Can We Make Artificial Intelligence Accountable? Posted on : Sep 20 - 2018

Lack of explainability of decisions made by Artificial Intelligence (AI) programs is a major problem. This inability to understand how AI does what it does also stops it from being deployed in areas such as law, healthcare and within enterprises that handle sensitive customer data. Understanding how data is handled, and how AI has reached a certain decision, is even more important in the context of recent data protection regulation, especially GDPR, that heavily penalizes companies who cannot provide an explanation and record as to how a decision has been reached (whether by a human or computer).

IBM may have made a major step towards tackling this issue, announcing today a software service to detect bias in AI models and track the decision-making process. This service should allow companies to track AI decisions as they occur, and monitor any ‘biased’ actions to ensure that AI processes are in line with regulation and overall business objectives.

If this software can truly explain the decisions taken by even the most complex deep learning algorithms, this development could provide the peace of mind that many companies need before unleashing AI on their data.

Breaking bad decision paths

‘Explainability has been a big focus for our research’ said Jesus Mantas, Managing Partner at IBM Global Business Services, in an interview with myself earlier today, and this software has grown out of that research. By measuring IBM’s predicted decisions against the actual decisions taken by an AI program, including the weight it gives and the confidence it has on that decision, the software can theoretically figure out whether the algorithm is biased and determine the cause of that bias.

This could allow companies to prove their compliance with data protection regulations by tracking how an AI program uses its data, and make sure sensitive results are not compromised by a biased model. Companies can also set their own decision parameters to track, so that flawed decisions do not affect either business objectives or regulatory requirements. View More