Six types of AI bias in HR... And what you can do to counter them

Whilst AI brings many efficiency benefits, it also comes with risks that should be approached with caution. AI bias is one of them, and we explore what this means in this week’s blog.

From stock market IPOs to the threat of advanced computer hacking, AI is never out of the news. It is a big thing in businesses generally, and HR specifically, too: with AI automation being used for a whole host of tasks, often poorly.

Creating and placing job adverts; CV screening; candidate scoring; production of HR documentation such as employment contracts; and procedural communication, for instance in a disciplinary or grievance process.

For all its efficiency, AI has a number of downsides: data protection, human redundancies, inaccuracy and hallucination, and what we're going to focus on in this blog: bias.

Types of AI bias

It is known that bias largely occurs in AI because the models are trained on vast datasets from the real world, which contain long-established biases themselves. These biases can then manifest in different ways depending on how the AI is used.

So let's start by summarising six different types of AI bias that can impact HR decisions.

• Aggregation bias -This is when a single model of AI is applied to a diverse group of people assuming that everyone is the same, when, of course, they are individuals – our differences matter.
• Automation bias -This is when you place too much trust in the AI, assuming that everything it says is true. In fact, you should be reviewing and challenging its output.
• Confirmation bias -This is when the AI tells you something that you want to hear and reinforces a bias already within you.
• Label bias -This is when subjective labels are fed into an AI training model based on discriminatory principles; for example, that men are better at leading than women.
• Measurement bias -This is when the metrics by which you judge someone do not actually reflect their worth to you as an employer.
• Representation bias -This is when the real-world cohorts whom you are assessing are not reflected accurately in the AI’s training data.

Combatting AI bias

The genie is out of the bottle with AI, and it is here to stay. If you do decide to use it within your HR function there are a number of safeguards which it would be wise to adopt.

As an overarching principle, start with accepting that you (as employer) maintain accountability for the outputs that AI produces. It will be no defence if things go wrong to say that it was the computer’s fault. With this mindset, be rigorous when making decisions about adopting AI – how do you justify it? By acknowledging that you are on the hook, it should help you take it more seriously.

Getting more specific, human oversight of AI systems is essential. The speed in which AI can complete tasks means that biases can be amplified significantly. A system which puts a human review into the process can help identify when bias is occurring and correct it.

Know when not to use AI. Relying on it for legal advice and documentation like employment contracts or redundancy is a minefield.

It may draw on information that is from UK law but outdated (employment law changes all the time), or from a jurisdiction that has no relevance to UK law; or convincingly make something up. The cost to you if it does this and you are taken to an employment tribunal could be eye-watering.

Professional help

For ensuring your HR practices and documentation are legally compliant, it is much safer to seek professional advice. We offer a practical solution, always telling you what you can do, not what you can’t. We can provide you with up-to-date documentation as well. All for an excellent-value retainer under our Advice Line package. Contact us to find out more.