In part three of our round table series, experts debate the ethics of handing over the highly developed expertise of pathology to AI. They also explore how AI validation data can be better managed and shared. Plus, they reveal the extent to which pathologists in their circles are becoming involved in AI development.
The panel
Where should the line between automation and human judgment be drawn?
“Trust isn't just something that develops between the technology and the pathologist; it's also part of the relationship between the pathologist and the patient.”
David West: I don't think there's a single point where we can draw the line between automation and human judgment. It's something that will evolve over many years as trust develops. Building that trust is as much a clinical and commercial challenge as it is a technological one.
One of the most instructive examples comes from outside pathology. In primary care, AI-powered ambient documentation tools are becoming increasingly common. A clinician can have a natural conversation with a patient while the AI automatically generates the clinical note, prepares the documentation for the electronic health record and even supports billing. The technology removes administrative work without interrupting the interaction between clinician and patient.
The important point is that these systems didn't begin by replacing clinical judgement. They started with a narrowly defined task and kept the clinician firmly in the loop. The physician reviews, edits, and approves the note before it becomes part of the medical record. As confidence has grown, these platforms have gradually expanded their capabilities, for example by surfacing relevant clinical evidence during the consultation.
I think pathology will follow a similar path. At every stage, there needs to be an expert human overseeing the process. Trust isn't just something that develops between the technology and the pathologist; it's also part of the relationship between the pathologist and the patient. Although pathologists often work behind the scenes, patients still place their trust in the specialist who signs the report and takes responsibility for the diagnosis.
I don't see that changing any time soon. The pathologist will continue to be the person who ultimately signs off the report and accepts legal and professional responsibility for the diagnosis.
Between today's largely manual workflows and that enduring responsibility lies a broad spectrum of opportunities for AI. As evidence accumulates and confidence grows, more tasks can be safely automated. But each step along that journey has to be earned through trust.
“The profession itself must decide which tasks can safely be delegated to AI, when the evidence is sufficient to trust those systems, and where human oversight must remain.”
Syed T. Hoda: Listening to David's example of ambient AI, I can certainly see the power of those technologies. But pathology presents a different challenge.
Pathologists spend years developing highly specialized knowledge that is both deep and extraordinarily niche. Much of that expertise isn't easily translated into the kind of plain-language interactions that underpin many of today's AI systems. If I walked outside my hospital, there might not be another person within a mile who shares the specific knowledge I've built in my subspecialty. That's the nature of pathology: it's a profession built on highly specialized expertise.
So, when we talk about where to draw the line between automation and human judgment, I don't think there's a simple answer. At the moment, the gold standard is still the judgment of experienced pathologists. The profession itself must decide which tasks can safely be delegated to AI, when the evidence is sufficient to trust those systems, and where human oversight must remain. In that sense, pathologists have to authorize the evolution of AI within their own discipline.
The real question is how pathologists choose to evolve into that space, and how much of their knowledge and decision-making they are prepared to entrust to technology while maintaining the standards of quality and patient care that define the profession.
“There's a risk that skepticism may give way to overconfidence.”
David Gibbs: One aspect we don't talk about enough is that AI has to account for the humanity of the people using it. Every pathologist, like every clinician, is subject to cognitive biases. At the moment, many of us approach AI with a healthy degree of skepticism, and that inevitably influences how we interpret its recommendations.
Over time, however, that bias may shift. As AI becomes more widely adopted and demonstrates its value, there's a risk that skepticism may give way to overconfidence. If workloads continue to increase while the pathology workforce remains constrained, pathologists may come to rely more heavily on AI simply because they have little alternative.
Even if the pathologist retains responsibility for the final sign-off, that doesn't eliminate the influence of human psychology. The way we review an AI-generated result will still be shaped by our expectations and experience.
You can see this in everyday practice. A pathologist reporting prostate biopsies, where a substantial proportion of cases are malignant, naturally expects to encounter cancer on a regular basis. By contrast, someone reporting endometrial biopsies, where only a small minority of specimens are malignant, is conditioned to expect benign findings. That difference in expectation influences perception and may affect the likelihood of overlooking subtle abnormalities.
If AI is to become a trusted partner in diagnosis, it can't simply model disease. It also needs to take account of the human decision-making process and the cognitive biases that influence it. Otherwise, there's a danger that the interaction between AI and the pathologist could inadvertently reduce, rather than improve, diagnostic performance.
What would it take for AI-skeptical pathologists to trust these technologies?
“The process is no different from evaluating a new clinical chemistry analyzer or any other laboratory instrument. “
Derek Welch: Ultimately, trust comes down to one thing: data.
Pathologists need to understand the fundamentals of how AI tools are developed and validated before they can have confidence in using them. That means asking the same kinds of questions they would ask of any diagnostic technology. What data was used to train the model? How does it perform in terms of sensitivity, specificity, and positive or negative predictive value? Has the training data introduced bias? Are there situations where the model produces misleading or erroneous outputs?
Those are exactly the kinds of questions we ask when evaluating new products. We've developed a formal due diligence framework to assess AI companies and their technologies before they're considered for clinical use.
In many ways, the process is no different from evaluating a new clinical chemistry analyzer or any other laboratory instrument. Every diagnostic platform has to undergo rigorous validation, quality assurance, and performance testing before it becomes part of routine practice. AI should be held to exactly the same standard. The methods of validation may differ, but the expectation is the same: the technology must demonstrate that it performs reliably and improves patient care.
That's how pathologists think. They're scientists. If you show them high-quality evidence that a technology works under real clinical conditions, they're far more likely to trust it than they are to be persuaded by marketing or impressive demonstrations alone.
“The technology may move rapidly, but trust has to be built one step at a time.”
David West: From a technology perspective, we're still at a relatively early stage. Companies like ours build on the capabilities developed by the frontier AI laboratories, and today's models are not yet able to solve many of the complex problems that pathologists deal with every day. Could they become capable of doing so in the future? Probably. But I think we're looking at a journey measured in decades rather than years.
We've already seen how this plays out with digital pathology itself. Compared with AI, digitizing glass slides is a relatively straightforward technological challenge, yet widespread adoption has still taken many years because of financial constraints, cultural change, and the need to build clinical confidence.
I think AI will follow a similar pattern. The underlying technology may advance extremely quickly, but adoption in routine pathology practice will happen much more gradually. Every new capability will need to earn the trust of pathologists through evidence, validation, and real-world experience.
In other words, the pace of progress in AI models and the pace of clinical adoption are unlikely to be the same. The technology may move rapidly, but trust has to be built one step at a time.
“Trust must be earned through rigorous validation, not assumed because a technology is promising.”
Syed T. Hoda: Before adopting an AI system, laboratories have to be confident that it works consistently on their scanners, with their workflows, and for their patient population. If that evidence isn't available, the clinical and legal responsibility still rests with the pathologist.
That's one of the reasons we haven't implemented AI for routine diagnosis at NYU. It's not a question of resources or a lack of interest. We evaluate these technologies closely, and our pathologists regularly ask why we aren't using them. The answer is that, at this stage, we haven't seen sufficient evidence that the available tools perform reliably enough within our specific environment to meet the standards of patient care we expect.
To reach that point, we'd need to undertake extensive local validation ourselves, and that's a substantial undertaking.
So, I think we're still some way from widespread adoption. What the field needs is a more robust and standardized framework for evaluating AI across different scanners, laboratories and clinical settings. Ideally, we'd also have objective ways of comparing the performance of different AI tools on the same cases.
“Involving pathologists throughout development is one of the most effective ways to build trust in the tools that eventually reach the laboratory.”
David West: Regulation is central to building trust, but it's only part of the picture.
One of the biggest challenges facing digital pathology is the lack of consistent technical standards. To some extent, that's a consequence of a competitive market. Scanner manufacturers, software vendors, and other technology providers all compete with one another, and that has slowed the development of the interoperability standards the field needs.
Even without AI, a digital pathology system is made up of multiple regulated components. The scanner, viewing software, and diagnostic display must all meet regulatory requirements and function together as an integrated system. If one component falls outside its approved specification – for example, if the display no longer meets the required standard – you may no longer be operating within a validated diagnostic environment.
That's why industry-wide standards are so important. We've worked closely with regulators and organizations such as the Digital Pathology Association to help develop more consistent approaches. If we can't establish robust standards for the underlying digital infrastructure, introducing additional layers of AI will only make the challenge more complex.
The other part of trust is cultural. At our company, we talk about a "pathologist-in-the-loop" development model. Rather than building software in isolation, our engineers work closely with practicing pathologists throughout the development process, testing pre-release versions, gathering feedback, and refining the product in rapid development cycles.
That collaboration is about far more than usability. It's about making sure we're solving genuine clinical problems. If the software doesn't address a real need in pathology practice, then we're building the wrong product.
Ultimately, I think pathologists should have their fingerprints on the evolution of these technologies. Their expertise is what gives AI its clinical relevance, and involving them throughout development is one of the most effective ways to build trust in the tools that eventually reach the laboratory.
“We need better mechanisms for sharing validation evidence.”
David Gibbs: At the moment, we're not using AI routinely to support diagnostic decision-making. Instead, we're focused on evaluating these technologies through extensive validation programs.
The challenge is that validating an AI product requires an enormous investment of time and resources. Even within our national health service, that work is still largely carried out at an institutional level. As a result, the same validation studies are often repeated across multiple organizations, sometimes producing slightly different findings. That's an inefficient use of expertise and resources.
I think regulation has an important role to play here. In England, the National Institute for Health and Care Excellence (NICE) is currently evaluating AI technologies for healthcare, and I hope that will provide clearer guidance for the field. But I also think we need better mechanisms for sharing validation evidence.
We've already moved in that direction with point-of-care testing, where validated performance data can be shared through a national verification and validation database. Laboratories can then build on that evidence rather than repeating the entire process themselves. I'd like to see a similar approach for digital pathology and AI, with rigorous validation data published in a national repository that other organizations can use to streamline their own verification processes while maintaining appropriate local oversight.
There's still a great deal of work to do before we reach that point.
“Pathologists naturally have greater confidence in technologies they've helped design, test, and validate themselves.”
Syed T. Hoda: One approach we're exploring at NYU is to develop our own AI tools rather than relying solely on commercial products.
Part of the reason is that pathologists naturally have greater confidence in technologies they've helped design, test, and validate themselves. Instead of evaluating a product that has been built elsewhere, they're shaping a tool around their own workflows and clinical standards.
We're starting with relatively straightforward, high-frequency tasks, such as counting and quantification. Those applications are well suited to local development because we can evaluate them rigorously using the same scientific standards we apply to any other aspect of pathology.
That process also helps build confidence. We understand exactly how the models were developed, how they perform and, just as importantly, where their limitations lie.
I recognize that this isn't an option for every laboratory. Developing AI requires access to technical expertise, software engineering, and close collaboration between clinicians and data scientists, resources that are more readily available in academic medical centers than in smaller institutions.
Even so, I think institution-led development could become an important part of the regulatory landscape. Our intention is to validate these tools rigorously and, ultimately, seek FDA clearance. If successful, it could provide another pathway for introducing AI into clinical practice – one in which pathologists play a leading role in creating the technologies they go on to trust and use.
“Even though we are not actively using AI in-house, it is finding a way into the workflow.”
David Gibbs: That approach is also appearing in pathology contract work. Some of our work is sent out, due to our demand–capacity imbalance, and we’ve noticed that our outsourcers are developing their own AI tools, particularly for quantification and quality control. Therefore, even though we are not actively using AI in-house, it is finding a way into the workflow.
I think we need to examine the tools that our contractors are using, against our workflow priorities, to inform the types of products we need to develop in-house.
If we were to reconvene in five years, what outcome would convince you that technology has made pathology more human rather than less?
“Developers should invest in the user experience.”
David West: When I talk to our customers about the move to digital pathology, a lot of the conversation focuses on the financial barriers. Those are certainly real, and places such as the UK have benefited from strategic investment and top-down support to build the necessary digital infrastructure. I've also admired what NYU has achieved with its digital pathology implementation.
What perhaps receives less attention is the user experience. As someone from the technology side, I think our industry sometimes underestimates just how significant the transition is for pathologists. They're deeply familiar with the microscope. It's an immersive, tactile, and highly ergonomic way of working, and that experience has been refined over decades.
Technology developers need to pay much closer attention to recreating that experience in the digital environment. If we invest in making digital pathology intuitive and enjoyable to use, we'll create a much better experience for pathologists, and that in turn provides the foundation for successful AI adoption.
I also think AI itself can help build confidence in digital pathology. It can assist with navigation, guide users through complex cases, and enable capabilities that simply aren't possible with a conventional microscope. The real opportunity is not to replicate the microscope digitally, but to preserve everything that makes it effective while adding entirely new capabilities. To me, that's the holy grail.
“Pathology should be more integrated, more clinically influential, and more closely connected to patients.”
David Gibbs: Five years from now, I'd hope to see pathology much more closely integrated into the patient's entire care pathway.
That means becoming involved earlier, helping to trigger the most appropriate diagnostic pathway from the outset, and remaining involved later by supporting treatment selection and ongoing clinical management. Pathology should no longer be seen as a service that simply provides a diagnosis, but as a continuous contributor to patient care.
I also think the nature of disease management will continue to change. While many cancers will become increasingly curable, many others will be managed as long-term conditions rather than acute illnesses. That shift will require a very different relationship between diagnostics and treatment, with pathology playing an ongoing role in guiding therapeutic decisions throughout the patient's journey.
If technology helps us achieve that – making pathology more integrated, more clinically influential, and more closely connected to patients – then I would say it has made the profession more human rather than less.
“When people are actively involved in shaping change, the profession becomes more connected, more collaborative and, ultimately, more human.”
Syed T. Hoda: When I think about what would make pathology more human, my focus is less on the technology itself and more on the profession.
Technology has transformed all of our lives, and it will undoubtedly continue to transform pathology. What concerns me is whether the profession will engage with that change. I sometimes worry that some of the most influential voices in pathology are still reluctant to embrace the conversation or recognize that the role of the pathologist is evolving.
What gives me optimism is seeing people who are willing to engage with those questions openly. I think everyone on this panel recognizes that the profession is changing, and that we have an opportunity to help shape that future rather than simply react to it.
For me, that's what makes a profession more human: engagement. Open discussion, a willingness to acknowledge both the opportunities and risks, and the confidence to challenge long-held assumptions. When people are actively involved in shaping change, the profession becomes more connected, more collaborative, and, ultimately, more human.
The patient will always be at the center of medicine. But I also think that if pathologists feel more engaged, more valued, and more connected to one another, they'll be better equipped to deliver the kind of compassionate, patient-centered care that we're all striving for. That's the future I'd like to see in five years' time.
“Trying to meet today's workload without AI is like chopping down a tree with an axe when a chainsaw is available.”
Derek Welch: If we reconvened in five years, I think we'd be looking at a very different landscape.
If you compare today's AI with where it was five years ago, the pace of progress has been extraordinary. The effort required to develop and train these models has fallen dramatically, while their capabilities have increased at an astonishing rate. I expect that trajectory to continue. Some AI applications that seem innovative today may become standard of care within the next five years.
Take breast cancer as an example. For decades, assessing biomarkers such as estrogen receptor, progesterone receptor, and HER2 has been routine practice. I can envisage a future in which the standard of care is not simply to perform those tests, but to analyze them using validated AI algorithms that provide highly accurate quantitative and qualitative assessments because the evidence shows they produce better, more consistent results.
To me, that's what makes pathology more human. If technology helps us answer the questions we're asking of every case more accurately, more consistently, and with greater confidence, then it is directly improving patient care.
In many ways, pathology is an informatics specialty. Every day we're answering an endless stream of questions: What is the diagnosis? Has the tumor reached the margin? How extensive is the disease? What does this biomarker show? Our responsibility is to answer those questions correctly, every single time. That's how pathologists care for patients, even though they rarely meet them face to face.
Digital pathology has already moved us in that direction by making it easier to collaborate with colleagues, seek expert opinions, and deliver diagnoses more quickly. AI has the potential to build on those gains.
Just as importantly, it could help address one of the profession's biggest challenges: burnout. The volume of pathology work continues to increase, while the workforce has not kept pace. Pathologists need tools that reduce cognitive load, improve confidence, and remove unnecessary burden from routine practice.
I often use a simple analogy. Trying to meet today's workload without these technologies is like chopping down a tree with an axe when a chainsaw is available. Both will get the job done, but one is far more efficient and far less exhausting. If AI can help pathologists deliver better care with greater confidence while reducing fatigue and burnout, then I think we'll be able to look back in five years and say that technology has made pathology not just more efficient, but genuinely more human.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
