Recently, an American lawyer was prosecuted for using quotes from non-existent cases in court. They were provided to him by ChatGPT. At the same time, according to the lawyer, the bot assured that these cases are still real. Though, ChatGPT is still great to use. If you enjoy reading and writing it is definitely a great helper. If you get bored and tired and want something more interesting. Then join the Bizzo Casino login and stay tuned.
This is not a unique case. Neural networks are prone to hallucinations, that is, inventing and distorting facts. We explain why this is happening.
AI hallucinations occur when it produces an answer that is as close as possible to the truth according to the data it has. This happens quite often.
Hallucinations start when there is a gap in knowledge, and the system chooses the most statistically correct information. That is what Pierre Haren, the CEO and founder of Causality Link explains.
The reasons may be different. But most often this happens if the bot was trained on a relatively small sample. It was not taught to say that he did not know the answer. According to Haren, AI says something that is very different from reality. But it does not know about it. Moreover, it does not lie. The question arises from such information. What is the truth about the ChatGPT brain? Is there a way for the system to know that it is deviating from the truth? he muses.
Arvind Jain, CEO of Glean, notes that it’s important to remember that the most likely answer isn’t always the right one.
What Can This Lead To?
First, to the spread of misinformation, since such answers seem realistic, although they are false. In addition, there can be a domino effect if the bot provides incorrect information on a topic that carries high risks. Therefore, experts recommend not using technology when errors are unacceptable.
The creators of large language models will not be happy if someone uses their system to control a nuclear power plant. Though, as humanity enjoys working with less effort, it can not be changed.
However, a bot can be unreliable in other situations as well. So, if an organization uses the misinformation provided by them in an advertising campaign, then they risk their reputation. Problems can also arise if the AI gives the wrong financial advice.
Overall, the AI hallucinations show that the technology needs to be developed and deployed responsibly, taking into account issues related to ethics and safety.
What Measures Can Be Taken?
How can you be sure this answer is correct? This question is up to the people. But now tech companies are introducing guarantees to explain this relationship transparently. If you slow down the development of AI, you can take measures that will help you use the technology safely. Some creators think that this is especially important because when a model gets the right answers most of the time, people start believing it all the time.
The first step is to have the models provide the sources from which the information was taken. According to experts, a good motto is to trust and verify. The absence of the latter can lead to the spread of disinformation.
If we don’t check, then we will allow something like the Mandela effect. It will seem true, simply because no one has checked, and everyone says and repeats it. Another way is to ask the AI to report when it doesn’t know the answer. According to its creators, large language models try to provide at least some result, because otherwise, they do not receive points. It needs to be made clear that non-response is acceptable if the information is not found in the training data and in the provided context.
In addition, it should be possible to report hallucinations and errors. Such a mechanism is implemented in ChatGPT. The user can put a “thumbs down” next to the answer and write how to improve it.
Eventually, such questions can become philosophical. Why? Do you trust all information on the Internet? How can you prove that what you see on the internet ever happened or provides you with the right information? As all bots are taking the information on the Internet you will never be able to prove what is true information and what is not. Besides, what about political influence? Such questions will have answers.