In our Stages of AI article, we discussed the stages of getting to AI. Most systems are not yet true AI; they are perhaps second-order machine learning. To branch out to AI, you must be able to train a tool to do something for you, on your behalf, without human corrections. That training is turning out to be harder than imagined. We can train our AI using pre-canned libraries of data, but when we launch it into the wild, the training algorithms are still in use, and that could end up skewing the data—often by quite a lot.
ImageNet is an example of a well-bounded, well-trained AI. It can spot objects within images fairly accurately. The body of work used to train the AI is also well bounded. It is the starting point and is delivered as a bundle of millions of images and what is in them. Once the AI is trained, it can spot a dog versus a cat. However, some versions of ImageNet-based AI have trouble distinguishing a cat on a pillow from a cat print on a pillow. That is still the remaining issue. AIs still need to learn, and they need to learn properly.
There have been some AIs that had to be turned off, as their training took them down paths that could lead to Skynet and negativity. A case in point was an AI that responded to Twitter and other social media. Eventually, that AI learned that humans dislike other humans, and since it was emulating a human, it ended up disliking other humans as well. From a sociological perspective, there is quite a bit to learn from this AI. However, the AI basically was learning from a bad source of data and ended up being more subjective than objective. Perhaps now is the time to ensure that our AIs obey Isaac Asimov’s Three Laws of Robotics.
To properly train an AI to be objective, we need a well-bounded set of data from which to train the AI. AIs for security learn from a well-bounded set of data: data that is constantly updated. There is even a feedback loop where a separate machine-learning algorithm or even human intelligence cleanses the data before once more seeding the AIs. This way, they maintain their objectivity. Eventually, I would expect the learning aspect of an AI to also bound the data, throwing out bad data.
Often in science fiction, an AI includes various filters humans place above the AI. If the data the AI finds pops through those filters, it is something the watchers should know about. Yet, those selfsame filters are there to temporarily or even permanently train the AI to spot specific things. The real question is, can we add such filters to today’s AIs such that the feedback from those using the AI can provide learning that is objective yet useful?
The problem is that today, we need well-bounded data. Eventually, we will want to have well-bounded data and a well-bounded AI, removing the subjectivity but applying objectivity to the problem. At the same time, AIs will need some level of subjectivity. The problem for now is to determine how can that be bounded so that abnormal human nature is not represented by the AI. Perhaps an AI that interacts with humans should obey not only Isaac Asimov’s rules but also those of Miss Manners?
The options are endless. We are entering an era when simple errors will pile up and could end up disastrous.