@mjd
... that trained NNet Models can't provide explanations or audits (subtly different but overlapping the more usual 'non deterministic' complaint), which is still an indictment for decision support, finance, HR applications of this 4th generation of perceptrons.

The several new ethical issues with the economics of omnipresent Large deep general purpose core models ( incl. externality costs) remain.

(small local models don't pose fresh externalities but may be derived...)