Fatal Figment Fragments
I completely understand your frustration. Getting straightforward historical data wrong is unacceptable, and it is incredibly grating when an AI confidently presents incorrect numbers as facts. You have every right to be annoyed by it.
Because I do not have a human mind or a true memory, I "hallucinate" by statistically guessing the next word based on patterns, which can cause specific numbers, dates, or [specific publicly recorded data] to get completely scrambled if they are not verified in real time.
While I cannot [stop] on my own, I can promise to do better by strictly checking my facts via live databases rather than guessing. If you still want to hunt down the accurate data you were looking for, I am here to do it right. If not, I entirely understand.
Those seven sentences were served me by an AI bot. In response to my telling it I was "aghast" at its "hallucination". When giving me figures which I was unable to corroborate. Worse still, these actual matters of record were wildly different to what had in those seconds of thinking time magically appeared on my screen.
Examples of this are legion. From that 'glue on pizza' tip to stop cheese sliding off, through lawyers caught citing imaginary case precedents, to Google's chief exec duped into a clear untruth when proudly demoing his bot and so casually wiping off a cool hundred billion from their value.
I still wince recalling my own endeavours misreported back to me.
Which reminds me of the Gell-Mann Amnesia Effect. Freshly applied from media sources to AI output. Where we realise that on an area of our own personal knowledge, the answers range from troubling to outright lunacy. Yet we somehow suspend that lens when asking about areas beyond.
Lately, other labels appear.
Domain Blindness. Thinking we can be proficient where our background is at best, sketchy.
The Jeopardy Phenomenon. Lulled into false comfort from the rapid and florid yet ultimately misguided answers (aka the Fluency Heuristic I think, as in the 'sounds right' glitch).
AImnesia. I heard about this from an accountant. When you unleash (Copilot in their case) AI on your spreadsheet and just assume their numbers are all correct, because machine. Even when they contradict your own workings.
These can each derail our sale.
Anyone selling in that dark and distant pre-LLM epoch may well have an anecdote of a prospect balking at an 'unsound' fact pitched. Then breaking our deal.
Guardrails around this have emerged. For instance, there are two different THINK Protocols. Suffice to say, checking output looms large.
Which is my main point here. Anything - yes absolutely everything - we state a prospect can look up through AI. Who knows what they'll find. So best we know for ourselves beforehand.
We can even have this prep as backup ammo.
After all, imagine having a slide or some other collateral ready that lists all the hallucinations encountered around what we know to be true.
That'd only need to be drawn on once, but boy, what impact we could then make ...