Machine learning and artificial intelligence have made huge strides in the recent past, from self-driving cars to Watson as Jeopardy champion to computers that can compose classical music and beat us at chess. But do any of these notable achievements mean that machines can anticipate our behavior online?

Humans aren’t predictable.

Not yet anyway. In fact, one of the biggest — and most enduring — challenges facing predictive modeling is that humans aren’t, well, very predictable. Compared withthe determinism of machines, we humans are endowed with generous amounts of free will.

And that means we’re not the totally rational creatures that traditional economic theory makes us out to be.

Predicting Religion but Not Running Shoes

For example, I am quite athletic, enjoy sports and engage in physical activity most days. But just because I consume content or buy products related to exercise and fitness doesn’t mean that a machine — or a human for that matter — could predict exactly when I will be ready to buy a new pair of running shoes or how advertising will influence me when I do.

That challenge becomes all the more tantalizing because some major aspects of our behavior turn out to be surprisingly easy to predict. For example, a 2012 study of over 58 thousand Facebook users showed that likes can predict religion, sexual orientation and numerous other personal traits and attributes with near certainty.

Buyers Don’t Resemble Clickers

Yet when it comes to predicting online shopping behavior, my team’s research into online display advertising suggests that buyers don’t resemble clickers much at all.

So although I may have researched running shoes online, my actual purchase decision could be impacted by many unpredictable factors: I may have purchased a new pair at the running store around the corner just before sitting down at my laptop. I may have sprained my ankle hours earlier. My boyfriend may have just given me a pair as a gift.

Since we humans can’t even predict these incidents ourselves, how can a machine?

Is Predictability Really the Goal?

But when it comes to online advertising, it’s important to establish right upfront whether predictability is really the goal. Machine learning research has shown time and again that while predictability can be used very effectively to identify target audiences in advertising or to recommend products on Amazon or Netflix, data can get noisy and errors of all sorts can occur.

Consider the click, for instance. Used originally as a campaign performance metric, the click has become problematic, particularly when machines start predicting clicks for every event and optimizing ad placement based on them.

Most Predictable = Least Valuable

But predictability and value to the advertiser do not necessarily go hand in hand. Clicks happen for all kinds of reasons and the most predictable clicks are often the least valuable. Fraudulent clicks, for example, often follow very predictable patterns, precisely because machines, not humans, are doing the clicking.

In fact, according to a recent presentation by digital ad fraud researcher Augustine Fou, every single ad shown on ghost or fake websites (with names such as and was clicked on by something — but not necessarily someone. Clearly, optimizing for bot-generated clicks on websites like these would be decidedly counterproductive for an advertiser.

Accidental Clicks Happen

There are also less nefarious scenarios, of course. Clicks can happen accidentally when users get distracted, aren’t computer-savvy or cannot see very well. For instance:

  • We see huge numbers of clicks on ads shown on mobile flashlight apps — no matter what the advertised product. But it would be wrong to conclude that these apps are great vehicles for advertising since it’s more likely that users are fumbling in the dark and accidentally clicking on ads as they search for buttons to turn the flashlight app on or off.
  • Certain mobile games have similarly high click rates but not necessarily because the user cares about the advertised products. We’ve found that often toddlers have been given the devices for entertainment and their fine motor skills (along with their understanding of digital display advertising) are still very much lacking.
  • Ads that interrupt the flow of a game until the user takes action also receive predictable clicks. From internet providers to credit cards, our research shows that close to 30 percent of these ads get clicked on, yet actual sign-ups are virtually zero.

Good Clicks, Bad Clicks

For reasons like these, it’s fair to say that the factors that make non-fraudulent clicks predictable are often more about the circumstances in which an ad is shown than a reflection of user interest in the product itself.

Bottom line: Since machines on their own cannot know a good click from a bad one, they’ll continue to optimize toward the clicks that are easiest to predict. Thus, it’s vital to remember that a thousand clicks don’t necessarily equal a thousand interested consumers.

In fact, our research has shown that models that optimize for clicks perform no better than chance with respect to predicting actual purchases.

Can Machines Make Moral Choices?

While being led astray by machine-generated online advertising strategies can be costly and disappointing to a brand, the same mechanisms can also govern user interactions where there is more at stake.

Since machine learning has a tendency to zoom in on whatever predictions are the easiest, those predictions risk not being consistent with a user’s cultural or moral values.

Algorithms Need Common Sense

We are still early in our understanding of what good and bad predictive modeling and artificial intelligence will bring to the table in the many areas that are exploring its promise — from advertising to medical diagnosis to predictive policing.

Inevitably our algorithms will improve as we teach machines to predict and avoid false and fraudulent clicks. For the foreseeable future though, it’s a good idea to balance predictive modeling with human intelligence, intuition and common sense.

Title image by Dmitry Ratushny