Building AI into the product - a good idea?
There is a lot of talk about AI at the moment, but many products are still "dumb". As a result, many development teams will be asking themselves whether AI should be built into the next generation of products. In the following, I explain why we should think twice about it.
Real or fake?
AI deals with the automation of intelligent behavior and machine learning. The definition is relatively broad. It is therefore often difficult (or even impossible) for the user to recognize whether a system is actually using AI. To name just two examples:
- Clippythe infamous Microsoft Office assistant, was supposed to appear intelligent but did not use any AI.
- The program written in 1966 ELIZA also contains no intelligence whatsoever, but has allegedly driven people up the wall in online forums who thought it was a real conversation partner. Have a conversation with ELIZA now >>
When analyzing a problem, we should therefore carefully examine whether the problem requires a "real" AI. This is certainly the case with topics such as autonomous driving or the analysis of medical images. But even with a chatbot, it doesn't necessarily have to be "real" AI, as ELIZA impressively demonstrates.
We might now ask ourselves: Why shouldn't we install AI? There are a whole host of reasons for this:
What it means to build AI into a product
First of all, the development of AI-based systems is a good deal more complex. This is because the training data for the AI is an integral part of the development. The training data must be versioned, compared, checked for quality, etc.
I wrote about all the things that can go wrong in a guest article in the HOOD block: The 11 traps in machine learning
HOOD Blog
The second problem is the testability or predictability of results. Just because AI delivers good results under laboratory conditions, we cannot be sure that this will also be the case in a productive environment. In particular, results in AI are generally probabilities. This makes it difficult to test for "right" and "wrong".
On the one hand, this means that AI-based systems should have an update mechanism. However, this again increases complexity. Furthermore, we have to test updates very carefully so as not to offend users. A doctor would be irritated for good reason if a system diagnosed a tumor on an X-ray image and a day later no longer did so on the same image.
The third problem is the need for resources, both in the product and in development. Tesla is another good example here. Tesla installs a powerful control unitto meet the requirements of autonomous driving. At the same time, Tesla is building Dojo one of the most powerful computers in the world. Tesla needs Dojo to process the training data for autonomous driving.
Alternatives to AI in the product
I have already mentioned a few alternatives to real AI, Clippy and Eliza. Ultimately, the focus should always be on customer benefit. Then we can check whether AI is the best solution or not.
By far the most common alternative to AI is implementing decision logic or looking things up in knowledge databases. This was known as Expert system a hype in the 1980s. Even today, a lot can be achieved with clever algorithms and well-maintained knowledge databases. Even Fuzzy logica hype from the 1990s, can look like AI.
Conclusion
There are major problems that really cannot be solved without AI in the product. But for many companies, the cheaper and often better way is to stick with classic algorithms.
The only question is what we should do if marketing insists on having to use the "AI" label. Then we should look for a small, unimportant feature where we can install a ready-made, low-resource library. But in this case, it would be better to convince marketing of the nonsensical nature of the action.
