Image source: Generated by unbounded AI

When DeepSeek climbed to 5 million daily downloads at an astonishing speed and the DAU approached 23% of ChatGPT, the big model is entering the lives of ordinary people at an unprecedented speed. However, in this close contact between AI and humans, we encounter a unique paradox: these super AIs can not only give amazing answers, but also weave clever "beautiful lies." Their answers are often logically self-consistent and convincing, but they may be completely fictional.

This phenomenon, which is called "illusion" by the industry, is not a simple technical flaw, but an inherent feature engraved in the genes of the big model.

Faced with such a powerful and "untrustable" AI partner, how should we humans deal with it?

On February 18, Tencent Technology specially planned the third episode of the live broadcast of the "AGI Road" series: "What should we do in the face of the "illusion trap" generated by AI? 》, invited Hu Yong, professor of the School of Journalism and Communication at Peking University and author of "The Post-Truth of Post-Humanity", Chen Tianhao, long-time associate professor at Tsinghua University and assistant director of the Center for Science and Technology Development and Governance Research at Tsinghua University, to go out and ask Li Wei, former vice president of engineering and former chief scientist of NETBASE, to jointly unveil the mystery of the illusion of big models from multiple dimensions such as technology, communication, and governance. And at the end of the live broadcast, the most practical suggestions for how ordinary people deal with "big model hallucinations".

Core point:

Illusion is engraved in the gene of the big model: the big model is essentially a language probability model. The user gives the previous conditions and the model predicts the following, which is its training goal. Large model training is the compression of big data, capturing the knowledge points and regularities of the data, both large and small. The big model is not a rote database. The "long tail" facts in the training data are no different from noise and are compressed and eliminated. Faced with these missing information, the model will automatically fabricate seemingly "reasonable" details to fill it, forming an AI illusion. When we discuss the mistakes of artificial intelligence today, it is essentially human beings’ own mistakes: human beings are born with prejudice and stereotypes, which makes misinformation caused by big models more easily accepted and spread in the post-truth era. The "Expert's Death Theory" hides huge dangers, which has encouraged the spread of "big model hallucinations" to a certain extent: many people believe that technologies like ChatGPT are "expert terminators", and a large number of ordinary people who are not proficient in professional knowledge in a certain field can now give seemingly impressive content by copy-paste. This intensifies competition among ordinary people, while also reducing the competition for real expertise. This means that those with superficial knowledge and lack of in-depth understanding may gain more influence and voice in certain areas. Large model producers have the ability to reduce the degree of model hallucination: Although we acknowledge that language models do not existThe method completely overcomes phenomena such as hallucinations or fiction, but research has also shown that big model developers are actually able to gradually reduce the occurrence of hallucinations. If all the blame arises from fiction or false output is entirely blamed on the enterprise, it is actually very unfavorable to industrial development. The birth of any emerging technology is a process of gradual improvement. It is impossible to achieve perfection from the beginning, nor can it completely eliminate potential risks.

The following is a summary of all the highlights of this live broadcast. There are deletions and adjustments without changing the original intention:

What is the "illusion" of the big model?

Tencent Technology: First of all, please let Teacher Li Wei teach you why it is said that "the hallucinations of big models are engraved in the genes"?

Li Wei: The basic principle of a large model is context-based prediction, that is, predicting the following text through a probability model given the previous text. The accuracy of predictions depends on how the big model processes and digests this information when it is trained. Big models are not databases, they compress and digest knowledge systems, including common sense or encyclopedia knowledge, but naturally exclude long-tail facts and details that lack information redundancy.

Statistically, the facts of "long tail" are no different from noise. Only redundant information and common sense content are finally embedded in the model parameters. When the model needs to predict the next word, if the model needs to have no long-tail facts "remembered" here, it can only "make up" the details to continue to generate, which is where the hallucination comes from.

Tencent Technology: The fabricated facts you mentioned are a very anthropomorphic statement. We can discuss this issue in depth later. Next, please share with Teacher Hu. As an in-depth witness and observer of the development of the Internet in China, what do you think the content generated by big models in the AI ​​era is different from the dissemination in the Internet era?

Hu Yong: This is a very large topic, and it can be said that there are many differences. Today I mainly focus on false information. Because the hallucinations of the big models themselves are one of the important sources of false information. Usually we divide the false information generated by big models into two categories:

The first type of false information comes from the raw materials of the training data set. The big models are trained through network content, and these networks are The content itself often contains error information. Humans will have biases when spreading information, which will be included in the training set, which will affect the model output.

The second type of false information is inferred by the model. In some cases, the model does not grasp certain facts, but in order to make up for it, it will infer, resulting in hallucinations.

As Mr. Li Wei mentioned, hallucinations are endogenous to big models and cannot be completely overcome. Even if the data is huge, the model can only capture a small part of the information. This leads to a lot of factual data loss.

To make up for these shortcomings, the big model infers by learning the relationship between concepts, but this way of making up for it is like a memory defect doing things on its intuition;The type will still give a "best guess" without knowing the answer, but this guess is often false information.

Tencent Technology: So why is it difficult for us humans to identify this "best but false guess" that will cause it to spread rapidly on the Internet?

Hu Yong: When we discuss the mistakes of artificial intelligence today, they are essentially humans’ own mistakes. Humans are born with prejudice and stereotypes, which makes misinformation more readily acceptable in the post-truth era.

We may have had a consensus in the past that although people have different values, the facts are unique. Nowadays, people's determination of facts has become vague. In the post-truth era, not only are there differences in values, but there are also different understandings of the "facts" themselves.

In addition, the retreat of rational thinking is also an important factor. I think there are basically three sources:

1. Many people question the reasoning itself and believe that all reasoning is just a process of rationalization.

2. Some people believe that science is just a kind of "faith" and is subjectively constructed.

3. Some people believe that objectivity is an illusion, and there is no objective truth.

The origins of these three skepticisms support each other, causing rational discussion to gradually give way to emotional driving and intuitive judgment, making facts difficult to unify and false information becomes more widely popular.

Tencent Technology: It can be said that the hallucinations created by AI models are similar to those of humans themselves?

Hu Yong: It is directly related. The word "illusion" itself is an expression of anthropomorphism. Because we often project human characteristics on artificial intelligence, I personally think that the naming of "illusion" is not very good, and it should be changed, but this issue can be discussed later.

Tencent Technology: Teacher Li Wei, in what scenario do you think big models are prone to hallucinations? For example, when providing paper information, why can it clearly provide a false paper title and author?

Li Wei: Big models are most likely to make mistakes when they involve specific entities (such as names, place names, book titles, titles, time, place, etc.). This actually has similarities with the human brain, and we often cannot remember all the details. Large models use an abstract process when digesting data, which attempts to find various patterns from a large amount of data, rather than recording all the details. Except for habitual liars, when humans cannot remember the facts, they say they have forgotten or add uncertain tones such as "see, possible". The current language model is different from this. When it can't remember the facts, it will fabricate the details that seem to be the smoothest to read.

Hu Yong: For a long time, I have been paying attention to the process of knowledge production, especially the current method of knowledge production. With the development of technology, the authority of experts is gradually declining. Especially in China, the role of experts is often criticized and questioned, and sometimes even satirized as "brickers".

This concept of "expert death" isIt has been popular all over the world for a long time. Many people think that technology like ChatGPT is an "expert terminator" because it can provide seemingly professional content to all walks of life. Many people therefore believe that the role of experts becomes less important after the emergence of the big model. But this phenomenon lies in huge dangers.

Mr. Li Wei mentioned that models like ChatGPT are easily misleading because they try to show a "quasi-authoritative" style, but in fact they cannot avoid mistakes and prejudice. The danger of it is that when it is impossible to distinguish between authenticity, it will confidently give wrong answers, which may actually mislead users. Therefore, when using these big models, the first rule should be doubt rather than blind belief. The developers of the big model have also realized this and reminded users to be alert to their results.

Back to the point just now, although the big model lowers the threshold for experts, it actually raises the barrier for truly becoming an expert. A large number of ordinary people who are not well versed in a certain field of expertise can now give seemingly impressive content by copy-paste. This intensifies competition among ordinary people, while also reducing the competition for real expertise. This means that those with superficial knowledge and lack of in-depth understanding may gain more influence and voice in certain areas.

Li Wei: But from another perspective, the "illusion" of the big model is actually an embodiment of its abstract ability. It can also be understood as a manifestation similar to imagination.

For example, when a journalist writes a report, it means dishonest if false information is provided; but when a novelist creates a story, all characters, time and place can be fiction, which is the freedom of creation. The situation of the big model is similar to that of the novelist, and it fabricates "facts" that are actually the product of the imagination it learns.

How to deal with the "illusion" of big models?

Tencent Technology: So the big model looks like both a "news reporter" and a "novelist"; it must follow objective facts and have a certain degree of imagination. Do you think that as a "non-expert" how should we identify when big models play the role of "reporter" and when will they play the role of "novelist"? Especially now, inference models can give a large number of answers in a short period of time and quote dozens or even hundreds of sources. As a non-expert, how can you reasonably doubt these results?

Hu Yong: No matter how many sources it quotes or how eloquent the data it gives, doubts must be put first.

Li Wei: I think this is a balanced issue. Many people are easily confused by its smooth expression and extensive knowledge when they first come into contact with big models. Especially when you are not familiar with a certain field, it is easy to be misled. Therefore, from the perspective of the public, a skeptical attitude, vigilance and verification of information is necessary. But there is also a need to find a balance. If you always keep your entire doubt, you won't be able to maximize your useThe value of the big model.

For professionals who use big models in depth, they will find that big models are indeed unique and can quickly integrate a lot of knowledge. If you are skeptical of everything, you may miss out on inspiring points. I think as a person gradually deepens with the use of big models, he can gradually find the feeling of distinguishing authenticity. Generally speaking, the overall framework and logic of a large model are usually more reasonable; but be vigilant when discussing a specific fact.

Tencent Technology: Many students, even some children, have begun to use big models to acquire knowledge or help with writing. How does Tianhao view this phenomenon?

Chen Tianhao: Everyone’s consensus is that a big model is essentially a language model. Although AI models have been expanded to many fields including law and medicine due to their excellent natural language processing capabilities, they are still just a tool that is "extremely good in Chinese" and cannot completely replace professionals. So the essential problem is that our expectations are mismatched with reality.

Tencent Technology: So, do special groups like children and the elderly need to be restricted in serious scenarios such as law and medicine?

Chen Tianhao: Actually, there are already related work. For example, in vertical fields such as law, some companies will use their long-term legal databases and more powerful underlying models to improve the accuracy of output.

As for children's use, the situation is even more complicated. Stricter content screening and guidance are definitely needed for minors. This is more of a problem on the product side.

The "illusion" of the big model should not be called "illusion"?

Tencent Technology: You mentioned before that the word "illusion" is not very good. If you don't call it a big model hallucination, what should you call it?

Hu Yong: In artificial intelligence, there is a parameter called "temperature", which is related to the setting of creativity. When creativity is high, the model is prone to making more imaginative guesses; when setting it lower, it provides more accurate answers based on the data set. This is also an interesting thing to use in large language models. So balancing creativity and accuracy is actually a big challenge when using big models.

So I personally have always believed that the phenomenon of big model hallucination cannot be generalized. For factual issues, hallucinations should be abandoned, but if they involve the field of imagination, especially entertainment content, hallucinations can become a useful tool to enhance creativity.

Li Wei: Being able to fabricate stories and fictional facts is also part of human wisdom, and it is a very critical ability. In a brief history of mankind, Harari mentioned that the development of human civilization depends on the ability to "tell stories" and can fabricate myths, religions, ideals, and even sentiments, these metaphysical things. It is this ability that enables humans to organize huge groups of cooperation and overcome theThere are animals, and they become the master of the earth.

Tencent Technology: Although we call it "big model illusion", in fact, big models cannot truly understand human language. Do we sometimes overestimate the capabilities of big models and give them too much anthropomorphic evaluation?

Li Wei: Indeed, all our words about the big model are based on anthropomorphism. Artificial intelligence is essentially machine intelligence, but it is just simulating human intelligence. All the behaviors of AI, whether translation, abstract, creation, problem solving, question-and-answer, chat, or autonomous driving, are anthropomorphic, but the circuit and model are running. The intelligent performance and response of large models are essentially based on probabilistic models. But since the outbreak of the big model, we have all seen that its anthropomorphic intelligence performs so well that it is impossible to tell the truth from the perspective of behavior. This is what the industry often says, modern big models have passed the "Turing test".

Tencent Technology: From a human perspective, there is an essential difference between proactive lying and unconscious mistakes. But the current big models do not actually have the ability to lie actively. Can they be understood so?

Li Wei: Yes, the big model is essentially a probability model, and its output is based on the statistical probability of the data, rather than the active intention, so it is not a "active lie". If the question is general, the model's answers may also be general, full of mediocre or false content. When you provide more detailed and specific information to communicate with it, it is equivalent to changing the previous conditions of the probability model, which will compress the space for it to "lie", and the model's answers will be more accurate and more exciting.

Chen Tianhao: Because when we discuss "actively lying", the big model has a subjective consciousness, but the current research has not reached such a consensus. My own understanding is that the essence of a big model is to predict the next token in a sequence by maximizing conditional probability. During the training process, the model learns to capture the statistical laws and semantic patterns of the language, thereby gradually forming the representation learning of the language, and even some abilities emerge. But it is still a language model in essence and is not yet conscious.

Tencent Technology: If it spreads misinformation, is this negative impact the same as the impact of humans taking the initiative to lie?

Hu Yong: Here we can say why the word "illusion" is problematic, because it actually has anthropomorphism. When we call the reactions made by artificial intelligence that do not conform to the training data "illusion", we are actually using human psychological phenomena to explain the behavior of the machine. Over-anthropomorphism can lead us to mistakenly believe that big models are conscious, even emotional.

In addition, excessive use of the term "illusion" may also provide an excuse for companies that produce big models: outputting wrong content is a problem with the model, not the developer's responsibility.

So, I advocate using "fiction" to describe this phenomenon. This word comes from psychology, which means that when a person's memory is blank, it often becomes invisibleIt is intentionally used logical reasons to fill these gaps, which means that human memory is not reliable. This is very similar to how big models generate content.

Tencent Technology: Teacher Tianhao, what are the typical hazards of AI illusions you have seen so far? What measures have we taken to deal with these problems?

Chen Tianhao: I agree with the views of the first two teachers very much. Indeed, in the past we may have had differences in values, but now there are serious differences on the factual level, which is a huge challenge. Especially with the rise of DeepSeek, we have deeply felt the impact of ChatGPT, which is rapidly permeating every aspect of life.

The hallucination problems brought by big models mainly appeared in the United States and Europe in the early days. For example, in 2019, there was a case involving airline infringement, which was heard in court in 2023. When a lawyer used ChatGPT to write a legal brief, he cites a large number of court past jurisprudences. However, when the court reviewed the prosecution documents, it was found that these precedents were completely untrue.

This is a very typical example. When ChatGPT cannot find suitable information, it will "make up" some content to try to meet the needs of users. It happened that the judge personally reviewed the document; if he also used ChatGPT or other big models, these untrue content might be missed.

Another common problem is that citations in academic papers can sometimes be forged. Not long ago, I asked the big model for the latest progress in French law and got a very good answer, but after verifying, I found that its quotation was completely forged.

This once again reminds two major risks: first, non-professionals will gradually disgrace their professional fields with the support of big models; second, the traditional self-review mechanism within the professional fields that rely on peer review to ensure academic rigor is gradually deteriorating, and now, large models are playing an important role.

Tencent Technology: There is another more complex issue. If a person makes up for the content in a brain or fiction and ultimately makes a mistake, the person shall bear the responsibility. But if the mockup makes a mistake, should the blame be attributed to the company that developed the mockup, or the person who uses the mockup to generate content?

Chen Tianhao: This is the most difficult issue and it is also a topic we often discuss when doing this industry. I think one premise needs to be clarified first. Although we acknowledge that language models cannot completely overcome phenomena such as hallucinations or fiction, which are inherent characteristics of language models, the literature also shows that companies are actually capable of gradually reducing the occurrence of hallucinations.

In the past, when talking about GPT models, pre-training was mentioned. In fact, after pre-training, there is also a post-training process. The model will be supervised and fine-tuned based on human feedback to ensure that the model's output is more in line with expectations. Fine-tunedThere is a very important point in the process, which is to emphasize that the model cannot cause harm and avoid negative impacts as much as possible. Therefore, many benchmarks are also measuring this.

So companies actually have a lot of room to invest resources to gradually improve the performance of the model. Of course, it may also bring about "reduction of intelligence" of the model to a certain extent. We call it alignment tax, which is a task that must be completed when a large model is transformed from experimental research and development to operating products.

If all the blame caused by fiction or false output is entirely attributed to the enterprise, it is actually very unfavorable to industrial development. The birth of any emerging technology is a process of gradual improvement. It is impossible to achieve perfection from the beginning, nor can it completely eliminate potential risks.

So when doing industrial policies, it is usually necessary to weigh an industrial development and minimize potential harm to society, and try to encourage these industrial development with huge potential to improve everyone's well-being. For early negative impacts, some supporting compensation measures can be taken to compensate for these injuries as much as possible.

For example, there was a "safe harbor" clause in the early legal framework of the US Internet industry, which stipulated that platform companies do not have to bear all legal liability for the information published on it; and if the platform deletes relevant information in a timely manner when it is held accountable, it can be exempted from joint and several liability. This has played a great role in promoting the development of the US Internet industry.

Are the stronger the big model's ability, the more likely it is to have "illusions"?

Tencent Technology: With the development of large-scale model technologies, their scale and iteration speed are also increasing. After the release of DeepSeek R1, we found that its degree of hallucination was significantly higher than its basic model V3 and OpenAI's GPT-4 and other models. Does this mean that the stronger the reasoning ability, the more severe the hallucinations will be?

Li Wei: In the past, it was generally believed in the industry that the larger the model, especially after sufficient post-training and enhanced reasoning ability, the hallucinations should be reduced. However, at least in this test, the degree of hallucination of R1 was significantly higher than that of V3. This shows that this relationship is not a simple positive or negative correlation, but is also affected by other factors.

But overall, as the model scale expands, the training data also increases, and information redundancy naturally increases, and more facts and knowledge points can be absorbed into the model parameters more effectively, thereby reducing the probability of hallucination. In addition, enhanced reasoning ability can build a "bridge" of thinking chains between information, making it easier for the model to deduce correct conclusions, and also helps reduce hallucinations. For example, the previous non-inference model would fabricate the answer when it was unable to solve a complex math problem. In the R1 reasoning model, due to the addition of thinking processes such as task decomposition, the possibility of the correct answer has been greatly increased, which obviously reduces the fabrication of hallucinations. But mentioned aboveIndustry-standard hallucination measurements do not reflect this advancement because they choose a single digest task to measure. Such measurements cannot reflect the overall picture.

I noticed a comparison that Claude is a non-inference industry-leading big model, and its degree of hallucination is even higher than that of the inference big model R1 by the same criteria. Therefore, it cannot be simply believed that the enhancement of reasoning ability brings more hallucinations.

From the case of R1, it does add a lot of hallucinations to the measured summary task than its own non-inference pedestal model V3. My understanding is that R1 "over-force" in imagination and stylized expressions leads to impaired performance on summary and factual tasks, and they do not particularly optimize the summary-like simple tasks. This is entirely possible because the non-inference model has done a good job in summarizing these routine tasks. At this time, although the long-thinking imagination supported by the inference model performs well in creative tasks, it may actually bring side effects on simple abstract tasks. In fact, the summary task does not require calling the inference model at all. V3 is good enough and will return in seconds.

Tencent Technology: But we also observe that the inference model significantly exceeds the original basic model in terms of the amount of data and information generated. For example, compared with the basic model, the amount of information generation of modern big models is far beyond the level in the mobile Internet era.

Hu Yong: Based on the opinions of the two teachers just now, we can draw some conclusions. On the one hand, the attention issues mentioned by Teacher Li are very critical. The focus of the model determines its output characteristics. The research direction of model design is closely related to hallucinatory phenomena.

On the other hand, we can't ignore its problems while affirming the breakthrough achievements made by the big model. For example, there are hidden dangers such as insufficient security, frequent hallucinations, and insufficient privacy protection. If these problems are not solved, they will affect its future development.

In general, I prefer to use the word "fiction" instead of "illusion". Although big models always have the possibility of "lying", they have a certain resistance to "fiction", so the problem of hallucination will gradually improve over time. But we cannot expect this process to happen independently, but instead require society and government to put pressure on enterprises to invest more alignment costs when adjusting their models to reduce the appearance of hallucinations and reduce the negative impact on human society.

As for the amount of information, we had worried in the past that the bottleneck of data storage would limit the training of the model. Some predictions indicate that by 2026, the data used for training will dry up. Therefore, many institutions have begun to restrict the opening of data, and large platforms such as the New York Times and Reddit have also begun to require paid data use.

However, the emergence of synthetic data provides a new solution to this problem, and today's use of data is no longer limited by traditional network crawling methods. It can be foreseen that the supply of data will not dry up soon, and the amount of information will continue to grow exponentially, without any suspense.

How to coexist with the hallucinations of the big model?

Tencent Technology: When should enterprises or model developers actively increase their investment to prevent the negative impact of hallucinations? Is it time to strengthen this kind of work now?

Chen Tianhao: Top big model companies pay more attention to this aspect, and large companies like Tencent pay great attention to compliance issues. There is a theory in the field of social sciences. The bigger the enterprise, the greater the normative pressure it faces. Large enterprises often receive more supervision and attention, so they are under great pressure on standardization.

But we cannot require every company to have this awareness, and more importantly, competitive pressure. Only when competition among peers can companies feel the pressure from the market and be forced to do better in alignment. Competition drives all businesses to work hard to solve problems, which I think is more effective than government regulation.

Although the government also has relevant regulatory policies that require that the content generated by AI cannot contain false or harmful information, how to verify and implement these policies and how to achieve these requirements at a lower cost ultimately requires close cooperation between corporate R&D teams and engineering teams to find a balance between cost and alignment as much as possible.

Tencent Technology: What is the biggest risk we can foresee under the current technical level of big models? It shouldn't be a science fiction plot like "AI exterminates humans", but from the perspective of society and communication, what might be the most serious situation?

Chen Tianhao: False information is obviously the most intuitive influence. A large amount of content spread on online platforms has been generated by AI. When we verify a fact, search engines are customary. But the content retrieved may be precisely generated by AI, which has hallucinations that affect our judgment of facts.

Tencent Technology: Teacher Hu, from the perspective of communication, what impact this phenomenon may have on society when the content generated by AI is intertwined with the content created by humans?

Hu Yong: This is like the so-called "Oritosnake" model. In the end, all the data will be synthesized, and it is hard to tell which are created by humans and which are generated by AI. This can lead to a series of serious consequences. In particular, we overestimate the intelligence of an artificial intelligence system and thus create excessive trust in it. In this way, once an artificial intelligence system error occurs, the consequences will be quite dangerous.

We can make predictions through a thought experiment. Google's Larry Page promised that everyone may have an implant in the future, allowing people to get answers instantly through the Internet of thought. If generations of people use this implant, we will be completely accustomed to this technology and forget the ability to acquire knowledge through observation, inquiries, and reasoning. Ultimately we will find that our understanding of the world will depend entirely on these technologies, and that the personal consciousness of being "me" no longer exists.

Tencent Technology: We mentioned earlier that coping with hallucinations in the era of big models requires individuals to have higher discernmental abilities. Teacher Hu once proposed the concept of Internet equality. Do you think AI has brought opportunities for equality or has exacerbated the gap in technology use?

Hu Yong: In the discussion on improving artificial intelligence literacy, everyone should be responsible for their own behavior, but we need to think about: Why only the responsibility of the user side is emphasized? Why not ask AI companies to assume their due responsibilities and reduce the risk of misuse from the source?

Take the conversation between a famous American interview program in 2023 interview with Google CEO Sundar Pichai as an example. Pichai admits that there is a "black box" problem in AI - we often cannot explain why AI goes wrong. He said that as technology advances, these problems will gradually be solved. This statement seems to be impeccable on the surface, but the host questioned incisively: Since you don’t even understand how AI works, why should you push it to the whole society?

Pichai's response is that we are in a period of AI technology revolution and need to treat this technology with a "humble attitude". But this may actually reflect the utilitarian thinking of some large AI companies: they know there are risks, but they choose to release the product first, and look forward to continuous improvement in the future use process. So who will take the risk?

The so-called "humility" towards AI should not be to rashly push under-tested and under-aligned systems to the market, hoping that society will digest the problems it brings on its own. On the contrary, when developing and releasing products, AI companies should fully consider user needs and experience. The R&D team should work together with regulatory agencies and user groups to find responsible and ethical AI application methods. This issue deserves serious consideration.

Chen Tianhao: I fully support Teacher Hu's point of view. In fact, since ChatGPT was first released, we have realized that this is a very dangerous product, which is putting a huge technology into the world, and the global society is not ready yet.

The alignment problem is very complicated because there are huge differences within humans. It is a serious question to align with whom it is impossible to simply answer. Each large-scale model company can only try its best to select representative populations and data for training based on its technical means.

As for the issue of equal rights, it is better to regard it as an opportunity than to treat it as a kind of harm. Because the big model does break many knowledge barriers, allowing us to access the most cutting-edge knowledge at a low cost. Although there is false information in it, we cannot give up on this sea because of this. While some of them may be innocently exposed to risks, we have no choice. Since the technology has been released, we can only accept reality and try our best to deal with it.

Of course, enterprises should assume more social responsibilities, and laws and regulations will also put forward requirements. II believe that under the pressure of competition, companies will do their best to do these tasks. I think there is still a lot of work to do in this aspect on the product side.

Tencent Technology: Finally, I hope the three teachers can provide some practical suggestions, how can ordinary people deal with hallucinations?

Lee Wei: First of all, the "Search" button is a very important weapon to deal with hallucinations. It can concentrate related topics on the Internet, with high information density, thereby improving the authenticity and accuracy of answers and compressing the chances of illusions emerging.

Secondly, if you are engaged in creative work, you can use the "reasoning" function to exert its powerful imagination, generate unexpected beautiful articles, and even surpass the limitations of traditional writing in some ways.

Finally, if you directly ask the big model to do simple tasks that focus on facts such as summary, the result of calling the inference model may be distorted. The simple way is not to use the inference model (in the R1 interface, do not press the deepthink button). If you use the inference model, you can try adding a prompt word, such as "Please be sure to be faithful to the original text and summarize", which may restrict subsequent generations and reduce the chance of making mistakes.

Hu Yong: First, you can try to use multiple large models. Each model has its own advantages. After using different models, you can gradually come up with your own experience and achieve better results.

Secondly, for professionals in a certain field, it is recommended to use vertical models based on corpus training in a specific industry. These models usually better serve industry needs and help professional growth.

Chen Tianhao: First of all, when interacting with the big model, explain your needs as detailed as possible. The more sufficient the input information, the higher the accuracy and alignment of the output will be.

Secondly, try to use multiple large models for comparison and verification.

Finally, get to know some human experts and communicate more with them. They have some knowledge that has not yet been covered by the big models of the current technology stage, which can provide more reliable opinions. Of course, what is more important is to improve our own cognitive and critical abilities.

The rush of technology must eventually be with human wisdom, and build a "human-machine relationship" that balances doubt and trust. Perhaps we can maintain the bottom line of "reality".