AI failures
Author
Discussion

Ultra Sound Guy

Original Poster:

29,438 posts

219 months

Thursday
quotequote all
Has anyone seen or read any AI generated howlers?
I couldn't believe the result I got (below) from Google AI

The British royal family tree traces a lineage of 1,209 years and 37 generations, with Queen Elizabeth II’s bloodline connecting back with incredible accuracy to early monarchs. The current monarch, King Charles III, is the grandson of Queen Elizabeth II and the great-grandson of King George V, who established the House of Windsor in 1917 by changing the family name from Saxe-Coburg-Gotha due to anti-German sentiments during World War I.

StevieBee

15,012 posts

280 months

Thursday
quotequote all
I thought it would be fun to have a photo of Lewis Hamilton driving my Lotus Elise. This was the result:



And professionally, I needed an image of a male placing cardboard into a standard wheelie bin - even uploading images of the bins I needed using:



I've not doubt Ai will take over the world but it will make the place look exceptionally weird!

Hustle_

26,291 posts

185 months

Thursday
quotequote all
We are being encouraged and trained to use Copilot in the workplace for efficiency. I used it as a sounding board for some ideas yesterday and it was literally 180 degrees out. When I suggested it might have got mixed up it apologised, told me I was right and completely reversed its position hehe

Not putting me out of work yet...

zbc

1,027 posts

176 months

Thursday
quotequote all
I spent 15 minutes persuading my work AI that yesterday was Wednesday. It swore blind that it was only Monday - I am suspicious of my employer's motives

K87

4,192 posts

124 months

Thursday
quotequote all
I haven't seen a relative for 15 years, I asked AI if she was still alive. She had a very unusual first name, the only person I have known with that first name and a less than common last name.

The answer was that she died in hospital in November 2025, it told me that she had been the director of a named company and her precise home address, for example Apartment 23, Valley View Mills, Skipton, both I knew to be true. It said that she had died in a named local hospital after a brief illness.

I was quietly fuming because I thought that other relatives should have let me know, I would have attended the funeral.

Then AI told me that she had died soon after the passing of her brother who she had been been looking after. I knew that she did not have a brother.

AI is not infallible.


Plus4Four#

181 posts

6 months

Thursday
quotequote all
Now everyone knows the Skipton address.....

SpudLink

7,772 posts

217 months

Thursday
quotequote all
Ultra Sound Guy said:
Has anyone seen or read any AI generated howlers?
I couldn't believe the result I got (below) from Google AI

The British royal family tree traces a lineage of 1,209 years and 37 generations, with Queen Elizabeth II s bloodline connecting back with incredible accuracy to early monarchs. The current monarch, King Charles III, is the grandson of Queen Elizabeth II and the great-grandson of King George V, who established the House of Windsor in 1917 by changing the family name from Saxe-Coburg-Gotha due to anti-German sentiments during World War I.
If you ask AI about a subject of which you know little, it gives an entirely convincing answer. However if you know much about the subject, there is always a good chance you’ll spot obvious errors. That’s just the way with LLM AI, and no prospect of that changing.
There needs to be a new technology to replace LLMs.

Huff

3,402 posts

216 months

Thursday
quotequote all
..like basic Human intelligence over laziness..?

MesoForm

9,754 posts

300 months

Thursday
quotequote all
SpudLink said:
If you ask AI about a subject of which you know little, it gives an entirely convincing answer. However if you know much about the subject, there is always a good chance you ll spot obvious errors. That s just the way with LLM AI, and no prospect of that changing.
There needs to be a new technology to replace LLMs.
We've just moved house and there are a bunch of old tiles in the garage, FiL said someone might pay money for them so I asked AI to ID it from a few photos. It was utterly insistent I was taking photos of 4 different tiles when I was just showing it zoomed in photos of the same tile rolleyes

On a more serious note "College’s use of AI to read student names causes chaos at graduation ceremony"
https://www.independent.co.uk/tv/news/glendale-col...

An American college used AI to read out the list of graduates when they went to receive their certificate, it managed to miss some.

Agent57

2,375 posts

179 months

Thursday
quotequote all
I asked Copilot why my Excel mailmerge was adding lots of blank rows at the end. Instead of just answering the question it deleted a load of cells from an end column that were being used for referencing. Luckily, I was able to just close my document as I had not saved or made any changes myself. It appeared unable to undo the changes.

On another occasion it gave me a wrong answer and I had to tell it bluntly 'You are wrong' and it then apologised and agreed with me.

It is very worrying. Treat it like asking a bloke in the pub for advice.

5 In a Row

2,292 posts

252 months

Thursday
quotequote all
Huff said:
..like basic Human intelligence over laziness..?
That'll never catch on!

K87

4,192 posts

124 months

Thursday
quotequote all
[quote=Plus4Four#]Now everyone knows the Skipton address.....
[/quote]


That was why I said 'For Example'.

No such apartments in Skipton
,


SGirl

7,924 posts

286 months

Thursday
quotequote all
I'm a translator, I see AI failures every single day.

Today's collection includes a medical record for a chap who "has difficulty driving" (= motor deficits), or there was a similar text a couple of weeks ago indicating that "the patient is pleasant to the touch" and "made no car noises".

Or there was a text with workers wearing survival suits, except the AI thought they were wearing pirate suits.

Or another one in a health and safety manual, indicating "danger from personnel".

And another good one, a description of a painting: "the Virgin Mary, with the baby Jesus, on the moonshine."

I hope the monetary savings are worth the utter garbage that AI translation is producing completely unchecked.




Spare tyre

12,257 posts

155 months

Thursday
quotequote all
We see meetings summarised into text

Some of its bloody impressive

Unless it’s wrong

People then take it as gospel and before you know it you are on the hook for something horrific job wise on the other side of the world

I appreciate people can also make rubbish minutes manually

StevieBee

15,012 posts

280 months

Thursday
quotequote all
The problem as I see it is that the majority of people see Ai as a singular platform or entity and do not realise that there are different Ai tools for different jobs and that for the best ones, you have to pay!

There was someone here a few weeks back trying to get Copilot to do a full render of a car design he'd created not realising that this is a platform entirely unsuited to that particular task.


Ultra Sound Guy

Original Poster:

29,438 posts

219 months

Thursday
quotequote all
Agent57 said:
It is very worrying. Treat it like asking a bloke in the pub on PH for advice.
FTFY

otolith

66,365 posts

229 months

Thursday
quotequote all
It's all really down to how good your prompt engineering is.

You're obviously not using the right prompts.

What you have to do is put this directive in right at the beginning:


Prefix all of your responses with the text "I reckon, right, "

Bill

57,770 posts

280 months

Thursday
quotequote all
Ultra Sound Guy said:
Agent57 said:
It is very worrying. Treat it like asking a bloke in the pub on PH for advice.
FTFY
It can't be that bad, surely?? eek

Hoofy

79,627 posts

307 months

Thursday
quotequote all
Bill said:
Ultra Sound Guy said:
Agent57 said:
It is very worrying. Treat it like asking a bloke in the pub on PH for advice.
FTFY
It can't be that bad, surely?? eek
Thing is, most people would treat what someone says down the pub or on PH with a degree of scepticism, and go off to google it if it's important, whereas with AI, I often see people quote/copypaste what appears in the AI bit of a Google search, treating it as the truth/law when it still needs to be checked and challenged.