About Confabulation

December 16, 2023 Cognitive Insights No Comments

Depending on how exact you want inferences to be, humans are born confabulators. Not surprisingly, we also see much confabulation in present-day generative pre-trained transformers (GPT).

Confabulation is a look-alike but inexact information. Note that in the GPT field, this is nowadays called hallucination.

Confabulation is inherent to mental ‘parallel distributed processing’ (PDP).

Humans and artificial transformers have quite distinct processing substrates, yet they share the principle of PDP. In this, ‘distributed‘ means that many processing units continually work together to realize their simulation of a conceptual system. Yet, it remains a simulation. There is no sharp distinction between genuine memories and plausible (re)constructions in our usual way of thinking.

In the extreme, this means that we are always confabulating — mainly pretending to be largely what we cannot be: a Platonic conceptual system. Ironically, much of our mental power rests on the same basis. The little bit of bad comes with immense processing benefits. We need our confabulation in our art and science, as well as daily life. On the other hand, trying to be too conceptually smart may come with some form of autism.

We are born confabulators.

Graceful degradation

In a distributed system, units can fall out without a breakdown of the system as a whole. The human brain has many non-optimally reliable components, so there must be a lot of error-correction. This is a good thing, of course, but it can also lead to confabulation.

Nevertheless, to nature, this is a best-of-kind solution. It makes the system robust to minor errors and enables working with partially incorrect input information. In other words, there is an automatic accommodation to small errors from inside and outside. The increasingly imperfect result is, in many cases, good enough for nature’s ways.

The same underlying mechanism is also at play when learning a language, then ‘forgetting’ and re-learning it. The second time goes quicker (positive transfer effect even with random relations) although with need of error-correction. You can almost feel the faded patterns become crisper again with reuse. One could call it ‘graceful regeneration.’

Confabulation also aids us to see the world more as we like it to be ― engendering hope and – to many – keeping depression at bay. This comes on top of the just-described processing benefits.

Dementia

This is marked by a higher degree of confabulation, mostly due to the massive fallout of processing units (brainy neurons) and resulting endeavors for error-correction. For instance, in Alzheimer’s disease, plaque formation gradually degrades neuronal integrity.

Thus, together with memory flaws, confabulations also gradually replace genuine memories. The dementing/confabulating person doesn’t directly notice that more and more, he produces content that only looks like real memories. To him, these are just the memories his brain produces. Not surprisingly, long-engrained (old) memories suffer less from confabulation since they are increasingly embedded. The demented person may even regress to childhood patterns of conducting himself. That too is a kind of confabulation.

GPT confabulations

For its being a massively distributed processor (like us, but differently), it’s expected that we encounter many confabulations in GPT output ― more so since the source of its knowledge is still mainly us. It’s also expected that the system tries to error-correct implicitly, without knowing by itself to be doing so.

Developers are busy trying to diminish the level of confabulation. They will almost certainly succeed in reducing this to below human level. From then on, we will see a massive increase in business use cases being rolled out (now still preliminary).

And more

This process of de-confabulation will get another boost when artificially intelligent systems start getting their information directly from the world instead of second-hand through us.

That’s a good thing, of course, but the vicious circle from this means that the pace of increase in the usefulness of A.I. will get another boost. That is, we will get to super-A.I. even sooner.

Are we ready?

Or will our proneness to confabulation close our eyes until it’s too late?