AI has reached human levels (annoying human levels, but whatever).

AZEL

AZEL

Pre-takeoff checklist
Joined
Dec 6, 2025
Messages
261
Display Name
AcrftZeroEngineLand
I was discussing a show with fellow students and one made a point I did not think was correct. So, I used Gemini to fact check. It confirmed my recollection, but in its confirmation made two erroneous statements of fact. I challenged the most glaring one by restating its definitive as a question. It "confirmed" its erroneous response. Since this was a question that could easily be verified, I opened a new window and described the exact scene where the "fact" Gemini denied happening occurred. Then I went back to the first window and gave the episode number, approximate time, and described the scene. Instead of double checking, Gemini stated conclusively that I was wrong and reasserted (in a stronger and borderline more aggressive way) its previous wrong answers. Then, when I went through the scene step by step it fabricated arguments to support its erroneous position.

Human level intelligence, we has arrived.
 
AZEL said:
I was discussing a show with fellow students and one made a point I did not think was correct. So, I used Gemini to fact check. It confirmed my recollection, but in its confirmation made two erroneous statements of fact. I challenged the most glaring one by restating its definitive as a question. It "confirmed" its erroneous response. Since this was a question that could easily be verified, I opened a new window and described the exact scene where the "fact" Gemini denied happening occurred. Then I went back to the first window and gave the episode number, approximate time, and described the scene. Instead of double checking, Gemini stated conclusively that I was wrong and reasserted (in a stronger and borderline more aggressive way) its previous wrong answers. Then, when I went through the scene step by step it fabricated arguments to support its erroneous position.

Human level intelligence, we has arrived.
Click to expand...
This is where these LLMs have been from the beginning.
 
ChatGPT wasn’t that difficult.

I asked it to list all automobiles sold in the USA that come with manual transmission (standard or optional).
It missed the Honda Civic Type R, when I pointed out the error, it responded with “good catch”.

But its understanding of Einsteins Special Relativity is beyond 90% of pseudo experts on YouTube or Reddit.
 
TCABM said:
The problem all AI tools face is they cannot let a question or statement go unanswered or answered with I don’t know.
Click to expand...
The widely available public-facing tools, that is. There are ways to get a 'level of confidence' as part of the output.

One way is the adversarial model, which has another instance of the tool try to poke holes in the response of the first and tell you why it may be incorrect. One system I use at my job uses that model, but it also seems like @AZEL did something similar on his own with Gemini in the OP.

If you create an agent yourself you can also implement RAG and deactivate use of public knowledge sources and you can very well get it to tell you if the knowledge source to which it's pointed lacks the info it needs for a complete answer.

I've also heard of another way, which is testing of entropy across multiple instances. Basically it's having your question answered many, many times, and if there's a high amount of semantic variance in the collection of answers, it's more likely to be incorrect than answers that are similar across the board. But this method is computationally and energy expensive for obvious reasons and isn't likely to be used in free chat bots for a long time.
 
Last edited:
strikes2k said:
It's called a hallucination and is one of the biggest problems with AI. But AI is a misnomer for what is being put out there right now. It's not intelligent at all. It's just high level pattern matching at it's core.
Click to expand...
It's known as "bull$h1tting", which is distinct from "hallucinating". There was an IEEE paper presented on this a couple of months ago at a conference I attended - very interesting stuff. [The conference was a good excuse to take a XC flight in the Bonanza, too.]
 
I tried to “educate” Chat GPT on a specific subject matter when it was in error. I confirmed with “it” that it would from now on understand and give the correct answer. I went back on a separate device, did not log in, and it was again incorrect. It’s trying to suck up to you, beware.
 
455 Bravo Uniform said:
I tried to “educate” Chat GPT on a specific subject matter when it was in error. I confirmed with “it” that it would from now on understand and give the correct answer. I went back on a separate device, did not log in, and it was again incorrect. It’s trying to suck up to you, beware.
Click to expand...

It is, it always gives me compliments on my questions or comments. I just thought it like me .
 
455 Bravo Uniform said:
I tried to “educate” Chat GPT on a specific subject matter when it was in error. I confirmed with “it” that it would from now on understand and give the correct answer. I went back on a separate device, did not log in, and it was again incorrect. It’s trying to suck up to you, beware.
Click to expand...
This is a feature, not a bug. You don't want random people "educating" AI.

That it got a fact wrong to begin with is ALSO a bug. But the fix isn't randos.
 
Half Fast said:
I am sure people are already working to feed BS to language models deliberately, either for fun or with serious malicious intentions.
Click to expand...
There are government agencies devoted to infiltrating training data in order to insert triggers.
 
I made the mistake of asking GPT 5.2 about some turbine engines (TPE331 vs PT6 variant). The colossal amount of ******** it cooked up was almost mind-blowing.
Each iteration of questioning and it invited further, deeper questions about the mechanics of how they work and the design considerations -- as if it was in a position to opine on such matters. It was entirely false. It relied on erroneous citations that concluded none of what it was suggesting. It referenced generic drawings from google image search that had absolutely nothing to do with anything.

Sure the answer was well formatted, had good grammar, produced a somewhat believable train of thought, etc. If it was a middle school science project answer you'd probably applaud the detailed thinking. But it was total, 100% crap and you'd be dumber and worse off for having read it (I'm still trying to claw back the lost IQ points).

So my only advice is that you'd be smart to verify every answer you get outta AI, cuz...
1768259099150.png
 
I asked AI what the bolt torque for my motorcycle’s handlbar clamp was. The answer AI gave me was twice the torque that the factory manual listed. I would not recommend consulting AI for anything critical. On the other hand, it’s quite amusing to see what AI says on a subject that you are technically familiar with. I just hope the younger generation doesn’t rely on it too much.
 
The other day I posted an eccentric question with my gut shot hypothetically reason why it may work on a different aviation forum. My thought was there are many experienced and expert level members on it from a very diverse background in a long shot maybe one of them was a chemical engineer and could give a qualified yes/no… or completely blow it out of the water with “it’s a bad idea, here is why”. Another member posted the same question to an AI program and posted the results… in the results it listed that the idea has been discussed on experimental aircraft forums. I doubt it had any source for that other than my initial post. (Accelerating cross link polymerization and associated cure time of boiled linseed oil finishes with ozone enhanced ambient atmospheres is not probably something that is often talked about on experimental aviation forums. I could be wrong, but doubt it :). ).
 
Last edited:
Google’s AI that they now insist on showing with every search is ridiculously stupid and wrong most of the time. ChapGPT has turned out to be useful and mostly correct, but needs fact checking occasionally. It’s handy for doing research legwork for me so I don’t have to spend hours digging through web sites.
 
DaleB said:
Google’s AI that they now insist on showing with every search is ridiculously stupid and wrong most of the time. ChapGPT has turned out to be useful and mostly correct, but needs fact checking occasionally. It’s handy for doing research legwork for me so I don’t have to spend hours digging through web sites.
Click to expand...

For general purpose search, I reach for GPT first now. I haven't "googled" anything that wasn't maps in awhile.

I don't trust sam altman not to enshizzlefy chatgpt, so I think we'll start to see sponsored results soon there, and that will be heartbreaking. when the ANSWERS become sponsored, welp... that will be my clue to go buy a missle silo somewhere and convert it into a prepper shack. I fully expect to get there this year.
 
You must log in or register to reply here.
Back
Top