Holding AI Accountable

Short answers to ‘What are you actually doing in your Ph.D?’

Patrick Altmeyer

Delft University of Technology

Arie van Deursen

Cynthia C. S. Liem

February 25, 2026

The Ground Truth

Figure 1: Predictors of default risk.

The Ground Truth

Figure 2: Ground truth outcomes across two predictors.

Black-Box AI

Figure 3: Classifier predicts correctly 8 out of 10 times.

Black-Box AI

Figure 4: Simple counterfactual explanation for the black-box AI.

Black-Box AI

Figure 5: One happy recourse recipient, many losers.

Black-Box AI

Figure 6: Plausible counterfactual explanations for the black-box AI.

Black-Box AI

Figure 7: One somewhat happy recourse recipient, no losers.

Big, Beautiful Black-Box AI

Figure 8: Classifier predicts correctly 9 out of 10 times. But …

Big, Beautiful Black-Box AI

Figure 9: Plausible counterfactual explanations remains valid. Happy days?

Big, Beautiful Black-Box AI

Figure 10: White-washed black-box: plausible CE hides bias.

Holding Models Accountable

Figure 11: A model trained to use plausible explanations for predictions

‘ok but agi bruh’