How Can We Get AI to Keep Its New Year’s Resolutions?

AI can produce better answers than ever. The harder question is whether any of us will actually do anything with them.

By Bruce Lloyd, London in conversation with Frank Peters, Cleebourg.

“How do you get an AI to fulfil its New Year’s resolutions?”

The question came up during a conversation about the knowing-doing gap, and my immediate answer was: with difficulty.

Artificial intelligence has become astonishingly good at helping us think. It can analyse a problem, suggest alternatives, identify risks and produce a respectable list of things that ought to be done. What it cannot guarantee is that anything will actually happen.

That may turn out to be one of its most important limitations, because the problem is not simply the technology. It is the world into which the technology has to operate: a world of vested interests, competing priorities, imperfect information, organisational politics, fear, incentives and human beings who do not always behave in the way a beautifully constructed plan assumes they will.

Knowing is relatively easy. Doing is much harder.

When three prompts suddenly work.

Frank had just had a revealing experience of this.

A few months earlier, he had tried to build a booking system for meetings. Members would pay through one route and visitors through another, but both would eventually arrive at the same meeting. He used AI to help construct it and found the whole process so complicated that he abandoned it.

Then, several months later, he tried again.

The AI had improved. His own approach had changed. This time, after a few prompts, he had something that more or less worked.

That led to a perfectly reasonable question: why did the AI not know how to do this a few months ago?

Part of the answer must lie in the extraordinary speed at which these systems are developing. Something awkward a few months ago can suddenly become routine. But there is a more interesting question behind that one.

Suppose AI becomes very good indeed at telling us what should be done. What then?

Organisations do not create value merely because they know things. They create value because something useful happens, and making something happen takes us immediately into the territory of power, behaviour and incentives.

There are vested interests. People protect their positions. Some want perfection before anything is attempted; others prefer a series of small improvements. Some search for one glorious silver bullet while others suspect that progress usually comes from 101 ordinary things being done slightly better.

And then there are incentives.

I have long believed that distorted incentives played an important part in the banking crisis of 2008. People inside institutions could be rewarded for decisions that looked successful in the short term while creating enormous problems later.

So the useful question is not simply, “What is the right thing for this organisation to do?”

It is also, “What are the people inside this organisation actually being encouraged to do?”

An organisation can declare one objective while its pay, promotion, measurement and status systems quietly reward something quite different. That is one reason the knowing-doing gap interests me so much. We often assume ignorance is the obstacle: give people better information and they will make better decisions.

Sometimes they will.

Sometimes they already know perfectly well what should happen. Something else is pulling in the opposite direction.

Facts are not as innocent as they look.

Even information can become part of the problem.

Frank remembered a conversation with a supply chain manager who had told him, in effect, that if he wanted information supporting one interpretation, the necessary data could be supplied. If he wanted information supporting another interpretation, that could be supplied too.

The facts had not necessarily changed.

The selection had.

More data does not automatically solve that problem. Whenever somebody presents information, it is worth asking whether there is an interest in presenting it in a particular way. That interest may be obvious, hidden or not even fully conscious.

This is one place where AI can be useful, not because it is magically free of bias, but because it can help us interrogate what is in front of us.

What assumptions are being made? Whose interests are affected? What evidence would contradict the conclusion? What incentives are operating? What are we not measuring? Where might the feedback loop have broken down?

None of those questions guarantees a perfect decision.

But perfection may be the wrong standard.

Many of the difficult problems we face do not have perfect answers. They have useful answers, incomplete answers and decisions that have to be made in particular circumstances with the information available.

Which brings us to wisdom.

Wisdom is easy to quote.

Some time ago I tried to define wisdom as reliable, useful information that has stood the test of time and contains a values dimension: something that helps individuals or society function better..

There is a difficulty with that definition.

Wisdom is remarkably easy to recycle.

I can quote a commandment, a philosopher or a piece of ancient advice. An AI can do the same thing faster and at vastly greater scale. Neither activity guarantees that anybody will behave wisely.

There is a significant difference between wisdom and being wise.

Being wise requires putting wisdom into practice.

That distinction matters more as information becomes abundant and cheap. We can move from data to information and from information towards knowledge almost effortlessly. Ask about history, philosophy, management or almost anything else and an answer appears in seconds.

The difficult part begins afterwards.

Does that knowledge improve our judgement? Does our judgement affect what we do? Can we still act wisely when interests conflict, evidence is incomplete and there is no perfect outcome?

That is a much higher standard than possessing the answer.

At one point our conversation wandered into constitutional systems, courts, politics and trust. The Mayor stopped and asked whether we should really be discussing things about which neither of us was an expert and over which our direct influence was negligible.

I would put the question slightly differently.

Should we be talking about something if there is nothing we are ever going to do with what we learn?

There is nothing wrong with conversation for its own sake. Enjoying a conversation is a perfectly respectable use of time, and not every worthwhile discussion needs to end with an action point. Something learned today may become useful years later. An experience can pass through a community until eventually it reaches somebody who needs it.

But the question is still worth asking because time is limited. Are we acquiring knowledge merely so that we can have a more sophisticated discussion next time, or is there some point at which knowledge is expected to change our behaviour?

Learning when information is everywhere.

I was not a particularly compliant pupil at school. I reacted badly to quite a lot of what I was taught because I could not see the point of it, which was perhaps an odd beginning for somebody who eventually became a university professor.

The experience stayed with me.

History and the classics, for example, can easily be taught as objects sealed in boxes: learn these dates, remember these names, reproduce this information.

The more useful question is far more interesting.

What can we learn from all this accumulated human experience that might help us with what comes next?

AI changes the educational landscape because access to information is no longer the main difficulty. A student with a phone can obtain explanations, comparisons, summaries, images and arguments almost instantly. That does not make education redundant. It changes what education ought to be doing.

When I taught postgraduate students at South Bank University in London, I was fortunate to have relatively small groups. We could assume that students had access to information and then ask what that information meant inside their own organisations. What was relevant? What were its strengths and weaknesses? What additional questions should they ask?

That, to me, is much closer to learning.

I still remember, with some horror, a university lecturer from my own student days who sat in front of the class with a textbook, read from it and expected us to write down what he said. Even then it seemed a rather strange way to learn. In the age of AI it becomes almost impossible to justify.

A teacher’s role cannot simply be to transfer information. It is to create curiosity, enthusiasm, discussion and the desire to know more.

That does not mean learning has to become entertainment. There is a danger in measuring teaching through popularity because the easiest way to improve the score may simply be to become more entertaining.

But enthusiasm matters. Interaction matters.

And usefulness matters.

The booking system meets human beings.

Near the end of our conversation, we returned to the Mayor’s booking system.

He had spent hours trying to make payments, confirmations and meeting registrations work properly. Eventually he and a colleague began testing it by sending bookings back and forth.

Each test exposed something else.

What if somebody cancels? What if they change their mind? What if they want a refund? What happens if they alter the booking?

The original system had been built around a perfectly reasonable best-case assumption. We offer a meeting. Somebody books it. They pay. They receive confirmation. They attend.

End of story.

Except that this is not how human beings work.

The AI had not thought far enough ahead.

Neither had the human being asking it.

That may have been the most useful lesson in the entire conversation. The real failure was not PayPal, and it was not really AI. The system had failed to take sufficient account of the human dynamics of what it was trying to do.

What made it better was feedback.

Something was built. Somebody tried it. Something unexpected happened. The system was changed and tested again.

There is nothing glamorous about that. It is not a silver bullet.

It is a feedback loop.

I suspect a great deal of useful AI will develop in exactly this way. Not because the machine produces perfection at the first attempt, and not because the human always knows the perfect question to ask, but because each can help the other reveal what has been missed.

We will still need judgement. We will still need incentives that encourage the behaviour we actually want. We will still need organisations capable of listening when something goes wrong, and people willing to change course when experience shows that the first answer was incomplete.

Perhaps that is the closest thing to a workable New Year’s resolution for AI — and for the rest of us.

Keep learning. Use AI to help us learn better. Make the best decision we can in the circumstances.

Then do something.

And pay very close attention to what happens next.

That may not be perfect wisdom.

But it is probably much closer to being wise.