The ELIZA Effect when intelligence is artificial

Abstract

Artificial Intelligence models can be informative; they can provide insight and accelerate scientific discovery. However, they can also hallucinate, reproduce biases from their training data, manufacture spurious associations, and they remain susceptible to malicious corruption. Critically, contemporary AI models lack the capacity to reliably and reproducibly distinguish such biased, malicious or counterfactual training data from legitimate data. Moreover, they do not have the capability to reliably validate their outputs. The curation of their training data and the validation of their outputs therefore remain the responsibility of their human operators.