A recent issue of New Scientist, a well-respected British periodical, led with an article entitled “The Coming Rise of Real Artificial Intelligence.” It was the latest in any number of articles and Wall Street reports touting the putative coming age of AI – machines that think.

But while AI and machine learning have roles to play in today’s world – and indeed play a big role in our own indicators – they aren’t ready, not by a long shot, for prime time, or to handle all that their boosters think they can. Computers calculate, they don’t think. They excel as long as a task can be broken down into a series of computations that can be done in a relatively short time frame. But even in games like chess, where machines routinely beat humans, there are anomalous positions that will cause the computer to miss a move that is obvious to a human player.

Ironically that same issue of New Scientist carried a briefer piece that reported: “AI tricked into seeing what’s not there.” There were no pictures, just the equivalent of about one and a half columns of text. But it was eye opening.

One of AI’s biggest accomplishments has been visual recognition: the ability to distinguish among human faces, different animals, and other visual tasks. But it turns out that these so-called deep-learning neural networks are over-programmed, making it far too easy to create what are dubbed “adversarial examples.” For example, if you overlay a picture of a panda with a bunch of random dots that a human would have to stare at to notice, the computer no longer can identify the picture as a panda.

Ditto for another AI triumph, voice recognition. Researchers at Facebook (Nasdaq: FB) inserted some random digital noise into various phrases that people were speaking to one another and having no trouble understanding. The computer, however, then completely garbled what was being said.

If this doesn’t trouble you, consider the following quote in the article from a Facebook researcher: “Image-recognition systems in driverless cars can be made to ignore pedestrians and parked vehicles.”

Clearly, no matter how sophisticated AI might be or appear, humans are still indispensable in overseeing machines – especially since we don’t know how to define the anomalies.

The studies on adversarial examples have reaffirmed to me that computers are calculating machines that are useful tools, not problem-solvers. When it comes to the role they play in our indicators, if I see that something anomalous is going on that computers can’t factor in, I won’t hesitate to act on it, even if it means going against our indicators.

This happened recently when we closed out a position very shortly after recommending it because we felt the heated rhetoric over North Korea would likely make investors nervous. However, we expect these kinds of situations will be exceptions, hopefully rare ones, and not the rule.

And as we reported previously, we have improved our indicators by using both weekly and daily signals. The weekly signals allow us to use economic numbers that are reported only on a weekly basis. But as evidenced by our ongoing search to continually improve our indicators, they still aren’t perfect. There will be losses, but we will learn from them.

Still, we now have enough confidence to start recommending more than one position at a time and to give positions more time to work out. We will probably limit our aggressive portfolio to no more than three or four positions at a time. If the indicators are all neutral, we’ll move to the sidelines.

We also have spread our wings in another direction, and with the recent recommendation of an option on Facebook (FB January 19, 2018 $180 call) have decided to include components of ETFs as occasional recommendations.

Currently the indicators on SPY are positive, while other indicators are neutral. Noteworthy is that for the first time in a long while the indicators both on oil, the commodity, and on oil stocks, are neutral. This could signal that oil stocks are close to a bottom.