By Alastair McAulay, PA IT and cloud strategy expert and Fred Johnsen, PA AI and automation expert
Back in 1968, Stanley Kubrick's sprawling masterpiece '2001: A Space Odyssey' was one of the first movies to have a computer in a leading role. HAL (shift each of the letters IBM one step backwards) is the artificially intelligent computer that manages the spaceship in the best interests of the mission.
A key point in the plot is when one of the crew needs to override the computer so they can survive. “I'm sorry Dave…I can't do that” is HAL's chilling reply. The automaton has encountered a situation no-one anticipated. And the logical AI response contradicts the ethical response.
Science fiction nonsense? I'm afraid not.
When Microsoft unleashed its self-learning chatbot 'Tay' onto Twitter in 2016, in less than 24 hours it was spouting racial and sexual abuse. It had been thoroughly corrupted by the way people use Twitter. Autonomic systems can't be left to their own devices. As systems train themselves they need supervising.
You need to think about the data they'll use for training, and the goals you set them. Much like you have to ease new hires into the business, you have to implement autonomic systems gradually: a security monitoring system will need to understand what patterns of ordinary user behaviour occur during the day, then during the evening and finally over the weekends before it can be trusted to flag potentially errant behaviour.
The problem is that the artificial intelligence built into the system isn't formally specified as a set of logical steps. Many autonomics systems train themselves by analysing huge amounts of data to discern patterns and build their own rules from these patterns.
CIOs need to make sure people in the organisation understand how AI systems are being used in the business. They need to be confident that the AI systems not only adhere to a defined code of conduct, but that they're capable of adhering to it. That means:
- testing autonomic IT systems enough to earn the trust of employees and customers
- managing expectations – no matter how well you test the system there will always be some flawed decisionsas the AI system tries to infer the best course of action. Therefore people must always be able to override or intervene
- auditing systems regularly – to make sure the reasoning behind decisions and the reliance on different data remains relevant long-term
- building rules into systems – to make sure the AI is using data fairly, ethically and complying with regulation. You need to manage security risks when combining and making decisions on data from different sources. For example, you'd need to check your AI isn't using information about employees' religion or gender to make decisions.
The ethical behaviour of a company will be heavily influenced by the design of autonomic algorithms delegated to automate business processes. The Microsoft 'Tay' chatbot was obviously going rogue, but other AI might misbehave in a more subtle way. You need people in the IT team who have emotional intelligence and customer empathy, rather than the traditional process orientation to provide appropriate governance and oversight of AI systems.
Today's CIOs should start working out whether their organisation has the right mix of emotional, philosophical and technical skills to assess the ethical implications of using AI automated systems. That should help get the balance right between people making decisions or machines. And avoid “Computer says no” at a potentially catastrophic moment.
Artificial intelligence (AI) and automation: is your IT department getting left behind?