DEV Community

Cover image for Is ChatGPT testing the users?
Alvaro Montoro
Alvaro Montoro

Posted on • Originally published at alvaromontoro.com

Is ChatGPT testing the users?

Why does ChatGPT sometimes give clearly wrong answers? I get it can provide incorrect code or get confused with the natural language in a problem, but today it told me that the square root of 8 is 2 (just 2, no decimals). It's impossible that an AI (or any machine) thinks that's the correct answer.

It was a simple question, too: "what is the square root of the cube of 2?" And it explained the logic correctly. It just gave a wrong result: 23=8, √(8)=2.

Question to ChatGPT: 'What is the square root of the cube of 2'. And an elaborate answer by the AI explaining that it is 2 (8 is the cube of 2, and 2 is the square root of 8)

That is definitely wrong.
 

After I asked to recalculate the result as it might be incorrect, ChatGPT provided the correct answer —with what looked like some "snarky" comment about how the previous one was wrong, and the right calculations this time: 23=8, √(8)=2.828...

Conversation with ChatGPT. Following the previous question, the user states that the answer may be wrong and to run the calculation again. ChatGPT apologizes for any confusion, and states that the previous answer is wrong. It does the calculations again, this time with the correct values and result.

Super-verbose answer, addressing the previous error more than running the calculation
 

So I guess ChatGPT is coded to provide wrong answers on purpose. But why? Is the system programmed to "test" the user? Is it some A/B testing? What and why?

It could be to get publicity online (people like me sharing about incorrect results and how unintelligent this Artificial Intelligence thing is). It's plausible, but I doubt it's just that —it would be bad publicity, too, but as they say, "Bad publicity is better than no publicity."

There has to be more to it. What am I missing?

Top comments (4)

Collapse
 
jonrandy profile image
Jon Randy 🎖️ • Edited

It's because it doesn't 'know' or 'calculate' anything. It's a dumb (but impressive) language processing model - it's not intelligent in the slightest

Collapse
 
auroratide profile image
Timothy Foster

I don't think it's on purpose; it's probably more an artifact of how this AI processes language.

I tried the same experiment, but just said "That is the wrong answer". It proceeded to double down.

The AI is told its previous answer is wrong, and it responds with the exact same answer.

I figured the difference between what you put and what I put was the word "calculate", so I tried asking "Calculate the square root of the cube of 2". Interestingly, it's correct now.

Asked to calculate, the AI gives the correct answer this time.

Something about the word "calculate" might be cluing the AI that an actual calculation is needed. So my hypothesis is that the AI is often wrong because 1) language is extremely hard, and while this is good it ain't perfect, and 2) it's optimized for conversation, not accuracy.

Collapse
 
auroratide profile image
Timothy Foster

Actually, they address this in the first two points of their limitations.

during RL training, there’s currently no source of truth

given one phrasing of a question, the model can claim to not know the answer, but given a slight rephrase, can answer correctly.

Collapse
 
raldincasidar profile image
Raldin Casidar

My friend also experienced this yesterday. We dont know why AI provides wrong answer sometimes