2026-02-13
Another article on AI, another "feature" identified.
Now it is true that not all AI engines (or LLMs: Large Language Models) are necessarily the same in all regards, so one must be careful not to generalise from the particular, but this is a rather concerning particular which rather undermines one's trust in such models.
They do share the need to be trained on as wide a breadth of input as may be obtained, but that very feature mitigates against the possibility of verifying that all the training material used is "true" and "correct" - it's falrly obvious that a lot of invalid material is likely to slip through any "truth filter", if only because there is probably no reliable way to construct such a "truth filter", so commercial ambition will likely result in a "bung it all in" mentality.
John Droz joins the fray to explore AI's features and limitations, and why we should be wary of its output.
