How to use AI to help DEI: pitfalls, key actions, and checklist

Two common pitfalls

1. Bias in the data

The data used to train AI for HR solutions often contains significant biases: samples do not offer an accurate cross-section of local or (for multinationals) global society and tend to focus on white male candidates.

You must ensure that the data used to train AI systems is diverse in terms of age, ethnicity, gender, nationality, and any other relevant and identifiable dimension.

2. Lack of data

There are aspects of DE&I where employers’ ability to collect and hold relevant data is limited. One reason is the law: in many countries, it’s illegal to collect data relating to sexual orientation. And, where data is available, it’s likely to be based on self-reporting, which carries its own challenges and may be incomplete and (very) unreliable. Remember that these laws exist for a good reason – same-gender sexual activity is considered a crime in many parts of the world and, in some instances, is punishable by imprisonment, or even death.

Four key actions to ensure AI solutions support DE&I

1. Increase pressure on providers

Ask potential suppliers tough questions when procuring AI solutions for use in your HR systems. What are the guarantees that the issues of bias, hallucination, and error have been ironed out?

2. Get assurances regarding data

Extract assurances about AI training data. Chief DE&I officers or DE&I counsel should be involved in procurement and set-up discussions. (See checklist below.)

3. Improve governance via AI boards

Consider creating “AI boards” to oversee AI use, with members including seasoned and thought-leading DE&I professionals.

4. Understand how evolving regulation affects you

As regulation emerges, you need to understand how it will work in practice. In the European context, the EU AI Act – the world’s first comprehensive AI law – represents a significant change.

Checklist: your AI inputs and outputs

Key questions to ask include:

• Have we taken thorough steps to eliminate all kinds of bias right from the start?
• Is our input sufficiently broad?
• Does the output make sense?
• Is this what we were looking for?
• Does it reflect DE&I in the ways that matter to us?
• Is it aligned with our DE&I strategy?
• What – or, more importantly, who – is missing from these results?