As AI becomes embedded across enterprise IT, a new risk is emerging: cognitive surrender, the tendency to accept AI-generated recommendations without sufficient scrutiny. For mainframe professionals responsible for some of the world’s most critical systems, that risk carries unique consequences.

The term comes from a recent study by Steven D. Shaw and Gideon Nave of the University of Pennsylvania’s Wharton School. Across experiments involving more than 1,300 participants and nearly 10,000 reasoning problems, people could consult an AI assistant that was sometimes right and sometimes deliberately wrong. When the AI was correct, accuracy rose sharply. When it was wrong, participants followed its faulty advice nearly 80% of the time and performed worse than people who had no AI at all. Their confidence went up either way.

The researchers distinguish this from ordinary cognitive offloading. Using a calculator hands off a narrow task, but the person stays in charge of the reasoning. With cognitive surrender, the judgment itself changes hands.

It’s not hard to see why this happens. AI tools are fast, fluent, and confident, and much of the time their answers are correct. But a tool that is right most of the time is not the same as one that is right every time, and AI rarely signals which is which. A wrong answer usually arrives with the same polish and certainty as a right one. That distinction matters deeply in environments where the cost of error is measured in dollars, downtime, or risk to critical systems and reputations.

Why Cognitive Surrender Matters on the Mainframe

Mainframes continue to run the backbone of global business. Financial transactions, healthcare systems, government operations, and supply chains all rely on these platforms to operate with precision and reliability. The expectations are not just high, they are uncompromising. Systems must be secure, resilient, and accurate. There is very little room for assumption or unchecked automation.

AI is already being used to assist with code generation, system diagnostics, performance tuning, and documentation. These tools are powerful and, when used well, can significantly enhance productivity. They can help experienced professionals move faster and assist newer practitioners in navigating complex environments.

But when AI outputs are accepted without sufficient validation, especially in a mission-critical context, the risks become real:

• A suggested WLM setting may starve critical batch and online processing of CPU resources.
• Generated COBOL program that executes fine but may incur performance issues from Db2 query and configuration.
• A query about APF program protection omits issues that may lead to a security exposure
• A recommended solution to a performance problem that may affect the performance of other processing.

These are not hypothetical scenarios. They are the kinds of issues that can arise when judgment is outsourced too quickly.

The Hidden Risk of AI Confidence

The Wharton research points to something deeper than a tooling issue. Cognitive surrender is not about AI being flawed; it is about humans becoming less engaged in the decision-making process. When people defer too easily, they stop asking the second question. They stop testing assumptions. And as the study shows, they can feel more certain even as they become less accurate. Over time, this can erode the very expertise that organizations depend on.

For the mainframe workforce, this presents both a challenge and an opportunity.

The challenge is clear. As AI becomes more capable, there will be a natural tendency to rely on it more heavily. New entrants to the field may lean on AI as a primary guide rather than a supporting tool. Even experienced professionals may find themselves trusting outputs that “look right” without digging deeper. If left unchecked, this could weaken the culture of rigor that has long defined mainframe operations.

But the opportunity is just as significant.

AI Amplifies Expertise, It Doesn’t Replace It

As AI becomes more capable, organizations may mistakenly assume less expertise is required. In reality, the opposite may be true. The value of AI in a mainframe environment is directly tied to the knowledge of the person using it. A well-trained systems programmer, developer, or operator can use AI to explore options faster, validate ideas more efficiently, and uncover insights that might otherwise take longer to surface. They know which parts of an AI answer need checking, what must be tested before anything reaches production, and when a recommendation simply doesn’t fit how their environment works. In other words, the professional, not the tool, remains in control of the decision.

Because AI is only as valuable as the expertise behind it, strong, up-to-date mainframe knowledge has never been more important.

Deep technical knowledge acts as a counterbalance to cognitive surrender. It provides the confidence to question, the context to interpret, and the discipline to verify. When a systems programmer understands how workloads behave under different conditions, they are less likely to accept a generic optimization suggestion without testing it. When an application developer understands the intricacies of transaction processing, they can spot when a generated solution doesn’t align with performance or reliability requirements. When an operations team has a strong grasp of system behavior, they can distinguish between a useful recommendation and a misleading one.

The Case for Deep Technical Skills

The mainframe community has long emphasized technical expertise, professional development, and the knowledge needed to support mission-critical systems. Those priorities remain, but the rise of AI adds a new dimension. It is no longer enough to simply transfer knowledge. Organizations must ensure that their entire teams are equipped to work alongside AI in a way that enhances, rather than diminishes, human judgment.

This means training that goes beyond surface-level familiarity. It means building deep expertise in core areas such as z/OS, performance management, security, and application development. It means creating learning pathways that reinforce not just how to use tools, but how to think critically about their outputs. It means encouraging a culture where questioning is valued, and where validation is part of the process, not an afterthought.

Importantly, this is not about resisting AI. It is about using it wisely. In practice, that looks like:

• Treat AI output as a first draft, not a final answer. Review generated code, JCL, and configuration changes with the same rigor you would apply to a colleague’s work.
• Ask the AI to explain its reasoning. Then check that reasoning against your own understanding of the system. If you can’t explain why a recommendation works, you’re not ready to implement it.
• Validate before production. Test AI-suggested changes in a non-production environment under realistic workloads, and route them through your normal change management process.
• Remember what the AI doesn’t know. Site-specific standards, exits, and configurations rarely show up in generic AI advice. Keep them front of mind.
• Notice your own confidence. If a recommendation feels unusually easy to accept, treat that as a signal to slow down, not speed up.
• Share what you learn. When AI gets something wrong, document it so the whole team benefits.

The mainframe community has always been pragmatic. It has adopted new technologies when they add value and integrated them into existing environments with care. AI should be approached in the same way.

Organizations that invest in their people will be better positioned to realize the benefits of AI without falling into the trap of cognitive surrender. They will have teams that can move faster without sacrificing accuracy, that can innovate without compromising reliability, and that can leverage automation while maintaining accountability.

Leadership Sets the Standard

Leaders have a responsibility to ensure AI adoption strengthens decision-making rather than replacing it. Leaders set the tone for how technology is used. When they emphasize the importance of skills, training, and continuous learning, they reinforce a culture where expertise matters. When they provide time and resources for professional development, they signal that knowledge is a strategic asset. When they encourage teams to validate AI outputs and share learnings, they create an environment where cognitive engagement remains high.

The Wharton research offers a useful clue here: when participants had a reason to get the answer right and received immediate feedback, they were far more likely to question and override faulty AI recommendations. Leaders can build those same conditions into everyday work through clear accountability, peer review, and fast feedback loops.

Adopting AI responsibly is not about creating friction. It is about maintaining standards.

The goal is not to slow people down, but to ensure that speed does not come at the expense of sound decision-making. In a mission-critical environment, the best outcomes come from a balance of efficiency and rigor. AI can help with the former, but it is human expertise that ensures the latter.

Moving Forward Without Surrendering Judgment

The ease of accepting AI outputs can be both a strength and a weakness. Recognizing the possibility of cognitive surrender is the first step toward managing it.

As AI becomes increasingly embedded in enterprise technology, organizations should embrace innovation, including AI, but double down on the skills and knowledge that have always been the mainframe’s foundation. Invest in training. Support continuous learning. Encourage critical thinking. Build teams that are confident in their expertise and comfortable challenging what they see.

The future of the mainframe will be shaped not by AI alone, but by the professionals who know when to trust it, when to challenge it, and how to use it responsibly.

Darren Surch, CEO of Interskill Learning, is one of only two Lifetime IBM Champions for IBM Z, an Open Mainframe Project Ambassador, and an active member of the Mainframe Skills Council. He leads the industry’s defining mainframe workforce training organization, delivering over a million training hours annually and powering more than 80% of all IBM digital badges awarded for mainframe education.