@FBM,
For one thing, the article overstates the actual results quite a lot. From the graphic provided in the link, it appears that the prediction is never better than 58% or so, which means that it failed 42% of the times. This is only slightly better than the null hypothesis which would be a success rate of 50%. That big ERM+computer apparatus only succeeds slightly more often than flipping a coin.
One would have to read the published article to verify that the difference was statistically significant and at what level of alpha risk. However, I assume there is some decent level of stat significance, or the article would not have been publishable.
Another issue is that the choice in question is a very simple, binary one: left or right, and that it has no particular consequences. No moral quandary of any sort, no risk involved. Most life choices are far more complex than that and would thus involve a maze of brain signals that would be impossible to read through an ERM, at least with current technology.
A third issue is that the person in the ERM is encouraged to think about his choice well in advance and then commit to it. Now, we all know that when making a choice, we LEAN towards an option for a certain time, before COMMITTING to the choice, usually right before we act. The ERM could simply be reading that leaning process.
I therefore submit that it is possible to trick the machine by complexifying the choice process. Ie by thinking / imagining "left left left left" for a few seconds before switching to right (or not, say once you choose what you lean for, the second time you choose the opposite, and etc. alternatively). Doing so would (I guess) lead to a success rate that would be less than 50%.