Q&A with Fourthline’s CTO, Ralph Post

How prevalent is AI human bias? 

It is common enough that anyone using AI to make decisions about people should be checking for it. There is no single figure that tells you how widespread it is. Bias can come from the data, the people labelling it, what the system is asked to optimise, or how its output is used. So ‘human bias’ is only part of the picture. 

What makes it potentially damaging? 

For the person affected, it can mean being unable to open a bank account, having to verify their identity several times, or waiting longer than other customers. For the business, that means losing legitimate customers, creating more support work and damaging trust. AI can repeat the same mistake on an enormous scale. And if you only look at average accuracy, you may never see which customers are having a bad experience. 

What’s the best way that a CIO can correct AI for human bias? 

Start with the outcome you want to prevent. Are legitimate customers being rejected? Are some groups consistently having to make more attempts? Measure those differences and make someone accountable for fixing them, with the authority to stop a release if necessary. Then investigate the cause. It could be the data or the model, but it could also be how the service is designed. The CIO needs to make this part of how products are built and run. 

What makes this approach effective? 

It makes the findings matter. A report accomplishes very little unless someone acts upon it. Looking at the whole customer experience also helps you fix the right problem. In identity verification, for example, poor image capture can cause some customers to fail more often. Improving how a photo is taken may solve something that further model training would not. 

How can this knowledge be shared with relevant AI and IT team members? 

Use real examples and make the lessons part of everyday work. Show teams what went wrong, who was affected, and what needs to change. Then turn that into tests, clear guidance and checks before release. Bring product, engineering, risk and customer support together to review those cases. People learn more when they can connect a technical result to someone’s actual experience. 

What’s the best way to detect AI human bias? 

Look beyond the overall accuracy score. Test the full process with data that reflects the people who will use it and keep checking once it is live. Who gets wrongly rejected? Who gets referred for extra checks? Who gives up halfway through? Compare those outcomes across relevant groups and conditions, while being honest when the sample is too small to draw conclusions. A difference is a reason to investigate. It does not, on its own, explain the cause. 

What’s the biggest mistake CIOs make when addressing AI bias?  

Assuming it is something the data science team can take care of on its own. Engineers can improve a model, but leadership has to own decisions about which errors are acceptable and who bears their consequences. Adding a human reviewer is not an automatic solution either. That person needs enough information and authority to challenge the system, and their decisions need to be checked too. 

Is there anything else you would like to add? 

I would be wary of anyone promising ‘bias-free AI.’ What matters is whether you can show how you find problems, reduce them, and help customers when things go wrong. That gives buyers confidence and investors a stronger basis for believing the business can grow without accumulating any hidden problems. This should be part of running a sound business. The competitive advantage comes from delivering a better experience and being able to prove it.