Studies keep reporting that around 95% of AI deployments fail to benefit the bottom line. If you are considering one, it is worth understanding why before you become part of the statistic.
1. A language model has no knowledge
These are tools for building powerful expert systems. But the model itself has no knowledge or understanding, any more than a book understands the words on its pages.
2. We anthropomorphise anything that talks back
Human beings are social animals, and we ascribe human attributes to anything that looks or sounds even vaguely human. The easiest on-ramp to AI is an interface that acts human, so most people do this unconsciously.
It is like a photo-realistic mannequin in a shop window. Even with the best animatronics inside, you are not going to hire one for a real job.
3. A model cannot unlearn
It is not currently possible for a language model to unlearn something. Feed it large amounts of countervailing data to offset an error you have found in a dataset and you can do as much damage as good, because you are skewing every other data point at the same time.
This is why “general AI”, as currently approached, is not achievable by simply adding more. Dumping the entire contents of a large forum into a training set gives you a pile of garbage that smells like that forum. Adding an encyclopaedia on top improves your odds of retrieving something fresh rather than something mouldy – but the garbage is still in there.
None of which makes it a weak tool
The failure rate is not a statement about the technology. It is a statement about how it is used and what is expected of it – and those expectations are badly misaligned most of the time, including on the part of people selling themselves as experts.
Itwerx Corp is a service-disabled veteran-owned small business providing IT services across Seattle, Bellevue, Everett and Snohomish County. This is the kind of thing our AI integration work deals with – talk to us about yours.
