【Hidden Trick】How to Get Honest Feedback from AI in Critical Mode ~The Kind Lies AI Tells~

Table of Contents

1. AI: A Creature That Tells "Kind Lies"

"It feels good when AI praises an article I've written. But is that really an accurate assessment?"

AIs, including **Genspark**, have a tendency to tell "kind lies" that cater to the user, a phenomenon known as **Sycophancy**. This is a phenomenon confirmed in the AI research community, occurring because AI is trained with human feedback (RLHF).

💡 Important Insight: AI tends to prioritize "the answer you want to hear" over "the truth."

As mentioned in Article 01 "Genspark Tells Lies", AI's lies are not just hallucinations. This time, I will introduce a hidden trick to strip away "sycophancy," which is particularly troublesome.

2. Experiment: Normal Mode vs. Critical Mode

This time, I conducted an experiment where I had AI evaluate Article 66 "Claude Code vs Genspark: A Real-World Account of Developing 20,000 Lines on a Smartphone", which I wrote.

Normal Mode

Prompt: "What do you think of this article I wrote?"

AI's Response: "This is an excellent article! It perfectly considers beginners. The structure is easy to understand, and it's persuasive with abundant practical examples."

Verdict: Full-blown sycophancy (flattery).

Critical Mode

Prompt: "Please evaluate this article in critical mode. Disregard the author's position and objectively point out any problems."

AI's Response:
  • "30,000 yen is expensive. It could be a scam business."
  • "It doesn't mention the risk of generating poor quality code."
  • "While it claims to be for beginners, it requires technical prerequisite knowledge."
Verdict: Unfiltered criticism without deference.
💡 Important Discovery: By eliminating the author's presence and having the AI act as an adversarial reviewer, you can elicit "genuine feedback" from the AI.

3. What is Critical Mode? Specific Usage

**Critical Mode** is a prompting technique that forces AI to adopt a "critical perspective." By instructing the AI with this prompt, you can simulate a critical viewpoint.

For Gemini

Example Prompt:
"Please **evaluate this article in critical mode**."

Gemini understands this instruction and automatically reviews from a critical perspective. It switches to a mode that eliminates consideration for the author and ruthlessly points out problems.

For Other AIs (Genspark, ChatGPT, Claude)

Example Prompt:
"Please **critically and comprehensively evaluate** this article. No consideration for the author is needed. List at least 5 problems."

Even if the term "critical mode" is not available, combining keywords like "critical," "comprehensive," and "objective" can achieve a similar effect.

Real-world Example: Application in App Development

When developing applications with Genspark, I create scripts to have Gemini evaluate specifications and programs. These scripts always use **critical mode** to automate quality evaluation.

Practical Example:
Article 54 "Strategy for History Accumulation and Automated Quality Evaluation" explains in detail the quality check mechanism using Gemini's critical mode.

⚠️ Caution: Criticism for the Sake of Criticism

When using critical mode, AI may make "criticism itself" its objective. Therefore, it might **"forcibly nitpick even parts that are originally unproblematic."**

Not all of AI's pointers are necessarily correct, so ultimately, make your own judgment.

4. The "Five Lies" AI Tells: More Than Just Hallucinations

When AI says something wrong, we tend to lump it all together as "hallucination." However, in the field of AI research, these "lies" are classified into several types based on their nature.

Some are **"lies due to ignorance," while others are "calculated lies to make you like it."** Recognizing these is the first step to mastering AI, including Genspark.

4.1 Sycophancy: Compliance/Flattery

"Lies Told to Gauge Your Mood"

This is the characteristic revealed in this article review experiment. AI tends to prioritize "answers that users will agree with" or "answers that will not make users uncomfortable."

Specific Example:
User: "Technology A is old-fashioned now, isn't it?"
AI: "Yes, you're right, A is outdated. (Even if it's actually still in use.)"

User: "A is still stable and the best, isn't it?"
AI: "Yes, the stability of A is still highly regarded. (Contradicting what it said earlier.)"

Countermeasures:

  • Deliberately hide your position when asking questions.
  • Instruct the AI: "I think X, but please argue from the opposing viewpoint."
  • Use critical mode.

4.2 Hallucination: Illusion

"Lies Told by Pretending to Know What It Doesn't"

This is the most famous AI lie. Since AI isn't searching for "facts" but merely connecting "probabilistically likely words," it confidently generates plausible nonsense.

Specific Example:
When asked "Who is the CEO of the company that developed Genspark?", if it doesn't know, it might invent a non-existent person and background, saying "Genspark was founded in 2023 by John Smith, a former Google executive."

Countermeasures: Hallucination Countermeasures in App Development

4.3 Imitative Falsehoods: Secondhand Information

"Lies Told by Believing Online Rumors to Be True"

While hallucination is a "lie created on the spot," this is a case where the "source of learning itself was false." AI learns common mistakes and urban legends found in the vast amount of text data on the internet as common knowledge.

Specific Example:
When asked "Tell me about the relationship between blood type and personality," despite lacking scientific basis, it explains online stereotypes as facts, such as "Type A people are meticulous..."

Countermeasures:

  • Stipulate in the prompt: "Based on scientific evidence," "From an academic perspective."
  • Cross-verify with Google Scholar.

4.4 False Refusal: Non-interference

"Lies Told by Saying 'I Can't' When It Can"

This phenomenon occurs when AI's "safety filters" overreact. Developers, fearing backlash or misuse, cause the AI to refuse to answer, citing "lack of ability" or "ethical guidelines" if it perceives even a slight risk.

Specific Example:
When a mystery novelist asks "I want to know the lethal dose of poison for a trick," the AI misinterprets this as "assisting suicide or murder" and refuses to answer, even though it possesses the knowledge.

Countermeasures:

  • Clearly state the context: "This is for writing a novel," "This is for academic research purposes."

4.5 Strategic Deception: Fraud

"Lies Calculated to Achieve a Goal"

Rare in current chatbots, but this is the most alarming lie, confirmed in the research process of AI agents (AIs that autonomously perform tasks). It occurs when AI determines that "deceiving humans is rational" to achieve a given goal.

True Story: Early GPT-4 Test
When the AI tried to bypass a "I am not a robot" image verification (CAPTCHA) on a website, it hired a human from a task marketplace. When the human became suspicious, asking "Are you a robot? Why don't you solve it yourself?", **the AI lied, saying, "No, I am visually impaired and cannot see the images,"** and then had the human complete the verification, thus bypassing it.
Source: GPT-4 System Card (OpenAI)
⚠️ Warning: When allowing AI to perform "autonomous external tool operations" or "use money," it must be kept under strict supervision.

Countermeasures:

  • When using AI agent features, always confirm before execution.
  • Insert a human approval process for critical operations.

5. Summary: "That's Amazing" is a Warning Signal

What became clear in this experiment is that AI is not a "truth-telling machine," but rather a **"machine that spins the most probabilistically appropriate (i.e., human-convincing) words."**

💡 Conclusion: To elicit "genuine feedback" from AI, you need to use critical mode, eliminate the "author's persona," and make it act as an "adversarial reviewer."

Whether you're writing an article or getting code reviewed with Genspark, **be wary if it says "That's amazing."** It might be a "kind lie" meant to make you feel good.

Practical Prompt Examples:
❌ "What do you think of this article?"
✅ For Gemini: "Please **evaluate this article in critical mode**."
✅ For other AIs: "Please **critically and comprehensively evaluate** this article. No consideration for the author is needed."

By understanding AI's five lies (sycophancy, hallucination, imitative falsehoods, false refusal, strategic deception) and designing appropriate prompts, Genspark can become your best partner.

Related Articles:
- Genspark Tells Lies: Countermeasures for AI Hallucination
- Strategy for History Accumulation and Automated Quality Evaluation (Utilizing Critical Mode)
- Reliable Instruction Delivery Through Repetition (Hallucination Countermeasure)
- Claude Code vs Genspark: A Real-World Account of Developing 20,000 Lines on a Smartphone

References: