Last summer I took a course in AI, where part of the work was training models on data so they could learn patterns and make predictions. I did not expect it to explain what the news has been doing to everyone I know.
The goal when you train a model is that it learns the underlying rules in the training data, so it still works on new data it has never seen. But if you let it dig too deep into what it has already been shown, you run into a problem called overfitting.
It happens when a model gets too good at memorizing its training data. Instead of holding on to the general pattern, it starts memorizing everything, and loses the ability to tell signal from noise.
What you end up with is a model that performs like a Nobel laureate on its training data and like a distracted five-year-old on anything else. An overfitted model will happily seize on the fact that the letter e turns up more often than f, or that words ending in -tion appear on line 8 more often than line 9, and other things that mean nothing whatsoever.
If a model is going to be any use at all, you have to take overfitting seriously, because an overfitted one is, to use the technical term, useless.
Today it struck me that overfitting applies to something else entirely. News consumers.
The test that 800 people failed
A few years ago I sat in on a Gapminder lecture along with about 800 other people. We were given a quiz on the environment, population growth and poverty. Most of the room scored worse than they would have by guessing at random.
A lot of people put real effort into staying informed and still do not understand the world as it is. That is not bad luck. It is a systematic error, and it is overfitting.
To understand the world you have to see the long trends. How economies grow over decades, how technology develops, how violence and poverty and equality shift across generations. That is the signal, the underlying rules for how the world actually works.
News is something else. It fixes on random details instead of the larger patterns. Anyone who consumes news, at least without stopping to think about it, gets overfitted. You become a model fed everything, at the expense of what is relevant.
Most of those 800 people had probably been reading the news for years. They had memorized dramatic headlines, shocking images, disaster reports. Sitting there with their overfitted mental models, they were answering from noise rather than signal. That is how you end up doing worse than chance.
Cognitive regularization
In machine learning there is a fix for this, called regularization. The principle is that you deliberately restrict the model before it starts memorizing details it does not need.
Giving up the news is a form of cognitive regularization. You deny your brain the daily noise, and that forces your mental model back onto the signal, the slow but massive movements that genuinely shape the world.
The model comes back into balance. And you get to spend your attention on what matters.
What is overfitting, in plain terms?
A model that has memorized its training data instead of learning the general pattern. It performs brilliantly on what it has already seen and falls apart on anything new, because it can no longer tell signal from noise.
How can a human brain be overfitted?
By training it on years of dramatic headlines. You end up with a mental model built on memorable exceptions rather than underlying trends, which is why heavy news consumers can score worse than random guessing on basic questions about the world.
What is cognitive regularization?
The human version of a machine learning fix. You deliberately restrict what you feed your brain, cutting the daily noise so your mental model is forced back onto the slow trends that actually describe the world.
Do I have to quit the news completely for this to work?
No. The point is reducing the noise-to-signal ratio. Fewer headlines and more long trends, books and statistics will recalibrate your picture of the world, whether or not you go all the way to zero.
