According to a research, ChatGPT is dumb as €gg, but its text is fine

10.08.2023 – Despite the huge success of the past months, ChatGPT is dumb as an ass. At least something very similar claim the Purdue University skulls, and yes they are with a detailed test They point to the Achilles’ heel of artificial intelligence: It makes absurd claims all too often.

But perhaps the biggest problem is not in his mistakes, but in the fact that he presents all this so convincingly that it does not seem to people the first time, or perhaps the second time, that he is writing nonsense. Which it pretty much does anyway: The Purdue team asked the AI ​​517 questions that can also be found on Stack Overflow, to which 52 percent answered incorrectly. However, in 77% of cases, he did it abundantly saliva and beautifully formulated that 39% of people participating in the experiment simply preferred ChatGPT answers to human answers, if only because they were more convincing and informed.

The only problem with this is that 77% of preferred chatbot responses this way are wrong.

According to the study, users are only able to identify the error when the error is blatantly evident in the ChatGPT response. It is true that at least 2 out of 12 people still choose AI. However, when things are less obvious, they easily ignore it or simply underestimate the potential error factor.

But this also has a specific explanation. Test participants said they subconsciously chose ChatGPT because long, confident texts indicate genuine knowledge, and therefore they found it more useful and informative. But perhaps more importantly, he was always polite and soft-spoken, which helped him win people’s trust, and thus it was easier for them to accept an incorrect answer.

Researchers also note that ChatGPT often responds to questions and requests with text indicating a sure-fire solution. Like the “Of course I can help.”Obsession This will definitely solve it… while no longer drawing attention to possible errors or dangers. In addition, the AI ​​makes conceptual rather than factual errors, that is, many answers are incorrect because it is unable to understand the underlying context of the question asked.

Stupid ass but good lyrics?

Basically yes. Although the researchers note that with a larger and more diverse sample, they would get a more accurate picture of the error level in ChatGPT, which may be slightly better than this, but rather worse. Since ChatGPT went public, many people have come forward to note that the AI ​​is lying.

I experienced all of this myself with Hungarian content producers, and indeed: ChatGPT’s filth confidently and convincingly uttered absurd nonsense about celebrities, and when I pointed out his mistake, he first admitted the mistake, and then came up with another clearly presented lie. But this is the best case, there have been many cases where, after pointing out the problem, he simply denied that he was wrong.

OpenAI’s developers have stated that the software is only hallucinating, that is, it tries to relate two data pieces that are far apart as coherently as possible from the little information available. However, the problem with this is that the AI ​​doesn’t pay attention to all of this, and oftentimes doesn’t even acknowledge the error, which then convincingly tries to explain it with another lie instead of saying, “I don’t have enough information on that question.”

However, ChatGPT is still in its infancy. Even when she appeared, she had abilities that we previously had a hard time imagining in reality, and then developed more over the past few months. If these initial problems are eliminated, i.e. the ratio of bad to good answers improves greatly, and errors are acknowledged or at least anticipated, then they will be much more suitable for replacing and assisting certain functions/operations than they are currently – although they are already ideal for some tasks .

But until then:

ChatGPT lied so much about the host that they sued him for it