Improving AI integrity

Artificial Intelligence can basically be classified into two branches, one that uses established algorithms to help make decisions (but no learning also called reactive AI machines that work on fixed algorithms) and the other which learns from the data supplied to it to help make intelligent decisions (also called narrow artificial intelligence, machine learning and deep learning etc.).

A hard fact we need to understand is that an AI is only as good as the data it is fed on or tested called the learning data and test data, usually in the ratio of 3:1 ... If there were humans on Mars, the AI which works with earthlings may not be effective with Martians, so simple .. The AI which was developed was fed on data on this earth. As such since we do not know how Martians (if they exist) would react to different stimuli, we cannot develop an AI systems that is all encompassing and universal.

Since AI is getting to be very widely used across domains and functions in the future, we need to figure what types of training data was the model fed on and for how long, what was the testing data the model was tested on, who supervised the learning etc.

Can faulty inputs to the model lead to faulty decisions ? It is very much possible, deficient learning data set or testing data set or an oversight could have corrupted or influenced the final AI based decision that was made. Have we ever thought what would be resulting implications of a faulty decision by AI ?

For example if we look at the pharma industry, US Federal Drug Administration, the most toughest pharma testing and passing authority in the world, does Randomised Control Trial experiments many a time in different settings before finally accepting or rejecting the  drug or molecule.would be then released for the public consumption.

The randomised control trial (RCT) is a trial in which subjects are randomly assigned to one of two groups: one (the experimental group) receiving the intervention that is being tested, and the other (the comparison group or control) receiving an alternative (conventional) treatment                                                     - www.bmj.com

 If the RCT experiments have worked well for the global pharma industry, ensuring integrity, reliability and accountability in the process, why should it not work for the AI industry to check for the integrity, the utility, reliability,  completeness and applicability of the software before it is released for the consumption to the people of world. 

If nations can collaborate in this direction, or UN taking the lead in this case, we can avoid grave mistakes and accidents that would befall the world because of traps, drawbacks and shortcomings in the software that have escaped the eyes of the regulator. 

This article in part derives its origin to the article in HBR of June 2021 titled  We should test AI the way FDA tests medicines by Carissa Veliz.

George..