95%) selected 1 response option, and 39.59% (601/1518) selected between 2 and 4 responses. A large proportion of respondents had experience of a close person or family member being diagnosed with cancer (617/1518, 40.65%), and 43.68% (663/1518) had experienced the death of a close person or family member to cancer. In total, 6.79% (103/1518) had been diagnosed with cancer themselves. A total of 4.68% (71/1518) of the respondents worked with people diagnosed with cancer (on a professional or voluntary basis), and 9.09% (138/1518) were health and social care professionals or academics with an interest in the subject of cancer.

This study sought to examine the UK public’s awareness of and attitudes toward the use of AI in pathology research and practice and ascertain predictors of support for its implementation in clinical practice and for the use of health data and tissue images in its development. We found that awareness of AI being developed for use in pathology was low, yet respondents were conditionally supportive of its use in the diagnosis and management of cancer if human oversight was maintained. Respondents were also supportive of their data and tissue images being used in the development of such technologies by public bodies; however, there was less support for use by commercial bodies. To our knowledge, this is the only nationally representative study on public opinion on AI in pathology, a field in which AI has been regarded as the “third revolution” [37].

Awareness of AI being developed for use in pathology was very low, which was not overly surprising. While awareness of AI in general is increasing, a 2023 nationally representative survey of public attitudes toward AI in Great Britain reported that people were less aware of the use of AI in health compared to the use of AI in other applications such as facial recognition, driverless cars, and targeted consumer advertising [31]. In this study, qualitative data indicated a broad appetite for increased awareness and understanding of how AI systems could be used in cancer diagnosis and management. Suggestions from respondents on how to deliver this included social media campaigns and information disseminated via leaflets, videos, and television media using real-world examples of how AI can benefit their care. Despite low awareness, the respondents in this study held positive beliefs about the potential of AI to improve patient outcomes and health care system efficiency. Accordingly, the most frequently identified benefits of AI in pathology practice were maximization of NHS resources and more accurate decision-making. However, the most frequently identified risks were error, lack of accountability, and lack of oversight. These risks and benefits of AI in pathology are in keeping with those reported in studies investigating the use of AI in other medical specialties such as radiology [38] and dermatology [39]. Almost all respondents supported a diagnosis of cancer or an assessment of biomarkers being made by a pathologist with the assistance of AI, but there was no widespread support for diagnosis using AI only. This was also evident from the qualitative analysis, with many respondents indicating that some form of human oversight would help increase their support for the use of AI in pathology practice. This position is consistent with previous literature suggesting that the acceptance of AI in health care is overwhelmingly based on its use as a tool to support clinician decision-making rather than as a replacement [30,40].

There was strong support for the use of deidentified personal health data and images of tissue in the development of AI technologies for pathology by public bodies. This is reassuring given the potential impact of controversial failed NHS data-sharing projects such as care.data [41] and the General Practice Data for Planning and Research program [42] on public opinion. There was slightly less support for the use of health data and images by private bodies (vs public bodies), confirming previous research findings indicating a level of concern among the public when it comes to private or commercial organizations obtaining access to health data [16,43]. Although there were high levels of trust in NHS security and privacy protocols, there was some concern that data would be used for other purposes by both public and commercial organizations, with qualitative analysis highlighting that transparency and openness in how data are shared and governed would help increase support for the use of data in the development of AI health systems. As outlined in a recent UK Government review into the use of health care data in research [44], our findings reinforce the need for candid and continuous engagement with the public about data sharing, particularly when this involves commercial organizations.

Identifying more benefits of AI in pathology practice increased the odds of supporting the development of AI in pathology research, whereas identifying more risks decreased the odds of supporting implementation of AI in pathology practice. Greater awareness of and more positive beliefs about AI increased the likelihood of supporting implementation, and men were more likely to support the use of AI in cancer diagnosis and management relative to women. This is a finding that is common to the wider AI literature, in which women appear to be more cautious in their opinions on AI in health care than men. Yakar et al [45] surveyed 2411 members of the public in the Netherlands in 2022 regarding their views on AI in medicine with specific emphasis on dermatology, radiology, and robotic surgery and reported more trust in AI among their male respondents than among their female respondents. Similarly, a large web-based survey (N=11,004) of Americans’ views on AI in health and medicine conducted by the Pew Research Center in 2022 concluded that men were more open to the use of AI in their health care than women [46]. Although our data cannot explain why women or people from a non-White background are less likely to support AI in pathology, our findings correspond with the existing evidence of a notable lack of support for AI in health care among these populations. Given the evidence of gender inequities in the ability of AI tools to detect liver disease [47] and the concern that datasets used to train AI algorithms will exacerbate ethnic health care inequalities if they are ethnically unbalanced or biased [48], it is imperative that those involved in the development and implementation of AI technologies in pathology engage meaningfully with these populations to better understand their concerns.

Fundamentally, AI has the potential to revolutionize health care. Although the number of AI systems currently being used in health care in the United Kingdom remains relatively low, this number will undoubtedly increase within the next decade given the rate of their development along with committed investment from the government to embrace and deploy AI technologies in the NHS [49]. However, the acceptability of AI and, ultimately, its success will depend on public confidence and trust in its development and implementation [50]. While the Academy of Medical Royal Colleges recommends early and ongoing public engagement to maximize the impact and acceptability of AI in health care [34], our findings suggest that aspects of this may have been left wanting to date, particularly in the field of pathology, where AI technologies are being developed at such a rapid pace.

Presently, the United Kingdom lacks specific legislation governing the use of AI, opting instead for a “proportionate and light touch” framework that will provide a set of principles for regulatory authorities to interpret and implement according to the requirements of their own sector [51]. This approach, deemed proinnovative, aligns with the rapidly evolving AI landscape. However, our research underscores that public support for the use of AI in pathology is contingent on maintaining human oversight in clinical decision-making; therefore, regulatory bodies and stakeholders who govern the application of AI technologies not only in pathology but also in the wider health care milieu recognize that any shift from this position may undermine public confidence and trust. To ensure support for the use of AI in pathology, stakeholders should prioritize initiatives fostering awareness and promoting positive beliefs about AI. Awareness campaigns should adopt a balanced perspective that acknowledges potential risks while concurrently emphasizing the benefits. Stakeholders must inform the public of the evidence supporting the use of AI in health care while reassuring the public regarding robust data security measures in place in its development. Simultaneously, efforts should be made to bridge gender disparities in support, and specific strategies should be developed to address the lower support among individuals from diverse backgrounds, ensuring inclusivity in AI discussions and decision-making. Through a balanced approach encompassing transparency, engagement, and targeted education, policy makers can cultivate a supportive and informed public stance regarding the responsible implementation and development of AI in health care, aligning with the evolving regulatory landscape.

This study is novel in that it is the first to examine public awareness of and opinions on the use of AI in pathology research and practice with a nationally representative UK sample. However, there are several limitations to acknowledge. First, selection bias was a possible limitation in this study due to the web-based nature of the survey as those without an internet connection or some level of digital literacy had restricted access. Second, a known limitation of using a panel survey is the risk to data integrity due to inattentive responses [52]. Although this risk was mitigated in this study by using attention checks, screener validation questions, and a Completely Automated Public Turing Test to Tell Computers and Humans Apart, along with the removal of responses indicative of low effort in responding, there remains potential for low-quality responses due to the uncontrolled nature of the survey. Third, it was also noted that a large proportion (992/1518, 65.35%) of the respondents in the survey were educated at the graduate level; this may have implications for the findings given claims that attitudes on societal and political issues change incrementally with educational level [53]. Finally, the survey questions used in this study were developed by the research team based on the study objectives and the existing literature. However, to the best of our knowledge, no validated standardized questionnaire on the development and implementation of AI in pathology or even health care was available when this study was conducted.

Future work to validate a tool to longitudinally assess public knowledge of and trust in AI in health care would be of value, in line with AI being implemented more widely in many aspects of health care delivery. There is also a need to further explore the gender and ethnic disparities that were uncovered in this study to recognize the specific concerns of women and those from a non-White background. Given the suggested variances between countries in their level of trust in AI in general [54], further research representative of wider geographical areas would be of merit.

Our findings showed a low level of awareness of the use of AI in cancer diagnosis and management. Respondents were positive about the potential of AI and indicated a high level of support for its use, albeit conditional on the fact that human oversight of the process is maintained. Analysis of qualitative data on support for AI reinforced the need for transparency and regulation regarding the use of images and personal health care data for research purposes. It also largely indicated that respondents would be supportive of assistance from AI in their cancer diagnosis and management provided there was human oversight of the process and evidence of AI’s ability. Notably, our findings demonstrated that challenges remain in addressing the concerns specifically of women and apprehension remains about the access to and use of health care data by commercial organizations.

The findings of this study have important implications for all stakeholders involved in the development and implementation of AI in pathology research and practice. They indicate the need for a balanced approach to the process, one that comprises openness, engagement, and education to cultivate a supportive and informed public. It is imperative that AI developers, researchers, and organizations deploying AI applications take this balanced approach into account to align with the expectations and concerns of the public regarding such technologies.