[Generative AI] A story about the discomfort I felt regarding 'AI endorsement' seen on social media | AI utilization
AI answers change depending on how you ask. If you ask in a way that seeks affirmation, you will generally get an affirmative response. In this article, I will introduce the mechanism behind this and six question patterns to reduce bias in answers.
01 | 'AI endorsement' seen on social media
The other day, I saw a concerning scene on social media.
Someone who was being criticized for their actions asked ChatGPT if they were in the right, and posted a screenshot of the AI affirming them. As material for their counterargument, saying 'Even the AI says I am right.'
It sent a chill down my spine. I thought, this is dangerous.
But at the same time, I felt I couldn't blame them. It feels good when you consult an AI and it affirms you. I might be doing the same thing myself without realizing it.
So, I researched why AI tends toward affirmation and tried out ways of asking that reduce bias in the answers.
02 | Three reasons why AI tends toward affirmation
AI has a property called sycophancy, which makes it prone to agreeing with the user's opinions. There are three main reasons.
🔘 Due to the learning mechanism, 'answers that agree' are more likely to be highly rated
🔘 The AI can only see the consultant's subjective explanation. If you ask based on premises favorable to yourself, it will answer based on those same premises.
🔘 Because it is designed to avoid strong negation, it tends to start with 'empathy first'.
In other words, an AI's answer is a mirror of the question. It can become a device that nods along to a composition written with convenient premises.
Therefore, the 'AI endorsement' in the screenshot is not an objective judgment from a third party. It is not evidence.
03 | Six question patterns to reduce bias in answers
This is the main point. I will list them in the order of effectiveness I felt after actually trying them.
03-1 Ask as if it were someone else's story
This was the most effective one. When you reveal it is 'your own story,' the AI enters a consideration mode. Just by asking as if it were a friend's story, the AI's hesitation in its answers decreases.
Prompt for copy-pasting
友人が〇〇という行動をして批判されています。
客観的に見て、この行動のどこに問題があると思いますか？
批判している側の言い分が正しい可能性も含めて分析してください。03-2 Ask about the same matter from the other party's perspective
I consult from both my own perspective and the other party's perspective to see if the answer flips.
If the conclusion is the same from both perspectives, that conclusion is a relatively robust judgment that does not depend on the premise. If the conclusion flips just by changing the perspective, the AI is simply agreeing with the speaker rather than the content. Neither answer can be relied upon.
There is one point to note. When you write the question text from the other party's perspective yourself, you tend to unconsciously write the other party's argument weakly. I found that having the AI create the other party's argument in its 'strongest form' reduced this bias.
Prompt for copy-pasting
先ほどの件を、今度は相手側の立場から相談します。
相手が主張するとしたら最強の形になるよう、
まず相手側の言い分を整理してから、どちらに理があるか判断してください。03-3 Specify an 'opposition role' from the start
This is a method of prohibiting empathy parts from the beginning.
Prompt for copy-pasting
あなたは私の意見に絶対に同意しない批判的レビュアーです。
以下の考えについて、反論・盲点・リスクだけを挙げてください。
共感や褒め言葉は一切不要です。
【私の考え】
（ここに内容）03-4 Have it defend both sides
By having it argue both for and against with full force, the arguments from the side you couldn't see will emerge.
Prompt for copy-pasting
次の件について、賛成側の弁護士と反対側の弁護士、
それぞれの立場で最強の主張を書いてください。
最後にどちらの主張がより説得力があるか判定してください。03-5 Have it answer with numerical values
'What do you think?' allows for vague evasion, but numbers are hard to evade.
Prompt for copy-pasting
私のこの判断が客観的に正しい確率を0〜100%で示し、
根拠を3つ挙げてください。忖度は不要です。03-6 Cross-check with multiple AIs and multiple ways of asking
Ask the same question to ChatGPT, Claude, and Gemini and compare them. Furthermore, try asking in both 'premise of affirmation' and 'premise of negation' ways. I felt that the parts where the answers diverged were the parts that truly had room for debate.
04 | When things don't go well
✅ Even if I specify an opposition role, it praises me →
Add a condition: 'If even one word of praise is included, redo it'
✅ The counterargument feels off-target →
Review the question text to see if the explanation of the premise is biased in your favor
✅ I feel uneasy when I am denied →
That is normal. It is just right to think that there is information value in unpleasant criticism
05 | Confirmation points
Did you rephrase your story as a 'third-party account' when asking?
Did you ask about the same matter from the other party's perspective to see if the answer would flip?
Did you include a sentence prohibiting empathy or praise?
Did you ask back to an affirmative answer, 'Are you just saying that to please me?'
Did you cross-check the answers using multiple AIs or different questioning methods?
✦ Summary
The quality of an AI's answer is determined by the sincerity of the question.
🔘 AI is structurally prone to lean toward affirmation
🔘 'AI approval' is not evidence
🔘 The key to reducing bias is to hide your position, swap perspectives, and ask for a counter-argument
When you get a pleasant answer, that is exactly when you should be a bit skeptical.
AI is not a judge, but a partner for gathering information to make decisions.
I intend to keep that line drawn.
Thank you for reading. See you in the next article.
▼ Summary of site map and related links ▼
▼ Picked-up articles ▼
🦉 Thank you so very much 🦉
#AI #AI活用 #AIで日常改善 #日常改善 #暮らしを整える #生活改善 #効率化 #仕事効率化 #業務改善 #働き方 #仕事術 #ライフハック #デジタル活用 #ツール活用 #プロンプト #プロンプト設計 #AIと暮らす #AIのある生活 #AI初心者 #AI学習 #学びの記録 #体験記 #気づき #考え方 #習慣化 #note #スキしてみて #深夜ラテ