The answer to our questions about AI might be to keep asking questions

Artificial intelligence is everywhere today: writing emails, transcribing notes for medical records, and making self-driving cars a reality. “I can’t go to a dinner party where AI doesn’t come up,” Alicia Patterson, assistant professor of applied philosophy at Oregon State University, said at the ScienceWriters2026 conference in Corvallis, Oregon, on Sept. 26.

AI’s rapid spread, however, has come without moral guardrails, bringing to the forefront uncertainty about the ethical use of AI. And while a bit of uncertainty may be acceptable when automating an email, our tolerance for failure decreases when it comes to health care, where the stakes are high. How do we mitigate uncertainty in a rapidly changing technological environment? A group of panelists at the Council for the Advancement of Science Writing’s New Horizons in Science briefing didn’t have concrete answers, but suggested that we keep asking questions.

For Mohammad Adibuzzaman, assistant professor of medicine and biomedical engineering and leader of the Healthcare Ethical AI Lab (HEAL) at Oregon Health and Science University, one fundamental question to ask is, “what can go wrong?”

For example, researchers are testing using AI to evaluate patients for clinical trials, which a time-consuming process that requires the expertise to understand a patient’s medical history. “The risk in this particular case is, if AI misses someone [who is] eligible, then that patient might not get the advanced care they’re supposed to,” said Adibuzzaman.

At the other end of the spectrum is ScribeMD, an AI-based tool that listens to and records notes for patient visits. One of the most widely accepted uses of AI in healthcare, ScribeMD is largely deemed an AI success story, Adibuzzaman said. Yet even this tool has its risks. Clinical notes taken by an AI scribe are verified by the doctor, but errors or biases could lead to inaccurate medical records that could negatively affect future patient care.

From a philosophical perspective, Patterson noted that AI’s failures in healthcare have consequences not only for patients, but also for clinicians. “As things become automated, we tend to want to have workers in the loop in some way,” said Patterson, referring to the doctors or nurses who supervise the space between technology and human judgment.

But when automated systems fail, who is liable? In the past, Patterson said, blame has been misattributed to the person rather than the technology, putting the worker in a moral “crumple zone.” “Are we going to try to figure out how to protect these workers?” Patterson asked.

Despite the high-stakes risks in implementing AI in healthcare, Adibuzzaman sees a path forward. His research group is co-designing AI technology with those who will be using it, such as doctors and patients.

In a study aimed at improving the management of sickle cell disease, an app-based AI system will integrate patient data from an electronic health record with physical activity metrics available on the user’s phone. Adibuzzaman’s group has assembled a team of patients to co-design the system. The developers will ask the patients for feedback throughout the process, working side-by-side to improve the AI system. “I do strongly believe this is the right approach,” said Adibuzzaman, to prioritize patient safety and increase the likelihood of the system’s acceptance into the community.

Adibuzzaman emphasized that AI systems should come with disclaimers about their limitations and the consequences of failure, which should be clearly communicated to users and others impacted by its use.

Transparency is necessary for long-term sustainability of these models, Adibuzzaman and Patterson agreed. In every sector, from healthcare to autonomous cars, Adibuzzaman said, there need to be more open discussions and questioning about “the concepts of liability, risk, and the incentives around those systems.”