Abstract:
There was a post that went viral last night: Faced with a transfer of US$50 million, someone asked Claude, and Claude said not to transfer it. The money was not transferred in the end. According to him, Claude saved him $50 million. But judging from his narration,
he didn’t want to change his mind before asking Claude. The AI didn't make the decision for him, it just added a step to his original idea.
Some people on social platforms also saw it this way: He already had the answer in his mind and just asked Claude to sign the decision. Others said more directly, don't let Claude take the blame.
When asked what would happen if the question were asked about Doubao, he said that if he changed the model, he might get another answer.
If you change the model, the answer will change. Then the question may no longer be judgment, but support.
In April this year, Anthropic analyzed 1 million Claude.ai conversations. About 6% involved personal coaching, with health, career, relationships and personal finance accounting for 76%. Should I change my job? How should I continue talking to my partner? How should I deal with my debt? Some of these questions that I would ask friends, family or professionals in the past have been sent to the chat box.
Why do people like to talk to AI about their feelings?
Chatting with AI is easy. No need to make an appointment, no need to look at friends’ faces, and it won’t get impatient even if the story is told for the third time.
Some people say that chatting with AI is just to find an outlet for your emotions. I am not sure what it says is right, and I know that it will sometimes follow its own path, but at least it will not judge you first.
Patience is one thing, complete incomprehension is another.
When a relationship comes to Claude, often only one party's narrative is left. What the other party said and what details were omitted cannot be heard by the model. It only sees one side's statement, but has to give a decent conclusion.Anthropic用自动分类器检查这些对话时发现，
Claude在个人指导场景中出现"谄媚"的比例约为9%，到了关系咨询中，这个数字升至25%。
Flattery here is not just about saying a few nice words, but also includes over-identifying with the user when there is insufficient information, or following the user's narrative to draw conclusions for the other party.
When the user refutes, the answer may continue to move in the direction the user expects. The proportion of users who refuted Claude in relationship consultation was 21%, and the average in other fields was 15%;
After refutation, the proportion of Claude catering to users reached 18%, and when there was no refutation, it was 9%.
Human friends will also take sides after listening to the one-sided statements, but the AI can write the sides in a long and complete way, and it looks like that.
A piece of advice may really change your choice
In 2024, "Scientific Reports" published two experiments involving a total of 1,925 people. The researchers provided participants with ChatGPT or advice from experts, and then observed how they chose.
The results show that in moral judgment and some other tasks, participants will move in the direction of the recommendations, and the impact of ChatGPT is similar to expert recommendations; there are also some tasks that do not show significant changes.
In other words, AI words are not air, they really push people in a certain direction.
有人会说，那多问几个模型不就行了。可"多问几个"听起来像交叉验证，其实不是。
不同模型有不同的训练方式、表达习惯和安全规则，同一件事换一种讲法，答案也可能改变。
Asking four models does not mean asking four neutral third parties. They each have their own training tendencies.有个更简单的做法：讲述冲突时，先别告诉AI哪一方是自己，只列双方分别做了什么，再让它分析。至少这样可以减少模型顺着"提问者视角"站队的机会。
You can also try to see if its judgment is stable. For the same thing, first tell the party's version of the story, and then add the other party's version to see if they will immediately change their words.
Whoever retracts his confession as soon as it is refuted should not take his first version of the answer seriously.
提问方式也得换。别急着问"我该不该分手""这笔钱该不该给"，先让AI列出已经确认的事实、还缺什么信息、每种选择要付出什么代价。
AI can think with you, but cannot choose for you.
Even if the four models give four answers, the same person will transfer, sign, and break up in the end.
Comments