A chain is as strong as its weakest link. And in AI, often the weakest link is the user

AI used to be the preserve of university research labs and tech businesses. Today, AI is accessible to anyone with a phone, from your children to your grandparents. And that’s both an opportunity and a problem.

Usage does not mean understanding

Despite the apparently simple facade, AI is hugely complex technology that looks familiar but isn’t. In some ways it seems like a person. In others, like a computer program. But it is different, in fundamental ways. When we talk about a human, we assume “common sense” and certain values. When we talk about a computer program, we assume something that is repeatable and consistent.

The dangers of familiarity

Already we see issues where people confuse talking to an AI with talking to a human. An AI is mimicking a human, it is not like a human. Apart from anything else, it lacks “common sense”. You can get sensible answers for a long period and then something completely left field. What’s more, the AI may then build upon that incorrect assumption with perfectly sound logic, which makes discovering the error quite difficult.

AI users need a “driving test”

You would never put a new driver into a Ferrari and tell them to just go, or a trainee pilot into a plane and ask them to take you on holiday. But that is exactly what we are doing with AI. We roll out tools like Copilot and just assume people will find ways to use them that add value. They probably will. But they may also fall into the traps.

Key things all users must know

AI answers are not 100% reliable. You have to check them independently. An AI will give you the wrong answer as confidently as it gives you the right one. 
Unless you are using an approved internal AI, assume everything you put in is going into the public domain. So if you put in your company’s sales data, your health records or the business plan, that information is almost certainly going out of the country and if you don’t have an explicit contract saying what can happen to that data, it may be used to train future models, and fragments could resurface in someone else’s answers.
AI won’t always do what you say, particularly agents. If you install an agent and give it access to your email, your contacts and your files, it may do a good job of tidying things up. But it might just as easily try to access data on your corporate servers. Users need to understand that anything other than formally approved tools may not be used. It’s like going onto the web to download “free” versions of popular software. It may work but it may also be malware used by criminals to harvest all your private data.
If you use AI to create a report and the information is wrong, you (not the AI) are responsible. You cannot sue an AI. It is just a tool. You need to take personal responsibility for everything you publish, whether you used AI to create it or not.

Training is not optional

We all think we understand AI. It is useful. It makes us more productive and gives the impression we are more capable than we are. That is the appeal – and the risk. But every company needs to roll out training on AI. And the more the AI is made accessible and its use encouraged in the business, the more that training is essential. Pair it with a clear acceptable use policy and approved tools people actually want to use – ban AI without an alternative and its use simply goes underground.

Train the right people

Training often starts at the top. The senior management, the Board. But the people that need it most are those that have access to sensitive data. Often that is the most junior staff. The bank teller. The practice nurse. The school assistant. The store manager. These are the people with access to the latest raw data – arguably the most valuable data of all. These individuals are also often highly incentivised to be able to present information in a compelling way and have found that AI is very good at this.

Make the training easy

Few people are going to sit through two hours of AI training (or if they do, they are unlikely to remember much of it). Use tools that break training into short blocks, like TikTok videos, and test at the end that people have understood what they have learned. That way people will engage and deliver the value that is required. Keep a record of who has completed what – it is exactly the evidence a customer or regulator will ask for.

Today, AI is accessible to anyone with a phone, from your children to your grandparents. And that’s both an opportunity and a problem.

Make it frequent

AI is developing at breakneck speed. Whole new technologies are appearing monthly. Most of them are accessible for free over the Internet. Training needs to keep up, which further reinforces the idea of “little and often” rather than one long training session.

Easy access to AI means many people are using it without fully understanding the risks, and this is an issue, particularly in business
Training is non-negotiable and becomes even more important for junior staff who have access to raw confidential data
Make sure that training is accessible for best results. Regular 2-3 minute videos beat a half-day formal training session every time.

To explore further

Take a look at our Resources page with many  other helpful articles on AI and governance

Try our tools and solutions

Discover Akili’s tools and solutions designed to help you understand,
assess and govern AI with confidence.