ChatGPT has been banned from entering schools. AI will change the rules specially. LeCun: How about we also ban small models?

If you want to ask which model is the most popular in the AI ​​industry in recent times, nine out of ten people will vote for ChatGPT. There is no way, this model is too popular, just because it can do things Too much. Many people describe it as a real "hexagonal warrior": not only can it be used to chat, search, and translate, but it can also be used to write stories, write code, debug, and even develop small games and take the American college entrance examination... Some people joke that from now on There will be only two types of artificial intelligence models in the future - ChatGPT and others.

On the one hand, ChatGPT’s capabilities are so amazing that it is loved by the public. On the other hand, it is also in constant trouble and has been banned by different agencies. Some time ago, there was news that Stack Overflow had banned ChatGPT for no other reason. The official stated that ChatGPT was temporarily banned because the accuracy of the answers it generated was too low. Publishing answers created by ChatGPT is very important to the website and users who query the correct answers. Harmful.

Not long after this incident passed, top artificial intelligence conferences also began to ban academic papers written using ChatGPT and AI tools. Let’s see what’s going on specifically.

ICML prohibits the use of large language models to write papers

A few days ago, the International Machine Learning Conference ICML announced that it is prohibited to use large language models (LLM, such as ChatGPT) to generate papers. thesis, unless the generated text is presented as part of an experimental analysis of the thesis.

##File address: https://icml.cc/Conferences/2023/llm-policy

According to ICML, although language models such as ChatGPT represent a future development trend, they are followed by some unintended consequences and problems that are difficult to solve. ICML said ChatGPT was trained on public data, often collected without consent, and who is responsible if something goes wrong.

There is also the issue of authorship, who "wrote" the paper: machine or human? This is especially important considering that ICML only prohibits text that is fully generated by AI. Conference organizers say they do not prohibit the use of tools such as ChatGPT to edit or polish text written by authors, noting that many authors already use semi-automatic editing tools such as grammar correction software Grammarly to polish their articles.

However, as soon as this news came out, it still triggered a big discussion on social media. Yann LeCun forwarded and commented: "Large language models cannot be used, so medium and small language models are Isn’t it still usable?"

Sebastian Bubeck, head of the ML foundation team at Microsoft Research, called the rule "short-sighted" and promoted it. It reads: "ChatGPT and its variants are part of future technological development. Banning is definitely not the optimal answer."

##ICML said , the ban on AI-generated text will be re-evaluated next year.

There have always been varying concerns about the harmful effects of artificial intelligence-generated text. One of the most common problems is that the output from these systems is simply unreliable.

AI models are trained to predict the next word in any given sentence, but they don’t have a hard-coded database of “facts” to draw from, so the model’s generated results often A paradoxical situation. Many times, the generated sentences conform to grammatical logic but do not conform to objective reality.

There is another problem that is also very difficult - it is difficult to distinguish whether the text is "polished and edited" by AI or completely generated by AI. This problem is very serious for papers. If the author uses an AI model to generate a concise summary based on the paper, is this considered editing the text or generating the text from scratch?

Of course, using AI tools like ChatGPT is not the only drawback. For example, in peer review, fluent English expressions tend to receive higher evaluations, and AI models can help non-native English-speaking paper authors generate more fluent English texts. This will save researchers time and create a more level playing field in the academic field.

But it should be noted that ChatGPT is a large language model (LLM), which is very different from simple grammar correction software such as Grammarly. The main function of LLM itself is not to adjust the structure and language of already written texts, but to generate some new texts.

In fact, almost no paper authors will actually use AI language models to generate academic papers. Because the requirements for academic papers are very strict, authors generally check the contents of the paper repeatedly before publishing. If there are errors in a paper generated using AI, the reputations of all named paper authors will be seriously affected, affecting their entire careers.

From the perspective of paper review, people need a fast and effective detection tool to distinguish whether an article is generated by AI or manually written. Recently, a developer released a new application called GPTZero that has this detection function.

GPTZero was developed by Princeton University computer science student Edward Tian and is evaluated using two different metrics Whether the text was written by an AI model: perplexity and burstiness. This may be using "model" to defeat "model".

New York City schools banned access to ChatGPT

At almost the same time, ChatGPT was banned by the school. The New York City Department of Education has blocked access to ChatGPT on its network and devices over concerns that AI tools could disrupt education.

Jenna Lyle, a spokesperson for the department, said: This ban is mainly due to concerns about the potential negative impact of ChatGPT on student learning, as well as the safety and accuracy of the content generated by ChatGPT. Worry. While this tool can provide quick and easy answers, it fails to develop the critical thinking and problem-solving skills that are critical to students’ academic and lifelong success.

In addition, ChatGPT also has problems that other language models have. Because its training data comes from the internet, it often repeats and amplifies things like sexism and racial bias in its answers. This type of language model also tends to fabricate information, from historical dates to scientific laws, but others generally cannot detect that it is fabricating information.

It is these factual errors that have educators particularly worried about this tool. Many teachers say that software like ChatGPT is essentially impossible to test students’ ability to write essays. If ChatGPT could help students write in a matter of seconds, they wouldn't want to put the effort into writing.

Others, however, argue that the education system will have to adapt to the emergence of this technology — just as it adapted to earlier disruptive technologies like Google Search and Wikipedia. But it obviously takes some time for everyone to adapt to ChatGPT.

Whether it is manually formulating and enforcing text writing specifications, or using some methods to tell whether text is generated by AI, it shows that people are aware of the urgent need to standardize the use of text generation models such as ChatGPT. The use cases and value of ChatGPT are still to be defined.

The above is the detailed content of ChatGPT has been banned from entering schools. AI will change the rules specially. LeCun: How about we also ban small models?. For more information, please follow other related articles on the PHP Chinese website!