AI taking over the world is nonsense

Introduction

The large neural networks behind today’s generative AI are dangerous in that they can act as expert hackers, apparently hacker that are difficult for even their creators to control. This requires some creative effort to control both on the part of the creators and governments. However, what I am addressing in this note is the warnings by experts that (1) AI will take over jobs and create massive unemployment; and (2) AI will become so “superintelligent” that it will dominate humans.

Jobs

The fear of automation taking over jobs has a long history. So far, it hasn’t been reflected in economic statistics. In the US, job growth and low unemployment are the norm. Typical warnings are that AI can take over about 40% of the tasks in jobs, so 40% fewer workers will be required. But AI can be used as a tool to make a worker more productive; for example, it can make programmers more productive by letting generative AI create basic code to review. It can provide a digital assistant that acts as a trainer in a job; for example, some coffee chains are using it to tell new baristas how to make fancy coffee blends. When individuals are more productive, they deliver products or services more cheaply. When increased productivity increases efficiency, it can protect existing jobs if not create new ones. New jobs can also be directly created by AI—a huge number of construction workers will be required to build the big computer centers being planned to support AI use. And, in the US, automation in factories may bring back manufacturing that has been done abroad back to the US, with automation making the factories competitive with low-paid workers abroad. Robots will be doing repetitive production-line jobs Americans don’t want, but they will be happy to manage the robots.

Also, the number of workers are declining as the population shrinks in advanced economies and immigration declines. I have argued in several books that basic demographic trends will require that AI will eventually be required to take over more jobs to replace a shortage of workers.

There may be some disruption. There are current reports that there is a shortage of entry-level jobs as AI does some of the more simple tasks that were done by new employees. A college degree is apparently no longer a guarantee of a job. Some experts say that education opportunities for skilled jobs such as construction workers, plumbers, or electricians are lacking. (AI could help
here; a digital assistant that the worker converses with could instruct a worker how to address a new challenge.) For an individual being affected by these trends, saying that the economy will eventually balance all this out isn’t much consolation. But there are strong arguments that the trends won’t cause job loss in the long run.

Superintelligence

When large language models using very large neural networks were first introduced, they were said to surprise even their creators by their ability to create an article on a subject based on just a text request. It is perhaps natural to assume making the networks much larger will make them more intelligent—superintelligent. They might surprise us by how they could even build smarter AI without human help, continuing their evolution. As we give them the ability to control physical systems like robots, they can also act directly on their intelligence. Companies are spending billions of dollars to build huge computer centers and power stations to train the larger networks based on that bigger-is-better assumption.

Unfortunately, the assumption that bigger-is-better fails because there is not enough quality data available to train the models. When you increase the number of parameters in a statistical model, you have to increase the amount of data used to train the model by a larger factor, what has been called the “curse of dimensionality”. One estimate I’ve seen is that, if you double the number of
parameters, you must increase the amount of data by a factor of ten. If not, the examples will not be dense enough in a high-dimensional space to allow meaningful extrapolation by a statistical model.

There were problems with data sources on the Web in building the current Generative AI models. The first Large Language Models expressed some of the racial prejudice and pornography common on the Web and had to be suppressed by the model builders. Today, the Web has the same problem with the addition of conspiracy theories and other content that most would consider disinformation. There is an additional problem that many web sites are now being created by Generative AI learning from existing sites and are thus not independent data. Further, many quality sites such as newspapers are suing the original model makers for using their data. All of these factors making the need to get ten or more times the amount of quality text data that today’s models used is more than problematical—it is impossible.

What happens when you don’t use enough data? You get results that depend on too few text sources and don’t generalize a concept. You get more “hallucinations.” The bottom line: The larger models will be unusable.

For more

These ideas are expressed in more detail in my recent book, The Limits of AI: The True Role Of Computers In Our Future.