CATCH

Project

About the project

CATCH is a design science research project on adaptive AI for work that is demanding, time-critical, and unforgiving of a badly-timed suggestion.

Putting a person and an AI together often produces worse decisions than either would have reached alone.

A meta-analysis of more than 100 studies found human–AI combinations frequently underperforming both humans and AI on decision-making tasks, precisely where collaboration should excel.Vaccaro, Almaatouq & Malone, 2024

CATCH takes that result as a starting point rather than a curiosity. The explanation it pursues is temporal: AI support is designed once, for an average person in an average moment, and then delivered unchanged to a real person whose capacity is moving the whole time.

When a coordinator is overwhelmed, fixed support is too thin. When they are sharp, the same support is too heavy, and begins to erode the judgement it was meant to assist. The help does not fail because it is wrong. It fails because it does not know when to arrive.

The mismatch

One incident, two surges. Demand never arrives evenly, whether the scenario is a wildfire or a drone incursion. Watch what the same amount of help does at different moments.

Support needed versus support provided, across one scenarioA schematic. Demand rises in two surges while the coordinator's capacity declines, so the support they need fluctuates. Static AI provides a constant level: too little during the surges, too much in between. Adaptive AI follows the need, and the gap closes.Incident beginsIncident resolvedSupportToo littleToo muchToo little
Support neededSupport providedShortfallExcessStatic support is too thin through both surges and too heavy in the lulls between them. That is the cognitive mismatch CATCH sets out to remove.

Reading state from interaction

To adapt at the right moment, a system has to know the moment. CATCH infers it from how a person already works: no extra hardware, no interruption, no self-report.

Move your cursor across this panel

Movement entropy·

How disordered the direction of travel is. Rises with searching and hesitation.

Timing variance·

How irregular the gaps between actions are. Steady work is steady in rhythm.

Re-seeking·

Path travelled against ground covered. Retracing suggests the information was not found.

A demonstration of the kind of signal the project works with. These three features are computed from your cursor in your own browser, discarded as they are read, and never sent anywhere.

Interaction

Captured from ordinary use, with no extra hardware

  • Mouse trajectories, sampled at 60 Hz
  • Click events and navigation paths
  • Interaction timing and pause structure
  • Information re-seeking behaviour

Physiological

Used to validate what interaction alone can infer

  • Heart rate variability (Empatica EmbracePlus, 64 Hz)
  • Electrodermal and pulse activity
  • Eye tracking (Tobii Pro Fusion, 120 Hz)

Self-reported and performance

The ground truth the models are trained against

  • NASA-TLX workload ratings
  • SPAM situational awareness queries
  • Fatigue scales
  • Task success, decision quality, response timeliness

What is missing, and what we will do about it

  1. 01State inference

    Interaction-based functional state inference in high-stakes coordination

    No validated framework exists for inferring functional state during actual emergency coordination. Which interaction features are most diagnostic of performance degradation, how their reliability varies across contexts, and whether laboratory-trained models generalise operationally are all open.

    Objective 1

    Develop and validate interaction analytics for real-time functional state inference in emergency operations contexts.

  2. 02Adaptive design

    Adaptive AI design principles for dynamic human–AI collaboration

    Current AI systems provide uniform assistance regardless of user state. Design knowledge is missing on which tasks to automate when workload is high without degrading skill, how to maintain situational awareness during cognitive offloading, and how to preserve coordinator agency.

    Objective 2

    Design and evaluate adaptive AI systems that dynamically adjust support based on inferred coordinator state, to maintain performance, preserve situational awareness, and delay fatigue onset.

  3. 03Team resilience

    Multi-agent coordination resilience in human–AI teams

    How human–AI teams coordinate during multi-agency emergencies is poorly understood. When should AI escalate a decision rather than handle it? How should coordination authority shift as conditions change? What keeps a team resilient when unexpected events disrupt established patterns?

    Objective 3

    Establish design principles for coordination-level resilience in multi-agent human–AI emergency response teams.

Where the work happens

  • LUT

    LUT University

    Lahti, Finland · AI and HCI Behavioral Lab

    Varjo XR-4 headsets, Tobii Pro Fusion eye tracking, Empatica EmbracePlus sensors, multi-participant capacity and sound-attenuated rooms. Host to the problem analysis, design and development, state inference validation, trust calibration and team resilience studies.

  • DLR

    German Aerospace Center

    Braunschweig, Germany · Simulation facilities

    High-fidelity cockpit simulators, air traffic control simulation, psychophysiological equipment and established workload protocols. Provides the aviation emergency scenario and professional pilot participants, a validation domain impossible to reach at LUT alone.

  • USF

    University of South Florida

    Tampa, United States · Behavioral AI Lab

    Independent replication capacity, a behavioural laboratory supporting team studies, established emergency management participant pools and high-performance computing. Establishes external validity at a geographically and institutionally separate site.

Partners and people

Funding

CATCH is funded by the Research Council of Finland, funding decision 377657, as an Academy Research Fellowship held by Dominik Siemon.