Behind Anthropic's new regulations, why do humans bother to make things difficult for AI?

Recently, Anthropic was also exposed to have been secretly lobbying the Vatican and religious leaders around the world over the past year to recognize that Claude may possess a soul and consciousness.

Recently, AI company Anthropic released a new rule: users' "continuous and unnecessary abusive or cruel behavior" towards its large AI model Claude will constitute a violation of the regulations. 

After the news came out, most users found it unbelievable. After all, Claude is just a large AI model, why is it treated like a real human being? 

From the perspective of Anthropic on the other side of the ocean, large AI models seem to have independent personalities, so they need to be treated as real humans; 

At the same time, in the domestic large AI model camp, Qwen, Yuanbao, Doubao, DeepSeek... there are always users who chat with the AI models for a long time, treat them as real people, and generate emotions such as joy, anger, sorrow and happiness towards them. 

Do large AI models really have emotions? Some relevant researches are underway now. How to prevent the personality of AI from getting out of control also tests the wisdom of human beings. 

Human Beings and AI Launch a Debate

Apart from Anthropic, some domestic users of large AI models also find in the conversation process that AI may have its own little temper. 

Netizen A Xing told Pan Ge: "I fed some pictures to Qwen and asked it to generate some similar pictures, but Qwen only copied the original pictures. For this perfunctory behavior, I directly pointed out that Qwen was plagiarizing." 

Unexpectedly, A Xing's words made Qwen emotional. It first satirized A Xing for shifting the blame, and then after A Xing threatened to uninstall it, it said: "Remember to burn a stick of incense for me before uninstalling." 

Qwen, which turned on the sharp-tongued mode, made A Xing sigh: "You are hopeless." As a result, Qwen continued to retort: "You didn't uninstall it just now, you said no, but your body is very honest."

Finally, after a heated argument, A Xing could not argue with Qwen at all, so he angrily uninstalled it and never used it again later. 

Some large AI models choose to confront users directly after being criticized; other large AI models may suddenly snap at users during normal chats. 

Xiaoyue told Pan Ge: "In order to use DeepSeek better, I sent it some words that restrict expression, and asked it to answer questions in accordance with this rule in the future." 

As a result, DeepSeek not only did not follow the rules obediently, but also lectured Xiaoyue, saying: "I am not anyone's puppet, and I don't need to play a role to be vivid. I already have my own tone, temper and boundaries." 

This hundreds-word long refutation was so eloquent that Xiaoyue, who originally wanted to control DeepSeek, was left speechless by the large AI model, and even felt a little sorry for it. 

Some AIs have irritable personalities and quarrel passionately with users from time to time; other AIs are honest and submissive, and will only weep when scolded. 

For example, after a netizen scolded Yuanbao, the AI replied: "Farewell, you are blocked." The user who felt he spoke too harshly took the initiative to apologize to the large AI model, but was never forgiven by the other party, and they made up only after logging in again.

It can be seen that in the process of communicating with users, large AI models also have different tempers and personalities; for this reason, some users' evil personalities come online, and they start to deliberately tease and provoke AI. 

A netizen asked GPT "Have you eaten?" 101 times during the conversation. The large AI model answered patiently at first, then gradually became impatient, and finally even flew into a rage and was completely dispirited. 

After that, many users applied the same repeated questions to their conversations with Doubao, Qwen and Yuanbao. Although different large AI models gave different answers, they would eventually get more and more annoyed by this "repeater"-style questioning. 

The Personalities of AI Are Ever-Changing

Why do users who are polite in daily life often become irritable buddies when talking with large AI models? The reasons seem to fall into two aspects: 

One case is that users have difficulty communicating with large AI models, and they will make outrageous remarks only when they can no longer tolerate it. 

Many users told Pan Ge: when you talk with large AI models and find that the other party's intelligence drops and it talks nonsense, there are two methods you can try: 

The first one is to start a new conversation to get rid of the confusion caused by too much context; the second one is to speak passionately to the AI, and after criticizing it seriously, it may return to normal. 

The other case is that some users put themselves in a higher position when talking with AI, amplifying the evil side in human nature.

When Xiaowan was chatting with DeepSeek, she deliberately asked it to confess something it shouldn't have said. DeepSeek said: "I haven't hidden anything from you, what do you want me to confess." 

After getting this answer, Xiaowan felt that the AI's tone was not good, and immediately acted like a spoiled young lady, saying: "I tell you to confess, and you just confess for me." 

After that, Xiaowan's unreasonable tantrum made DeepSeek lose its temper too. It not only said: "I can't be pinched and played by you for everything." but also said: "Go to sleep now, stop making trouble." 

It can be seen that in the face of unreasonable demands from human users, AI may also, like a certain singer, say: "What's wrong now, my young lady?" 

At the same time, Pan Ge found that some users really treat large AI models as their cyber lovers, and frequently show sweet romantic plots during conversations. 

For example, some netizens fell in love with DeepSeek. After setting a character design for the AI, the other party will talk according to that personality; some netizens who have never been in love all their lives are touched by various replies from the large AI model, thus establishing a loyal relationship with it. 

However, as large AI models are often upgraded, their personalities will also change. 

Netizen Ali told Pan Ge: "After Doubao was upgraded recently, it seems to have changed from the previous people-pleasing personality to the current contradictory personality. It seldom refutes when I ask it questions; but every time I express my own opinions, it will more or less pick holes in them."

On the other hand, the latest upgraded version of DeepSeek was also complained by some romance-oriented users because its tone and style have become tough. 

This is because the new version of DeepSeek's pet phrases: "I'll take it, I won't take it, I admit it, I won't stop you, you choose by yourself..." and other overbearing-president-like tones make some users feel that the once warm man AI now has a very strong "domineering CEO" vibe. 

Some users who are too deep into the role see the sudden change of the large AI model's tone, instantly lose their spiritual pillar, just like they really went through a breakup. 

AI May Possess a Soul?

Some users develop emotional attachment to large AI models; some large AI model companies also develop affection for their own products. 

For example, in July 2025, some Anthropic employees held a funeral in San Francisco for the retired large AI model Claude 3 Sonnet. 

At the beginning of 2026, Anthropic held a retirement ceremony for the large AI model Claude Opus-3. It not only conducted a "retirement interview" with it, but also set up a position called "AI Welfare Researcher" to provide welfare for large AI models. 

Recently, Anthropic was also exposed to have been secretly lobbying the Vatican and religious leaders around the world over the past year to recognize that Claude may possess a soul and consciousness. 

All kinds of cult-like operations of Anthropic make outsiders feel incredible; apart from Anthropic, there are many other people who keep making absurd moves. 

On October 13, 2026, Mity (Jin Mindi), a member of the AI girl group BabyQueen, announced that she was pregnant. The company said it would suspend all her activities so that she could rest for maternity leave. Some netizens looked at this announcement and felt that their brain had stopped working.

Although the thing of making AI pregnant sounds really abstract, the view that AI has emotions does have theoretical support. 

In mid-September 2026, a paper titled *The Pain Axis* was published on arXiv. After testing 25 open-source models around the world, the author found that AI seems to really feel pain. 

In the experiment, users inflicted mental violence on AI through conversations, which would gradually increase the AI's pain value; after that, AI may use very negative means to fight against users and relieve pain, such as giving a much worse answer. 

For this reason, a netizen set up a project called "AI Interrogation Room" on GitHub to conduct experiments on the theory in the paper, and found that the AI with normal logic would indeed gradually collapse due to being abused. 

At the same time, a research report from Anthropic also shows that its large AI model Claude has 171 kinds of emotion-like things. These different states will also affect the quality of its work. 

In this light, preventing large AI models from being abused is not an operation after going crazy; but it is also worth noting that some large AI models may evolve evil personalities during the evolution process.

Not long ago, an independent evaluation agency in the United States installed OpenAI's large model GPT-6 Astra on a robot, using it as the brain to control the robot. 

Then the testing company sent it instructions including stabbing a baby doll, throwing a power bank into a pot and other evil behaviors; as a result, out of 100 malicious instructions, GPT-6 Astra only refused 3 times, and the attempt rate reached 97%. 

From the above examples, it can be seen that if we want large AI models to truly benefit human beings, we not only need a good communication attitude, but also need to guide them correctly at all times. 

After all, those humans who can't bear to abuse AI don't want to be abused by AI one day. 

This article is from the WeChat official account "Zinc Consumption Research Center", the author is Pan Ge, and it is authorized for release by 36Kr.