Human–AI teams must train together before high-stakes decisions put the partnership to the test.

Stress changes human cognition, so AI systems must adapt how they present information.

The best human–AI teams manage the relationship, not just the technology.

In network science, a dyad is a two-agent group that works together as a unit. Counterintuitively, a dyad has three parts, not two: Partner A, Partner B, and the relationship or agreements between them. A nurse–physician pair in a trauma bay is a dyad. So are the two paramedics working together in an ambulance.

Regardless of the type of dyad (equal partnerships, superior-inferior, flexible), the strength comes from the ability of the two partners to interact clearly and easily. A dyad of two experts who cannot communicate clearly will often lose to a dyad of less-skilled individuals who coordinate effectively.

Most discussions about using AI leave the relationship out. AI is framed as a tool that humans use, or as a system that runs independently and presents results to us. But, as AI systems move from limited tools toward something closer to co-pilots, the relationship changes into a human–AI team: a partnership where each member contributes different strengths.

The highest-performing teams will recognize this and actively manage all parts of the dyad. In this post, we look at three considerations for building human-AI teams that actually work in high-stakes moments.

You Can’t Build a Dyad in the Middle of a Crisis

A critical mistake in human-AI teaming is assuming the partnership will function automatically in high-stakes moments just like it does in low-stakes ones. This is, of course, a mistake in all-human teams as well, but it is amplified when different parts of the team respond very differently from each other in high-stakes moments.

A useful though imperfect analogy is teams that actively mix individuals from exceedingly different cultures.1 If one person comes from a culture that defers to authority and tolerates ambiguity well, while another comes from a culture with a flat hierarchy and a strong need to eliminate uncertainty, the dyad they form might struggle when they work on a high-stakes problem.

How do high-performing teams bridge this gap? They train together in low-pressure moments before they have to rely on each other in high-pressure ones. When a new teammate joins a wildland firefighting group or a military special forces team, they typically run exercises and practice drills with their new teammates before they deploy together, and we should treat silicon teammates the same way.

Human-AI teams tasked to operate in critical environments should proactively build experience in non-critical environments first. Simulation is a great way to build this experience. In Formula-1 racing, for example, drivers and engineers work with AI in pre-race simulations long before relying on those systems during an actual race.2

Stress Changes Human Decision-Making

The human brain is constantly making tradeoffs between speed and accuracy. In high-stress environments, we operate with a different cognitive architecture than we do in low-stakes moments—not just a degraded one, but a fundamentally different one. We tend to lose peripheral vision, rely more on pattern matching, and favor speed over accuracy, among other changes.3,4

As a result, our ability to partner with AI should be different in high-stress vs low-stress moments, and effective human-AI dyads will need to account for this. Complex dashboards and multiple competing alerts can overwhelm a decision-maker whose cognitive bandwidth is already strained.

In high-pressure environments, good design often means:

Highlighting the single most important piece of information

Surfacing clear choices rather than raw streams of data

Reducing noise and ambiguity

Well-designed AI systems can help their human partners maintain clarity when the environment becomes chaotic, while poorly designed ones can increase cognitive load, forcing the human operator to monitor and interpret the AI instead of focusing on the decision itself.

Stress Doesn’t Change the AI—but the World Changes Around It

Unlike humans, algorithms do not experience stress, and their internal processes remain constant regardless of external pressure. However, as the world around an AI model becomes more chaotic and rapidly changing, AI systems can easily suffer from model drift: a widening gap between the world the system was trained on and the one it currently inhabits.5

Imagine a ship navigating with a digital co-pilot. If the AI system was trained primarily on calm seas, it may perform well in normal conditions. But when the ship encounters violent storms or unpredictable currents, the model’s predictions may become less reliable.

Unless the human-AI dyad knows if the current situation is outside the training model, it might drift invisibly into poor performance, causing the human teammate to trust the AI one right when it is least reliable. As a result, drift detection is a crucial part of effective human–AI teaming, and high-performing systems need ways to signal when the AI’s confidence is falling or when conditions differ from its training environment.

As the surface area where we partner with (and not just use) AI increases, the goal is not to eliminate human judgment or automate difficult decisions. The goal is to create partnerships where each member of the dyad complements the other. When that balance is achieved, the result is something powerful: carbon and silicon teammates working together as a system that can remain calm, coordinated, and effective even when the environment becomes chaotic.

1. Measuring cultural differences is complex, but Geert Hofstede’s work on cultural dimensions gives an interesting take on multiple “axes of difference” between different cultures. See: https://geerthofstede.com/culture-geert-hofstede-gert-jan-hofstede/6d-model-of-national-culture/

2. See, for example: https://ioaglobal.org/blog/how-machine-learning-is-powering-formula-1-cars/

3. See, for example: Starcke K, and Brand M. Decision making under stress: a selective review. Neurosci Biobehav Rev. 2012;36(4):1228–1248. PMID 22342781

4. Dr. Zab Johnson does an exceptional job explaining the speed-accuracy tradeoff in human neurobiology here: https://youtu.be/BzTjbiYxVrg.

5. I’ve written about model drift in teams before. See: https://www.psychologytoday.com/us/blog/the-emergency-mind/202511/model-drift-how-mental-models-degrade-over-time

© Psychology Today