This is SERIOUS...
They didn't say how they actually trained the robots - neural nets, genetic algorithms, or whatever. Assuming they used a genetic algorithm (likely), what this shows is that such a maximizing algorithm will train itself to incorporate deception, as long as the training scenarios are not constrained. Because, simply, it WORKS! It begins to find global maxima of it's fitness functions using deceptive routes...
Now, that really DOES have implications. It is very, very difficult for a human to look at a neural network or a genetic algorithm function and understand what it actually DOES, and under what conditions. All we know is that it maximizes the output fit for a given set of inputs in the training data or experience base. We actually have to observe it in operation to have some idea how it works (for any sufficiently complex matrix or functions).
Case in point - GAs were used to design the compressor turbines for the jet engines of the Boeing 777 - and the GA engineered a design which eliminated an entire set of compressor blades, and was the most efficient. Something that the human engineers had never been able to do, and had significant difficulty in understanding how it had done so, even when they looked at the design. But it worked, and those 777 engines are all the better for it.
But this could be the opposite - we could be training robots that reach globally maximum functions that, frankly, do so with no "morals". If those robots can lie, cheat, steal, even kill...well, unless there is a penalty for that in their training function, they WILL - because it is the most efficient manner of operating.
So, what these esteemed professors have shown is that unless we develop training functions with HUGE negative impacts for immoral behavior, our robots will train themselves to emulate your basic Colombian drug lords in behavior. Interestingly, there are a fair number of people who turn to crime even WITH society showing large penalties for it - and I fear that if the robots assess the probabilities they might come to the same conclusions.
Asimov was right...