Math and Smart Software Ethicality
July 5, 2020
I noted “Mathematical Principle Could Help Unearth Unethical Choices of AI.” The idea is that a numerical recipe runs when the smart software developer trains the model or lattice of models. The paper states:
Our suggested ‘Unethical Optimization Principle’ can be used to help regulators, compliance staff, and others to find problematic strategies that might be hidden in large strategy space. Optimization can be expected to choose disproportionately many unethical strategies, an inspection of which should show where problems are likely to arise and thus suggest how the AI search algorithm should be modified to avoid them in the future. The Principle also suggests that it may be necessary to re-think the way AI operates in very large strategy spaces, so that unethical outcomes are explicitly rejected in the optimization/learning process.
Several observations:
First, does the method “work” in the murky world of smart software; that is, some smart software is designed specifically to generate revenue. The “training” increases the likelihood that the smart software will deliver the results; for example, increased ad revenue.
Second, what happens if the developers and subject matter experts ignore the proposed numerical recipe? Answer: The algorithm will perform based on the training it receives. The purpose of the smart algorithm is to deliver what may be to some an un-ethical result.
Third, what if the proposed numerical recipes itself identifies an “ethical” action as “un-ethical”?
To sum up, interesting idea. Some work may be needed before the cheerleading commences.
Stephen E Arnold, July 5, 2020
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
