← Blog

Intelligence amplification: AI that helps people decide

Most of the AI that pays off in business helps people make better decisions. In this post, I explain intelligence amplification, an idea a professor shared with me years ago, and why it’s still the most useful way to think about AI at work.

A conversation with Yoav Freund

I learned the term from Yoav Freund, a professor at the University of California, San Diego, who shared the 2003 Gödel Prize for creating the AdaBoost algorithm with Robert Schapire. He’s kind, easy to talk with, and passionate about his work.

We were talking about the progress AI had made in medicine. I mentioned how well AI was detecting the spread of breast cancer on pathology slides. Google’s model, called LYNA, scored 99% on a standard measure of detection accuracy in a 2018 study, as VentureBeat reported.

Then I made a leap a lot of us make. Maybe, with a little supervision, computers could take over this kind of diagnosis entirely, and maybe replace the specialists who read the slides.

Yoav pointed out what that view misses.

Intelligence amplification

It can feel like AI has only two futures: it takes all our jobs, or it makes our lives effortless.

Yoav told me about intelligence amplification, usually shortened to IA, which is AI spelled backwards. It even has its own Wikipedia page. The idea is to use technology to make people’s thinking stronger. In medicine, that means helping doctors work faster and catch more. The researchers presented LYNA the same way: as a tool to help pathologists, who make the final diagnosis. Its name says so too. LYNA stands for Lymph Node Assistant.

We use intelligence amplification every day. Can’t remember the Spanish word for cat? Search for it. Gato. That’s assisted recall. Voice assistants work the same way. They extend what the human mind can do.

How this applies to business

The most impressive AI systems need enormous amounts of data. Speech recognition is a good example. At 90% accuracy, one word in ten is wrong, which is unusable. At 95%, it’s one word in twenty. Getting from one to the other takes far more data and much better algorithms.

Google’s speech recognition went from a 23% word error rate in 2013 to 4.9% in 2017, according to VentureBeat.

Line chart of Google's word accuracy rising from about 80% in 2013 to 95% in 2017, reaching the threshold for human accuracy
Bar chart of Google's speech recognition word error rate falling from 8.5% in July 2016 to 4.9% in 2017

Most companies will never have data on that scale. They don’t need it. The models that medium and large companies rely on every day, like predicting which customers will leave or recommending products, are lighter-weight cousins built to help people decide. That’s intelligence amplification.

Say a salesperson has 5,000 leads to sort through. A model that ranks them and hands over the 100 most promising helps that person decide faster and with better information. Netflix works the same way. It doesn’t pick one movie for you. It narrows thousands down to a handful, and you choose.

Why it still matters

The black-and-white view of AI, either a threat that will take over or a cure for everything, falls short. Intelligence amplification treats AI as a tool that helps us do better than our minds alone, and keeps improving as we need it to.

Today’s AI assistants make this more relevant. We build every system so that a person on your team approves any consequential action before it happens. You can see why in what Bennett ONE does. The same idea runs through why explainable AI matters: people trust tools they can understand and check.

Let's talk about what you want to improve

Talk to an engineer