Today’s expert panel on Data Ethics took a fascinating turn: to consider what a healthy relationship between human and AI would look like. Although we tend to discuss characteristics and affordances of technology, proper use of technology depends on the human side of the partnership, too.
When choosing or using any tool that uses AI, we must remember that it is a tool. A complex one, perhaps, but not significantly more so than a motor car or a medium-sized organisation. To work safely and effectively with those, we need to understand their capacities – what they are and are not suitable for – and be able to detect when things are going wrong. At least that level of understanding should be possible with any AI: vendors who claim “black box” (or “magic”) status are damaging both their product and their customers. For some applications – for example to understand what a student needs to do to improve their predicted performance – we may, of course, need products that let us understand more.
That understanding shouldn’t be limited to what information the AI takes as input; we also need to know what inferences it makes and how it uses them. Those may well be more sensitive for individuals than the raw data, and more damaging for groups and society. Even if an individual is happy to release their own data, we need to understand what impact that might allow the AI to have on others with similar characteristics or behaviour patterns. It occurs to me that this is yet another reason not to base such processing on consent: we can hardly expect individuals to consider the impact on others when making their own personal decisions what to disclose. Privacy is a collective endeavour.
When building an understanding of AI, our human view of the world is unlikely to be a good starting point. Despite analogies such as “Intelligence” or even near-human forms, we are dealing with computers. It occurs to me that science fiction is littered with puzzled robots: we need to remember those and expect our AI’s to seem as weird to us as we seem to them. They don’t know what a student is, or an exam result. It’s all just numbers. They can – indeed probably should – let us know whether the situation in which we place them is one they were designed for. John Naughton calls for machines that know when they don’t know. But in education, maybe a machine that is certain should also raise human doubts: many of us in the working group had followed career paths that might well have been ruled out by a school AI.
Achieving this seems to require improved awareness all round. If users of AI need more awareness of how much their suggestions can be trusted in different circumstances, then vendors of such systems need to be aware of this need and provide documentation and tools to help. And perhaps those procuring AI systems need a deeper level of awareness of the questions to ask and information to expect since – as was pointed out – once you’ve paid for it, it’s hard not to use it.
Many thanks to all those involved – this “catalyst” was definitely changed by the experience.