These days, I respond to hearing “generative AI” with a bit of an eye roll. The smart ass in me thinks: “As opposed to ‘degenerative AI’ that is going to make all of us degenerates.” I can be a little (?) quirky sometimes.
There are things we do and do not like about AI, generative, degenerative, or otherwise. We like it when it automates tedious tasks or when it anticipates us for a smoother experience. We appreciate it when it crawls through a ginormous corpus of information to find things we have not thought of. And we are endlessly entertained by the funny images it generates.
(The image included here is one of the responses from OpenAI DALL-E 3, accessed via Microsoft Copilot, to the prompt, “Please generate 2 to 3 images of what degenerate AI robots would look like.” Not only do they look not at all like degenerates, but I have no idea what the significance of the chessboard is; it was in all the images. Ironically, it appeared to be confused by the word “degenerative.”)
However, we do not like that they are not transparent to us about how it works, what data they learn from, etc. We may distrust AI with some things while not with others, and some of us distrust it more than others.
Through all this, we have created lofty expectations for AI and analytics in general. We want perfection all the time. We become upset when they get it wrong.
Worse, sometimes we do not even know when they get it wrong. Ignorance is bliss; we carry on.
Maybe it is more degenerative than we think.
“All models are wrong”
This is broadly attributed to the famous British statistician, George E. P. Box. The point is that there is no such thing as a perfect model in the real world.
All empirically derived models are wrong now and then, pretty much by design. This includes the most sophisticated, cutting-edge AI algorithms. When you think about the word “model,” it makes sense. It is a representation of a reality, not the reality itself.
We may have subconscious or even explicit expectations of perfection. However, models work on chances rather than certainties. We do not know when and by how much they can be “wrong” until they actually go “wrong.” We may or may not ever know why they go wrong. And they can be very wrong.
In essence, all empirically derived models give you tailored advice based on what they “know” and have been “told.” Even the best GenAI algorithm spits out what it “thinks” to be the most plausible outcome or response. While they may be very good, they do not guarantee they have the right answers. The models depend heavily on where their knowledge comes from and who tells things to them.
For the user, there are two key questions. (1) How often is too often to be wrong? (2) How wrong is too wrong?
Sustained accuracy and its (often unplanned) costs
Higher precision or accuracy is not categorically better with algorithms, as it almost always implies greater instability. They tend to go wrong more often in a shorter amount of time, thus requiring more frequent updates and adjustments.
For one thing, this translates to greater maintenance costs and greater resources required just to support them once they are in production. Some models may be designed to update themselves, but this comes with added risk and, therefore, costs, once operational.
The best models strike the right balance between accuracy and stability for specific needs. I have seen very well-thought-out, straightforward predictive models outlast complex ones by over a decade, even surviving multiple economic crises.
In my experience, the total cost of ownership is almost always underestimated in analytics. The post-development costs, specifically operational costs, are particularly underestimated—if they are estimated at all.
The eventuality of “bad advice” given by algorithms
Although algorithms give their best advice, they can, and eventually will, give you bad advice.
It may do so simply by chance. The real world has natural variability built in, simply because it is the real world. There is no way around this unless you live in a different world than the one where we mortals live.
Then there are systematic errors when the generation (not just collection) of the data is not specifically designed and controlled for the sole purpose of developing the algorithm. Given how most of the data we use for analysis is generated today, there is not much getting around this, either.
Finally, but not trivially, there are systematic errors introduced by analysts who develop the algorithms. Things like suboptimal application of techniques and the failure to recognize and adjust for the myriads of biases and suboptimalities are some of the big ones. These are rarely intentional, and a lot can be thoughtfully done, but very often they are not.
The “wrongs” go both ways—missed opportunities on one side and loss/damage on the other. And while some wrongs are relatively inconsequential, other wrongs are much wronger. To quote Box: “Since all models are wrong the scientist must be alert to what is importantly wrong. It is inappropriate to be concerned about mice when there are tigers abroad.”[ 1 ] At the extreme, there is a line to be drawn when wrong results in anything catastrophic, fatal, unethical, immoral, etc. If you cannot afford to get the wrong advice even once, it should give you some serious pause. It is the proverbial “that you can does not mean you should.”
The implication is that anyone using algorithms must be prepared for the eventuality that they will be “wrong.”
The risk of degenerating to mental laziness
It does not mean users have to understand the technical details of the algorithms. It does mean users need to be smart about what it means to leverage an algorithm in each specific use case. The users are ultimately responsible for the consequences of using the algorithms.
Any use of empirically derived algorithms is a risk-reward exercise. The upside of being right must outweigh the combined cost of being wrong plus the development and maintenance.
For both users and developers, the upside is much more attractive and exciting. But we also need to develop an appreciation for the cost of the possible consequences. I find that many are not sufficiently prepared for this, which often has risk and even ethical implications.
It is becoming easier by the day for the average person to develop, access, and use algorithms without much thought and with almost utopian expectations. AI can become a path to mental laziness if we are not careful. That may very well make AI “degenerative.”
[ 1 ] Box, G. E. P. (1976). Science and Statistics. Journal of the American Statistical Association, 71(356), 791–799. https://doi.org/10.1080/01621459.1976.10480949.
