Artificial Intelligence (AI) has revolutionized many areas over the last years, from autonomous driving, cancer detection, to recommending video series to binchwatchers. However, advantages of AI also come with tradeoffs that are often not talked about.
Speed VS Correctness
One of the big benefits of AI techniques is the speed. However, this comes with a tradeoff of taking shortcuts. For Machine Learning techniques, this e.g. means relying on probabilities instead of facts to achieve calculations that cannot be done otherwise. For meta-heuristic search algorithms, this means looking at a specific part of the overall search space that most probably contains the optimum solution. In the end, all of these shortcuts lead to the “most probable” correct or best result. However, this might not be enough if you want a result that is 100 % correct, as it is often the case with algorithms used in traditional software engineering.
Data-Driven Knowledge VS Human Intelligence
Humans decide what to do in a specific situation based on their knowledge and experience. Whereas experience is built by past outcomes of similar decisions (this has worked in the past, so it will also work now), knowledge is gained from experiences of trusted other parties. Deciding which sources to trust is a crucial activity every individual has to perform. Because every decision is based on own experiences or knowledge gained from trusted sources, it is usually easy for humans to reflect on these decisions to give explanations on why some decision was wrong, and how to make it better in the future.
On the other side, AI techniques (machine learning in particular) usually rely on historical data about some event that is adapted by humans to opt for certain decisions. The result thus depends on the data that an algorithm has available. If something is wrong in the data, there is no way to automatically build a “bullshit filter” that detects inplausible connections. So if a text generator is trained on fake news, it will most likely produce fake news, again. However, as the way that an AI algorithm retrieves a certain result is usually a blackbox (there’s a dedicated research area concerned with this, called “explainable AI”), it is almost impossible to detect such errors once they are introduced in the algorithm, or track the source of a wrong decision.
How to deal with AI
If we want to apply some Artificial Intelligent algorithm to a specific problem, we should never
solely trust that “AI solves everything”. Certain benefits always come with tradeoffs (there’s even a name for this: scientists refer to it as the “No Free Lunch Theorem”). If we are aware of these tradeoffs, we can also weigh them against the potential benefits to decide whether we should actually opt for a certain decision. This level of human intelligence is still required. And it usually requires us to have at least a basic understanding of the algorithms that we make use of. But if with this knowledge, we can put AI into the right place, and live with it without hesitation!