Project Veritas leaked bombshell documents yesterday concerning Google’s “Machine Learning Fairness” program, where fairness is defined by their social agenda. (Disclosure: I largely support Google’s social agenda, just not promotion of that agenda through mass, hidden manipulation).
In my recent article “What Journalists Need to Know About Artificial Intelligence” , I discussed the insertion of unintended biases into a given machine learning process through poorly designed training sets. For example, if an algorithm training to recognize whether a photo contains a human is not shown enough photos of European-descended people relative to individuals of other descent, the resulting algorithm may not perform well on future (and previously unseen) images of persons with European descent. This truth holds no matter which heritage is underrepresented in the training data set, just whom the bias impacts will vary. Remember when Google erroneously classified images of black people as gorillas? .
In the same article , I asserted that trained machine learning models prove difficult to audit, to find cause and effect pathways of decision making within.
Now suppose one wanted to purposely bias the future predictions of a trained machine learning algorithm, perhaps for social engineering. The easiest way to accomplish this is to intentionally bias the data set used to train the algorithm in the first place, so that outcomes you do not want to see cannot be learned due to under-representation in the input data. Therefore, these outcomes will not be produced during production use of the algorithm. In the “attention economy”, a small attenuation of certain classes of information, engineered in this way, will prove far reaching.
If outsiders try to “catch” a machine learning practitioner in the act of intentional bias manipulation, they will experience a severely difficult time determining potential cause and effect relationships by examining the trained model itself, as per the difficulty with auditing machine learning algorithms described above.
Easy-peasy. Social hacking so simple even a child prodigy can do it!