At this year’s Human Factors and Ergonomics Society (HFES) ASPIRE 2026 conference, one theme stood out: AI is increasingly being integrated into decision-making as a recommender, in which AI generates a recommendation, and the human is expected to verify the output, validate its accuracy, and ultimately take responsibility for the decision.
This creates a challenging dynamic. The technology is intended to make decision-making easier, yet the human must often do substantial additional work to determine whether the AI’s output is useful, accurate, and appropriate for the situation. Success is attributed to the power of AI. Failure ends up being blamed on human ‘over-reliance’, laziness, or a slew of other factors even when the AI recommendation is poor.
Brittleness in implementing AI as a recommender
During our panel and throughout the conference, the concept of brittleness repeatedly resonated with people working to integrate AI into their systems. As people have gone down the traditional tech-centric approaches of AI implementation, they can feel the tension between AI’s power and benefit, and its critical flaws.
When AI is designed as a recommender, its outputs can become opaque additions to an already complex decision-making process. Users may receive an answer without having sufficient visibility into the underlying evidence, assumptions, relationships, or uncertainties that produced it.
This is particularly concerning in complex, dynamic environments, where the quality of a decision depends on more than the accuracy of a single recommendation. Context changes and priorities shift. A recommendation that appears reasonable in isolation may be inappropriate when considered alongside the broader situation.
Building safer, more resilient systems requires us to move beyond simply adding AI-generated recommendations to existing workflows. We need to understand how AI capabilities can be integrated with the information, expertise, and relationships that people rely on to make decisions.
This means designing AI as part of the broader decision-making system, rather than treating it as a separate entity that produces answers for humans to check. The goal should not be to give people more recommendations to evaluate. It should be to help them make better decisions.
““Design focus should not solely be on the quality of AI recommendation, but on the quality of the decisions that come from it.””
At Resilient Cognitive Solutions (RCS), we apply Cognitive Systems Engineering to design decision support around the cognitive work people actually need to accomplish. We recognize that AI can contribute far more than a recommendation. When its capabilities are thoughtfully integrated with domain expertise, relevant information, and the user's decision requirements, AI can help people navigate complexity without losing sight of the bigger picture.
AI does not have to be an opaque addition that introduces new brittlenesses into an already complex system. With deliberate, decision-centered design, we can build systems that use AI to support effective decision-making while remaining responsive to the changing conditions of the real world.
As AI adoption continues to accelerate, we need to ask a different question. Instead of asking, “What can AI recommend?” we should ask, “How can AI best support the decisions people need to make?”