Eliminating Bias in Algorithms
Artificial intelligence is basically the execution of algorithms or formulas that tell a program what to do, when to do it, and how to do it.
At the most basic level, an algorithm follows very simple rules. Examples might be balancing a checkbook or adding a column of numbers and presenting a total. At a more advanced level, it can navigate you to your destination, tell you the speed limit, and what time you will arrive. It will process highly complex strings of directions containing hundreds of rules and steps. Some algorithms are written by humans but more and more they are generated by the software itself. These are called Learner algorithms and this is where problems arise.
Back in 2015, Amazon was using an in-house developed algorithm to rate potential software developers. They soon realized that the algorithm was far from gender-neutral in rating women.
This was caused because the algorithm was using historical data on who had been hired, and the vast majority of applicants had been men. The algorithm “assumed” this was desired because of the frequency of its occurrence and then used gender as a screening criterion.
An algorithm is only as good as the data it has access to. According to MIT, less than 1% of all data is currently being analyzed. And even less of the data organizations have accumulated about employees is accessible to be analyzed. If Amazon’s algorithm had had access to a larger and more diverse database, the results might have been different. But, because men have dominated software development and other technical fields, it may have reached the same result.
Very few algorithms have intentional bias built-in. The bias is the result of past human behavior being mimicked, so to speak, by the program. If more conscious thought and understanding had gone into the defining the output desired, the algorithm could have been instructed to include women or minorities.
Bias can be prevented in algorithms but it takes thoughtfully defining the desired output and understanding what is inherent in the dataset being used by the algorithm.
The articles below provide some insight into how to develop more objective algorithms and how they become so biased.
In this article, discover three tips to leverage AI in support of your diversity goals and ensure AI neutralizes–and does not perpetuate–human bias.
Artificial intelligence is supposed to make life easier for us all – but it is also prone to amplify sexist and racist biases from the real world
Where do gaps exist in your training data? Good advice on preventing bias
From job ads to salary offers.
About This Newsletter
Hand curated articles, videos, podcasts, and other media on the future of work, talent, recruitment, and learning. If you find this useful, please share on Twitter. You can always reach me at firstname.lastname@example.org. I’d enjoy your comments - positive or negative. Send me an email and let me know what you think.