KD2 Lab, Team Rooms A&B

Additional speaker information: https://www.tu.berlin/kke/ueber-uns/team/tobias-feldmann-wuestefeld

Measuring trust in AI with neural markers of cognitive offloading

As Artificial Intelligence (AI) systems take on increasingly active roles in human workflows, a critical question emerges: how do people decide how much to rely on them? Trust in AI is a latent, dynamic construct — yet notoriously difficult to measure. Self-reports are slow and biased, and behavioral proxies are easily confounded by task demands. This talk presents a series of EEG studies developing implicit, continuous neural markers of trust-dependent cognitive offloading. A key methodological innovation is the exploitation of the brain's lateralized organization — the left hemisphere processes the right visual field and vice versa — which allows us to assign one hemifield to the human and the other to the AI, and directly read from the neural signal how much cognitive responsibility each party carries.
Using the N2pc component, we show that attention was redirected toward the AI's task when the AI was unreliable, rather than being fully offloaded. Using the contralateral delay activity (CDA) — a marker of working memory load — we tracked the full trust cycle in real time: as AI reliability was first established, then withdrawn, and then restored, CDA amplitudes rose, fell, and recovered, mirroring trust formation, violation, and repair. We further show how transparency about AI performance shapes this cycle. These findings demonstrate that the brain's lateralized organization offers a powerful lens for studying dynamic trust calibration, with direct implications for biosignal-adaptive system design.