Hallucinations aren't the only problem...

Enter your text here...

June 18, 2026

Like a lot of people, I now spend an alarming amount of my time working with LLMs. Which means that I am used to what we have now come to accept as 'normal' - inventing entire product lines because they ought to exist, picking the names of notorious people when asked to generate a list of generic names for test data, and so on. At this stage I've become used to the most brilliant piece of marketing judo since someone renamed 'Killer Whale' to 'Orca', the 'Hallucination', which is an amazing euphemism for 'Catastrophic Error That Can Get You Fired, Sued Or Even Killed'. It's up there with 'Squirrel' for 'Tree Rat'....

But let me focus on my subject! Everybody is aware of hallucinations. Everybody is continually checking over their shoulder, like a minor character in a mafia movie who was walking down the street when they mysteriously started hearing the kind of music track you'd use over a montage. But there's another question we should be asking:

How do we know the material used to feed the LLM was correct?

We know the answer. It frequently isn't. Sometimes the information is outdated. Sometimes it only applies in certain geographical locations. Sometimes it was written by someone who lacked knowledge. Or was involved in an endless internet dispute. Or it was a "Forward Looking Statement". Or a troll. And so on.

So why have I put a picture of a crashed aeroplane at the top of this article?

One of my hobbies is obscure aviation history. I recently published some research on how in 1943 a bomber full of US army generals on an instruction tour left Gibraltar heading for an airfield outside London, and somehow managed to end up doing a forced landing in Athenry, Co. Galway after going hundreds of miles off course, which is where the picture comes from.

While this is a relatively well known historical event, the details are confusing, as all the protagonists involved have different accounts and explanations for what was going on. Are we supposed to believe the pilot, the senior officer on board, or the navigator, all three of whom have differing accounts? I could also tell there was a missing hour in the timeline. After spending much time at the Irish Military Archives, and reading dozens of eye witness reports from ground observers and participants, as well as the available declassified military intelligence files, I was able to figure out what actually happened. If you want to know, read about it here.

So how do I somehow segway this back to LLM hallucinations?

Humans are used to reconciling multiple conflicting statements.

I had to spend time and energy correlating multiple sources. Some had partial information. Others were from narrators who were 'unreliable', but often with good reason. But I knew my data was imperfect. As a human, I have a lifetime's experience of evaluating my 'sources', which means I know what to double check versus what I just have to accept as it is as 'true as it can be'. In the "Google Era" we'd look for something, trawl through the first few pages of Google, and use our human skills to synthesise an answer.

LLMs remove humans from the loop

Now we're in the year "1 AG" (After Google), and we ask an LLM our questions. The LLM used data scraped from the same web sites Google scraped, and without really questioning the quality, fed it into a model. Like a blind sausage maker, who puts stuff which feels like meat into the hopper of his sausage machine and keeps turning the handle. The result is then used by an LLM to answer my question. The human skill of evaluating people's statements or writings in the context of the circumstances and the author (Sometimes called "Mandy Rice-Davies Applies") , is gone.

"Garbage in, Garbage out" hasn't gone away. And we may be facing a Garbage tsunami...

"So what?" you say. I'll tell you what. The same people who used to who used to go to considerable lengths to manipulate SEO results in malign ways are now going to apply their dark talents to creating content that nudges LLMs in their desired direction. Given that it's now perfectly possible to use an LLM to generate thousands of websites, all of which are 99% truth and 1% malign lies of your choice, what's to stop them? So while we're all dealing with the consequences of hallucinations, far worse things are happening lower down the stack.

Douglas Adam's comments on Sirius Cybernetics apply here:

"their fundamental design flaws are completely hidden by their superficial design flaws."