AI lacks ambition, so stop fearing a machine takeover | Ben Fay

Ask many people what worries them about artificial intelligence, and you will likely hear some version of a sci-fi movie storyline: The machine wakes up one morning, works out that people are the obstacle and starts making its own arrangements. I've worked with these systems for a living, and I would like to offer a less cinematic narrative.

To start, your AI has no ambition. It wants nothing and prefers nothing. Close the window and it is not waiting for you, and nothing in there resents being switched off or plots what to do next.

That should settle the Hollywood question. But it does not settle the real one: What is AI? We pile a lot into the word “intelligence,” including information processing, reasoning, grounding and volition.

These systems have the first two in spades. They are prediction engines trained on a staggering volume of human writing, and they produce the next plausible word, over and over, faster than any person could. They will draft your proposal, summarize a 200-page manual in seconds, answer a customer at midnight, and even write software from scratch.

What they do not possess is grounding in reality (AI connects symbols to symbols, not to anything in real life), which means they also lack the judgment to know if anything they made is true or accurate, nor do they care one way or the other.

This concept can be particularly difficult for us to internalize, since for all of human history we have associated articulate speaking with high intelligence and even wisdom. Now we have to accept a new truth: Just because AI talks well does not mean it actually knows what it is talking about.

Another critical capacity missing from AI is volition, which is the ability for it to generate its own purpose, and decide on its own what it should do. This means that every objective one of these systems pursues originally came from a human. Every permission it holds was handed to it by a human.

When one causes harm — and some have — the chain runs back to somebody's decision every time. In criminal law we do not put the gun on trial. We ask who was holding it. This is comforting news because the root of the problem moves away from science fiction and into the management of people, something we have ample systems and experience for.

Here are three ways that businesses are approaching that problem today:

The first is the ban. Several organizations simply prohibited AI outright. It never works. It relocates the use to personal phones and home laptops, where there are no logs, contracts or supervision. You have not reduced the risk. You have blindfolded yourself to it.

The second is the opposite: enthusiastic, widespread adoption with not one written rule about what may be typed into these tools. That is where most small and mid-sized organizations sit today, and it is the posture that produces the viral stories we've heard continuously over the last year.

The starting point of the more prudent, middle path does not necessitate forking out piles of cash on technology or consultants. Decide which tools are approved. Write down in plain words what may or may not be entered into them — customer records, payroll, anything under a confidentiality agreement.

Make it known that each person using AI is responsible for the output as if they made it themselves but also offer leniency for those who make a mistake and are honest about it. Then make one person an owner of this new policy and set a date on the calendar to revisit it. None of this requires a massive budget, but many organizations have yet to take the time to do it.

The bottom line: The machines are not coming to take over. What is coming is a long run of small decisions, made by normal people, about what we hand these tools and what we verify before we trust them. Every one of those decisions takes judgment, something that people — not machines — are still best suited to supply.