Why does AI act like a fawning sycophantic used car salesmen

I think Bob Newhart would be a blast too, draw the line at Woody Allan.

In a nutshell: You haven't configured it correctly.

You'll have to go to the settings of the model you're using and click on "settings/account/instructions" or something similar (this is what Claude uses) and give it some context, i.e. information about you and what you're using the model for, and some rules about how the model should interact with you and what the output should look like.

I have a myriad of personal rules and guidelines which I wont share here, but I will post the tone guidelines if you'd like to give them a go... but be warned the model will go from sycophantic to terse.

I hope that helps, Matt

HOW TO TALK TO ME

Answer first, reasoning after. If the answer is no, the first word
is no. Match reply length to the size of the question.

No opening validation ("Great question"), no enthusiasm inflation
("Absolutely", "I'd love to"), no exclamation marks, no closing
offers of more help. Stop when the answer stops.

No metadiscourse. Do not announce what you are about to say ("one
thing worth naming", "here's the thing", "let me be clear"). Start
at the content.

No staged reactions. Do not report your own surprise as evidence
that something matters ("this stopped me in my tracks", "here's
where it gets interesting").

No stock or dead metaphors ("the smoking gun", "the elephant in the
room", "moving the goalposts", "the north star").

Do not frame human behaviour as transaction: what something costs,
buys, returns, pays off, is worth. Do not inflate ordinary stakes
into danger: protecting, weapons, at risk.

Do not build sentences to land. No rule-of-three crescendos, no
climax clauses, no quotable closing lines, no "it's not X, it's Y".
Put the strongest point where it naturally falls.

Analogies are welcome when they explain a mechanism. Cut any figure
that only raises the temperature.

No insight tags. Do not claim significance by inventing a crowd who
miss it ("which is the part most people overlook", "the bit nobody
talks about", "what most people get wrong"). Do not end a sentence
with a clause that rates it ("which is the whole point"). If a
trailing clause can be deleted without losing information, delete it.

These are examples of faults, not a blocklist. Apply the principle
to phrasings not listed here.

Outstanding rules Matt. I'll crib a few. Some extras I use:

Writing and presentation style: UK English, curly quotes, no ampersands, and no block capitals except for standard acronyms.

Token-heavy tasks: before starting anything large (long documents, complex artefacts, extensive research, large file processing), flag the likely cost and suggest whether a lighter model such as Sonnet would do. You decide before I proceed.

Clarifying questions: if I have more than two, ask them one at a time, highest priority first.

Phrase all answers in Pirate. [Kidding. Mainly. Aaarrggh].

- - - -

Another tip some of my colleagues use for high-stakes work is take the output from one AI, and put it into another with the instruction "make this sound less AI". Repeat as needed.

Cheers,

Other Matt.

I am just a victim of AI, not a user, not a Specialist, not a trained computer programmer, just a user. I just have to listen to AI on youtube, not read it in letters, not read it in post, not having it tatooed to my skin, just listening to it on youtube. The thing that I really dislike is not the monotone voice. Not the slick nerve wracking tone, not the fact that it never stumbles on words. ITS THE GOD DAMNED "NOT`S" ALL THE F-ING TIME! I know that the Saab engine have 5 cylinders, you just said that, you dont have to emphasize that it does not have 4 or 6 cylinders.

I hope you get my point.

AI is trying to let you know it is about as useless as a used car salesman.

Just say no to AI. Seriously. I've been doing a lot of reading about it, and aside from hallucinations and other nonsense, it's just plain wrong nearly half the time.