The Naughty AIs That Gamed The System

Artificial intelligence (AI) is undergoing somewhat of a renaissance in the last few years. There’s been plenty of research into neural networks and other technologies, often based around teaching an AI system to achieve certain goals or targets. However, this method of training is fraught with danger, because just like in the movies – the computer doesn’t always play fair.

It’s often very much a case of the AI doing exactly what it’s told, rather than exactly what you intended. Like a devious child who will gladly go to bed in the literal sense, but will not actually sleep, this can cause unexpected, and often quite hilarious results. [Victoria] has created a master list of scholarly references regarding exactly this.

The list spans a wide range of cases. There’s the amusing evolutionary algorithm designed to create creatures capable of high-speed movement, which merely spawned very tall creatures that generated these speeds by falling over. More worryingly, there’s the AI trained to identify toxic and edible mushrooms, which simply picked up on the fact that it was presented with the two types in alternating order. This ended up being an unreliable model in the real world. Similarly, the model designed to assess malignancy of skin cancers determined that lesions photographed with rulers for scale were more likely to be cancerous.

[Victoria] refers to this as “specification gaming”. One can draw parallels to classic sci-fi stories around the “Laws of Robotics”, where robots take such laws to their literal extremes, often causing great harm in the process. It’s an interesting discussion of the difficulty in training artificially intelligent systems to achieve their set goals without undesirable side effects.

We’ve seen plenty of work in this area before – like this use of evolutionary algorithms in circuit design.

22 thoughts on “The Naughty AIs That Gamed The System

  1. I’m applying for a job where they want to apply machine learning in medical image analysis. I keep thinking of these cases and how there’s going to be a very interesting conversation if they choose to interview me.

        1. I’m not in the US. However no matter where one is, this data is exceptionally valuable not just for progress in diagnostic tools but also for how that can be used for profit by corporations so it is tightly held onto by those who seek to profit from it.

          We need an “open source” like sharing system in medicine.

          Perhaps we will have websites for MRI one day like we already have for DNA and genealogy.

          But then there are the same hurdles with privacy (perhaps less with MRI than DNA).

  2. My neural network for identifying basketballs in images learned that if you just guess the top middle of the image you’re usually right, because a significant number of photos were of people dunking.

    1. In school I aced a multiple choice iq test that seemed to have bias in the questions, despite not knowing many of the answers. The same phrases were used in positive and negative questions such that it was possible to infer the answers in a “what colour is napoleons white horse” fashion. Better tests tend to randomize questions from many authors.

  3. Good luck specifying the reason for playing a game. If I were forced to play the game with the boat and was told to score as much as I could then I would do exactly what the AI did. as I have absolutely zero interest in playing computer games much like the AI probably.

  4. I made genetic algorithm to design simple logic circuits. It worked pretty well, fitness reaching 1 in a few generations and gates left to spare. It even figured out to isolate superfluous gates by disconnecting them from circuit and “wrapping” them into each other.

    Then I decided to hand check the results in spice… I never cussed more in my life, probably. Turns out, that GA was generating highly specialized sequential circuits. Since it was tested with sequence of inputs, and it’s state was preserved during the testing, and the sequence **was always same**, it just produced the same sequence of outputs every time.

    Interesting thing was, I never wanted to simulate sequential circuits. I didn’t even design the simulator to accommodate them.

    I had to introduce randomness obviously, and it worked out in the end. But it was clear that I can’t rely on my own simulation.
    If I ever start work on that again, I’ll have to rewrite the whole thing to use Spice of some

  5. This reminded me of my daughter many years ago. My wife, child and I were walking to a nearby store after a rain shower. As any child likes to do she was jumping from puddle to puddle when mom got annoyed and pointed at a puddle and said “Don’t jump in that puddle!”. Without skipping a beat my kid walks around the puddle and jumps in the next one. Mom was about to lose her stuff when I pointed out the child simply followed her directions. We got to the store and about an hour later headed back to our place. Once again kiddo was jumping in the puddles. When she got to the puddle mom had said don’t jump in, she walked around it and jumped in the puddle on the other side. I pointed this out ot my wife. Where she realized that kids take what is said literally. A few years later that lead to me having a teaching moment when said child realized and said “YOU CAN’T MAKE ME!@” – @ 6 years old. It was fun to point out that she was absolutely correct but that there were means at a parents disposal to compel compliance without any force.

    1. lol,

      Reminds me of when I gave my daughter a data stick with a .iso file on it and asked her to burn it to CD because my CD burner wasn’t working.

      She returned a CD with a .iso “file” on it.

  6. I think these could be great solutions for the right problem. We had a middle school “egg drop” type competition, but the setup was to accelerate a rubber-band car into a wall and get the fastest car without breaking the egg. Nothing could come within 2 inches of the egg, so we were led in the direction of adding a crumple zone to the car. One team brilliantly had the egg holder mounted such that on acceleration the egg would fall out… into a nest of thick fiberglass-like insulation. I was amazed.

  7. It seems like most of these “illegal solutions” were caused by faulty frameworks, missing environmental constraints, or poorly randomized inputs. I find that the solutions generally seem to be defensible in light of the flawed environment that gave rise to them.

    If researchers want AI solutions that are relevant to the real world, they need to properly model the constraints and consequences of the real world. :)

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.