A medical device can produce the right answer and still be unsafe.
That is becoming an increasingly important regulatory problem as medical devices move from physical controls and displays to software interfaces, clinical decision support, and AI-generated outputs. The FDA’s latest human factors guidance makes the point clear: safety is not only about whether the technology works. It is also about whether people can use it correctly in the real conditions in which the device will operate.
In May 2026, the FDA finalized Content of Human Factors Information in Medical Device Marketing Submissions, establishing a risk-based framework for the human factors information manufacturers should include in 510(k), De Novo, PMA and HDE submissions. In August 2026, the FDA also finalized an updated Applying Human Factors and Usability Engineering to Medical Devices guidance. Together, the documents put renewed regulatory emphasis on the interaction between the device, its users and its environment. The message for manufacturers is straightforward: a technically correct device is not enough if the user interface can turn a correct output into an incorrect clinical action.
The “right AI, wrong outcome” problem
Consider an AI-enabled diagnostic device that correctly identifies a suspicious lesion.
The algorithm performs exactly as intended. Its sensitivity and specificity are validated. Its output is technically correct.
But the physician sees the result on a crowded interface, misunderstands the confidence indicator, overlooks an alert or interprets the recommendation differently from what the manufacturer intended.
The AI was right. The clinical outcome can still be wrong.
This is precisely where human factors engineering becomes important.
The FDA defines human factors/usability engineering around the interaction between users, the use environment and the device user interface. That interface is much broader than a touchscreen. It can include displays, alarms, controls, physical components, software logic, packaging, labeling and training materials.
For AI-enabled devices, that means the regulatory question is no longer simply:
“Is the model accurate?”
It increasingly becomes:
“Can the intended user correctly understand and act on what the model produces?”
FDA is turning usability into submission evidence
The May 2026 guidance is significant because it focuses not simply on whether manufacturers perform human factors work, but on what human factors information should be documented and submitted to FDA reviewers.
The guidance applies across major marketing pathways, including 510(k)s, De Novo requests, PMAs and HDE applications. FDA describes the framework as risk-based and intended to improve consistency and efficiency in its review of human factors information.
The timing also matters.
For submissions received before August 1, 2026, FDA generally does not expect manufacturers to have incorporated the newly recommended information. But for submissions going forward, manufacturers need to plan around the new framework.
That shifts human factors from a “design team exercise” toward a more visible part of regulatory strategy.
The user is now part of the safety equation
The FDA’s approach recognizes something engineers sometimes underestimate: users do not interact with devices in laboratory conditions.
A physician may be working under time pressure. A nurse may be interrupted. A patient may have limited health literacy. A home-use device may be operated in poor lighting or a noisy environment.
FDA specifically identifies factors such as vision, hearing, dexterity, memory, cognitive ability, health literacy, language skills and familiarity with the device as characteristics that can affect safe operation.
The environment matters too.
A device may be used in a crowded clinical room, in low light, with multiple similar components nearby or while the user is distracted. FDA considers these conditions part of the device-user system rather than external problems that manufacturers can simply ignore.
That is particularly relevant to AI.
An AI model does not operate in isolation. Its output enters a human decision-making chain.
Data → AI output → interface → human interpretation → clinical action
A failure anywhere in that chain can create risk.
FDA’s own example shows how serious “use error” can be
The FDA gives a striking example involving an infusion pump.
A nurse was changing the concentration of a prescribed medication. While programming a bolus, the nurse misunderstood the default settings and accepted the bolus concentration as the final dose. The patient consequently received a three-fold overdose.
The pump did not necessarily have to malfunction for harm to occur.
The problem was the interaction between the device and its user.
That distinction is central to FDA’s approach. A “use error” can occur when the outcome differs from what was intended without the device itself malfunctioning. FDA notes that poor design or conditions promoting incorrect use can contribute to such errors.
For AI-enabled devices, the same principle creates a new category of risk.
The model may be functioning correctly while the human-machine interface fails.
Why training is not the complete answer
One of the most important points in FDA’s human factors framework is that manufacturers should not treat training and labeling as substitutes for good interface design.
FDA’s guidance states that modifying the user interface is generally more effective for addressing use-related hazards than relying on labeling or training. Training depends on memory, while labeling may not be accessible at the moment a decision is made.
This matters enormously for increasingly complex devices.
If a clinician has to remember a long sequence of instructions to correctly interpret an AI recommendation, the problem may not be the clinician.
It may be the interface.
The regulatory expectation is therefore moving toward designing out foreseeable use errors, rather than simply warning users about them.
The new guidance also strengthens the business case for testing early
The FDA recommends identifying critical tasks and evaluating user interactions during development, rather than waiting until the end of the product lifecycle. Simulated-use testing can uncover errors that analytical assessments miss because real users can behave in unexpected ways.
This is particularly important for devices with changing software interfaces.
A seemingly minor modification, such as changing the location of an alert, restructuring information on a screen or altering how an AI recommendation is presented, can change user behavior.
For manufacturers, that means usability cannot be treated as a one-time validation exercise.
It needs to track the product.
That is already reflected in the FDA’s broader regulatory ecosystem. The agency recognizes IEC 62366-1 as a consensus standard for applying usability engineering to medical devices and specifically describes the process as one for identifying and mitigating risks associated with normal use and use error.
AI makes the problem more complicated, not less
AI introduces a particular human factors challenge: automation can change how people behave.
A clinician may over-rely on an AI recommendation. Another may ignore it because the system provides insufficient explanation. An alert may be technically accurate but poorly timed. A probability score may be mathematically meaningful but clinically confusing.
These are not necessarily model-performance problems.
They are human-machine interaction problems.
The FDA already recognizes that AI-enabled devices require different assessment approaches depending on their intended use, including diagnosis, triage, prognosis, treatment response prediction and risk assessment. The agency also identifies new questions arising from natural-language processing and large language models in medical devices.
This creates an important regulatory distinction:
Model validation asks whether the AI performs correctly.
Human factors validation asks whether people can use the device safely and effectively.
Both can be necessary.
What manufacturers should take away
The practical impact of the 2026 guidance is less about adding another document to a regulatory submission and more about changing when human factors enters the development process.
For manufacturers developing AI-enabled or software-driven devices, four questions should be asked early:
- Who will actually use the device?
- Not just the intended user on paper, but users with different levels of expertise, health literacy and familiarity.
- What are the critical tasks?
Which mistakes could cause serious harm?
- How will the AI output be interpreted?
Is the information presented in a way that supports the intended clinical decision?
- What happens when the user is under pressure?
Real-world conditions, interruptions and imperfect environments need to be considered.
The FDA’s human factors program already connects these issues to post-market surveillance, adverse-event reporting and recalls. The agency encourages manufacturers and healthcare facilities to report use errors and near misses because the information can lead to changes in device design and improved safety.
That creates a feedback loop:
Design → usability testing → regulatory submission → real-world use → use-error data → redesign
The bigger regulatory shift
The most interesting part of FDA’s 2026 approach is that it challenges an increasingly outdated definition of device performance.
A device does not operate in a vacuum.
For an AI-enabled medical device, the “product” is effectively the combination of the algorithm, software, interface, user and environment.
That means a model can have excellent accuracy and still create unacceptable risk if the interface encourages the wrong action.
For manufacturers, the lesson is therefore bigger than “prepare better human factors documentation.”
Design the device for the person who will actually use it, not the perfectly trained user imagined in the engineering lab.
Because in the next generation of medical devices, the hardest safety problem may not be getting the AI to give the right answer.
It may be making sure the human recognizes what that answer means, at exactly the moment it matters.
