I am to be part of a panel at the Digital Leadership Forum event today discussing AI and the Enterprise. In my opinion, the AI debate has become dominated by the AI technology and the arrival of products sold to Enterprise as “AI solutions” rather than the ecosystems and contexts in which AI algorithms will operate. It is to this that I intend to talk.
It’s ironic though that we should come see AI in this way – as a kind of “black-box” to be purchased and installed. If AI is about “learning” and “intelligence” then surely an enterprises “AI- Baby”, if it is to act sensibly, needs a carefully considered environment which is carefully controlled to help it learn? AI technology is about learning – nurturing even – to ensure the results are relevant. With human babies we spend time choosing the books they will learn from, making the nursery safe and secure, and allowing them to experience the world carefully in a controlled manner. But do enterprises think about investing similar effort in considering the training data for their new AI? And in particular considering the digital ecosystem (Kindergarten) which will provide such data?
Examples of AI Success clearly demonstrate such a kindergarten approach. AlphaGo grew in a world of well understood problems (Go has logical rules) with data unequivocally relevant to that problem. The team used experts in the game to hone its learning, and were on hand to drive its success. Yet many AI solutions seem marketed as “plug-and-play” as though exposing the AI to companies’ messy, often ambiguous, and usually partial data will be fine.
So where should a CxO be spending their time when evaluating enterprise AI? I would argue they should seek to evaluate boththe AI product and their organisation’s “AI kindergarten” in which the “AI product” will grow?
Thinking about this further we might recommend that:
CxOs should make sure that the data feeding AI represents the companies values and needs and is not biased or partial.
Ensure that the AI decisions are taken forward in a controlled way, and that there is human oversight. Ensure the organisation is comfortable with any AI decisions and that even when they are wrong (which AI sometimes will be) they do not harm the company.
Ensure that the data required to train the AI is available. As AI can require a huge amount of data to learn effectively so it may be uneconomic for a single company to seek to acquire that data (see UBERs woes in this).
Consider what would happen if the data-sources for AI degraded or changed (for example a sensor broke, a camera was changed, data-policy evolved or different types of data emerged). Who would be auditing the AI to ensure it continued to operate as required?
Finally, consider that the AI-baby will not live alone – they will be “social”. Partners or competitors might employ similar AI which, within the wider marketplace ecosystem, might affect the world in which the AI operates. (See my previous article on potential AI collusion). Famously the interacting algorithms of high-frequency traders created significant market turbulence dubbed the “flash-crash” with traders’ algorithms failed to understand the wider context of other algorithms interacting. Further, as AI often lacks transparency of its decision making, so this interacting network of AI may act unpredictably and in ways poorly understood.