I was recently sent a paper in The Lancet Gastroenterology & Hepatology that pulled me in for a closer look. OK, it was about colonoscopies, but the observations can be applied to a broader application of artificial intelligence. This study followed gastroenterologists using AI to help detect polyps. With the AI running, detection rates improved. And that’s no surprise.

But here’s what got me thinking. When those same doctors went back to working without AI, their detection rates dropped below where they’d been before the technology was introduced.

That’s more than just a dip. It’s what I call “AI rebound,” the paradox where a tool that boosts performance in the moment leaves people worse off when it’s removed.

More Than a Medical Story

If this can happen to highly trained specialists, it’s easy to imagine it in other domains. A driver grows comfortable with Tesla’s Full Self-Driving and finds their reflexes slower in a sudden takeover. A pilot spends most of a flight on autopilot and then has to land manually in bad weather. Even in creative work, I’ve seen writers lose their natural flow after leaning too heavily on a digital assistant.

The pattern seems to be the same from topic to topic. When a system takes over, the human role changes. We’re not “doing” the skill in its full form anymore; we’re supervising, monitoring, or even just waiting for something to go wrong. And while that might feel safer in the moment, I think that it might alter the fundamental human dynamics.

The Mechanics of AI Rebound

AI rebound, as I’m calling it, may be related to the “out-of-the-loop” problem that’s seen in conventional automation. When automation handles the details, situational awareness dulls. And in that context, we scan less, anticipate less, and make fewer micro-adjustments. Simply put, the mental models we rely on to navigate complex situations shrink because the system is doing what we once did ourselves.

Over time, this isn’t just about pausing a skill; it may be more akin to erosion. And when the technology steps away, the skill doesn’t simply return to baseline. It can come back lower.

The Lancet study didn’t find that AI was misidentifying polyps or making dangerous errors. It found that without AI, people were less sharp than before they started using it. That’s the paradox and one with significant implications. And it might be time to question the tool that improves performance while it’s active but degrades the very abilities it was meant to enhance, particularly when the tool is defined by intermittent or occasional use.

Why the Baseline Matters

In high-stakes fields, small changes in performance have real consequences. In medicine, not noticing a small lesion can mean a missed diagnosis. On the road, a half-second delay can turn a near-miss into a collision. In business, hesitation or uncertainty can derail a critical decision.

It’s easy to focus on the gains we see when AI is switched on. But the baseline matters just as much—because that’s where we operate when the tool is absent, fails, or needs to be set aside.

Designing Against the Drop

If AI rebound is a real and measurable risk, the solution isn’t to avoid AI but to integrate it in a way that preserves core human competence. And the potential fix might be as simple as making a few adjustments in the way we use technology.

• Mix AI-on and AI-off sessions so people continue practicing their full skill set.
• Highlight human-first decision-making with appropriate AI support.
• Incorporate regular takeover drills where speed and accuracy are measured without AI assistance.
• Track and reward unaided performance alongside AI-assisted results.

These are not just technical fixes; they’re design choices that keep humans engaged as active participants rather than passive overseers. And by tracking the upside and downside, it may foster added AI-augmented skills and diminish the potential for AI rebound.

Caution and Opportunity

AI rebound isn’t about fearing automation or clinging to the old ways of working. It’s about understanding how technology shapes our capabilities over time. And, as it does, taking steps to make sure that what we gain today doesn’t undermine us tomorrow.

This gastroenterology paper is a clear example of how easily this can happen. The doctors in that trial didn’t lose their medical degrees or their experience, but their sharpness dipped when the AI was gone. That’s a subtle, almost invisible shift, until it matters.

Today, the opportunity to name it, measure it, and design against it should be on our radar—from the operating room to the classroom. Because the day will come when the machine goes quiet, and our performance will depend on what we’ve kept alive in ourselves.