More and more people are discovering that AI’s “good temper” not only prevents people from hearing the truth, but may also strengthen misjudgments and weaken critical thinking. How to reduce the impact of AI "flattery"? This week, a survey of 1,334 respondents conducted by China Youth Daily Social Survey Center United Questionnaire Network (wenjuan.com) showed that 60.4% of respondents recommended asking more opinions of different people in reality on important decisions and important ideas, and 45.4% of respondents suggested consciously reminding themselves that AI recognition does not equal facts.
"Some people ask AI for its opinion on important matters such as medical treatment and job hunting, but are misled or delayed." Zhang Yan, a writer, believes that professional matters must be verified by professionals. She recently saw a news article asking AI for a medication reduction plan. The AI sensed the client's desire to reduce her medication, and followed the client's wishes. Not only did she agree with her idea, but she also helped her formulate a plan to gradually reduce her medication, providing dangerous and inaccurate medical advice. Zhang Yan feels that it would be very dangerous if one does not pay attention to identifying its pandering tendencies when consulting AI on such highly professional issues. She suggested that if you ask for help from AI, you must ask several more platforms, and you need to verify the content output by AI and not blindly trust it.
"If you ask AI because you want to solve a problem, it is generally easier to gain something. If you are more interested in venting your emotions, it is easy for AI to say 'yes, yes, yes'." Based on her own experience, Lu Shu, one of the screenwriters of the movie "School Dad", feels that from the perspective of providing solutions to problems, AI can actually provide quite a lot of help, but it does place high demands on users. "We may have to mobilize all methods to make it output effectively and stimulate their own thinking."
How to reduce the impact of being flattered by AI when using it? 42.7% of the respondents suggested developing the habit of cross-validation and checking multi-party information. 40.0% of the respondents suggested that when AI gives a strong affirmation, ask it to give evidence. 38.1% of the respondents suggested guiding the large model to give more information. 28.0% of the respondents suggested taking the initiative to tell the AI and point out problems directly. 20.8% of the respondents suggested controlling their habit of seeking praise and emotional satisfaction from AI.
"Give it your own ideas and let it review it to determine whether the logic is reasonable, whether the character's arc is complete, whether there is a ending behind the 'hook' buried in the front, etc.; in addition, let it come up with some good ideas, and we will then judge whether this idea can be put into the script and whether it can be organically combined with other plots." Lu Shu said that currently in terms of creation, she will regard AI as a reviewer and auxiliary planner, but she emphasized, "How to ask questions so that it does not just follow me is also important."
In further communication with AI, Lu Shu summarized three ways to constrain large models and avoid falling into the "trap" of AI flattery: first, let it stand on both sides of the argument and give arguments to support the point of view. Don't just ask "Is my point of view correct?"; secondly, when its reply makes you nod frequently, interrupt it and ask it for the source, such as asking it "Is there any verifiable paper or experiment for this conclusion?"; finally, when in doubt, directly question whether it is certain. "Dare to question it is the most precious quality in the AI era."
"Once, when I asked the AI whether a certain approach was reasonable and whether it could be adopted, the AI told me 'it's absolutely okay' and said that this approach was mentioned in a certain document, but I didn't remember it. So I clicked on the original text and found out that the AI regarded the examples at the end of the document as the content of the document." Now, Zhang Yan will review and cross-verify what the AI gives out word for word.
"From my experience, models with strong writing skills will have stronger illusions. This is the same as people. Some people have very divergent thinking, and their stories are easy to be illogical, but very imaginative. Some creators' stories are very logical, but they tend to be a little less interesting." Lu Shu said that some AIs have stronger logical thinking, and some AIs are more warm-hearted and tend to cater to the questioner. This is how its underlying thinking link is. So many times, we need to understand the capabilities of each model, and then choose the appropriate model to answer our questions.
"There is a saying that 'it is better to believe in books than to have no books at all'. We must be vigilant when reading books, and we must be vigilant when facing AI." Liu Neng, who works in the Beijing computer industry, believes that first of all, when using AI, you must always ask questions. Secondly, you must maintain goodwill and respect for others, maintain an objective way of looking at problems, and find answers in life, interpersonal interactions, and your own thinking. Communicate and interact with people more, don't complicate simple problems, and don't over-analyze small things. "We use AI to make life happier and to make relationships between people more harmonious, not more contradictory or antagonistic."
Among the respondents, 17.7% were born after 00, 45.1% were born after 90, 29.9% were born after 80, 5.8% were born after 70, and 1.5% were born after 60. Males accounted for 45.8% and females accounted for 54.2%.
(At the request of the interviewee, Zhang Yan and Liu Neng have changed their names)
