Research

My current programme, Human Cognitive Autonomy in AI Interaction, studies when AI dialogue supports users’ reflective agency and when it instead captures, substitutes for, or weakens it. This is a dimension of AI safety distinct from the accuracy, harmfulness, or appropriateness of any single model output.

Using text-derived measures of integrative complexity (differentiating and integrating multiple perspectives), intellectual humility (recognising the limits of one's own knowledge), and decentering (stepping back from thoughts and frames rather than being absorbed in them) within dialogue transcripts, we study how AI interactions affect subsequent reasoning and reflective agency. The programme combines theoretical work, public research instruments available on GitHub, and preregistered analyses to build an empirical basis for evaluating autonomy-preserving AI. Prospective experiments test when these effects persist beyond the interaction itself.

Recent work includes a reanalysis of 1,782 human-AI persuasion dialogues showing that text-measured integrative complexity (IC) moderates belief revision in an inverted-U pattern, with safety-relevant adverse movement concentrated in low-IC dialogue contexts (preprint, 2026), and a theoretical Perspective with Sara W. Lazar (MGH/Harvard), currently under review, introducing receiver-side examinability as a second axis of AI safety distinct from content compliance.

As part of that collaboration, a multi-method validation of the Integrated Decentering measure used in this programme, with convergent behavioural, neural, and longitudinal evidence across non-overlapping samples, is currently under revision, building on earlier work published in Scientific Reports (Nature portfolio, 2024).

This programme continues a long-standing thread in my work: computationally modelling and measuring cognitive processes that matter for agency. That thread runs from cognitive-architectural simulation with robotic embodiment (LIDA), through text-based developmental measurement in the Loevinger/Cook-Greuter tradition, and draws on years of applied machine-learning research at Amazon Web Services and McKinsey, much of it focused on extracting clinically meaningful signal from complex unstructured data at scale.

The common thread has been modelling how people form, use, revise, and step back from internal representations. My current work brings that question to AI dialogue: how can these systems create conditions for people to think with them, without thinking for them?

Selected Publications and Manuscripts

• Madl, T., & Lazar, S. (under review). A receiver-side blind spot in AI safety.
• Madl, T. (under review, preprint available). Text-measured cognitive complexity predicts belief revision in AI persuasion.
• Madl, T. (2024). Exploring neural markers of dereification in meditation based on EEG and personalized models of electrophysiological brain states. Scientific Reports, 14, 24264.
• Madl, T., Xu, W., Choudhury, O., & Howard, M. (2022). Approximate, Adapt, Anonymize (3A): A framework for privacy-preserving training-data release for ML. AAAI Workshop on Privacy-Preserving AI.
• Nadarajah, N., ... Madl, T., ... and Haferlach, T., (2021). Automated Disease Classification Using Whole Genome Sequencing (WGS) and Whole Transcriptome Sequencing (WTS) Data with Transparent Artificial Intelligence (AI). Blood, 138, p.275.
• Trapp, M., Madl, T., Peharz, R., Pernkopf, F., & Trappl, R. (2017). Safe Semi-Supervised Learning of Sum-Product Networks. Conference on Uncertainty in Artificial Intelligence (UAI).
• Madl, T. (2016). Deep Neural Heart Rate Variability Analysis. NeurIPS 2016 Workshop on Machine Learning for Health (ML4H).
• Madl, T., Franklin, S., Chen, K., Montaldi, D., & Trappl, R. (2016). Towards real-world capable spatial memory in the LIDA cognitive architecture. Biologically Inspired Cognitive Architectures, 16, 87-104.

A more complete list of publications is available on Google Scholar.

Email

tamas.madl@(1)
substitute (1) by ofai.at