Use AI as a Mirror, Not an Answer Machine (Research That Resists Hallucination)

AI doesn't give you the answer — it forces you to ask the better question.

A scenario where asking AI just wasted time

Imagine this. A founder asks ChatGPT about a new SaaS idea: "What collaboration tool is most-used in the Korean developer community?" The answer comes back in 5 seconds: "Slack and Jira are dominant in Korea too, with combined market share of about 82%."

It was a clean number. 82%. The founder spent 3 days building a competitive analysis on that figure. 5 more hours to write the conclusion. A week later, he asked a developer: "Do 82% of Korean dev teams really use Slack?" The dev laughed. "Our team uses Notion and KakaoTalk. Where did that 82% come from?"

That's the moment you realize what AI hallucination is. AI doesn't lie. It just pretends to know while not actually knowing. And it does it in that confident voice.

The core of this post is this: instead of asking for the answer, force AI to ask the better question.

Don't ask for the answer — refine the question

The research trap is simple. We go to AI looking for an answer. But AI isn't an answer machine — it's a mirror. A mirror returns the quality of the question you brought.

"How is the Korean SaaS market doing?" is the worst question. The scope is too broad. AI spews all kinds of information randomly. Some of it is roughly correct, some completely made up. You can't tell which is which.

Ask like this instead.

My SaaS targets marketing team leads.
Competitors: HubSpot, Mailchimp, ContentMark.

Compare these three services' pricing.
- Basis: Korean market
- Data: 2024–2025

Specific, with defined comparison targets and bounded scope. Asked this way, AI becomes more careful. And it's more willing to admit "I can't be sure about post-2025 data."

The core is this. AI succeeds at your research only as much as you expect. Prompt engineering sounds fancy, but it comes down to three things:

1. Bring context: "I'm a B2B SaaS startup founder. Main customers are SMB marketing teams."
2. Be precise about the goal: "I want to know, for 3 competitors: (1) base plan pricing, (2) main features, (3) target customer segment."
3. State the constraints: "Only post-2024 Korean market disclosure data. Exclude guesses or older versions."

Claude vs ChatGPT vs Perplexity: role division

Which tool to use? Simple answer. All of them. But for different purposes.

Start with Perplexity. This is research's entrance. Pulls live web info and cites sources. For exploratory questions like "what tools are popular with Korean marketers?", Perplexity is fastest.

Iterate with ChatGPT. Take the base info from Perplexity and ask ChatGPT something creative like "given this market, what's a differentiation strategy for entry?" ChatGPT is excellent at developing your ideas.

Wrap up with Claude. Dump everything you've collected — info, ideas — into Claude in one go. "Here's the market info, competitor analysis, and our hypothesis. What angle did we miss synthesizing all of this?" Claude handles long context best and is the most precise at pointing out where you've overestimated.

"Just use one tool, right?" is a forbidden thought. Use one tool and that tool's blind spot becomes yours.

This role split can finish in 30 minutes. Perplexity 10 min, ChatGPT 10 min, Claude 10 min.

30-minute market research workflow

Step 1. Define the topic (2 min)

Write the research topic in one sentence. Example: "Korean B2B marketing automation tools — key players and pricing."

Step 2. Map terrain with Perplexity (8 min)

I'm researching automation tools targeting B2B marketing teams in Korea.

(1) Top 5 competitors
(2) Base plan + main features for each
(3) Post-2024 evaluation in the Korean market

Cite sources for each item.

Live web search gets you the latest with sources. If "I don't know" comes back at this step, you already know what to do for the remaining 30 minutes.

Step 3. Develop ideas with ChatGPT (8 min)

I got this market info. (paste)

My service:
- Target: Korean SMB marketing teams
- Differentiator: integration with Korean payments (Inicis, Nice)

Give me 3 perspectives that could help us survive this market.

Step 4. Catch critical errors with Claude (8 min)

I'm looking at the analysis from Perplexity and ChatGPT so far. (paste)

- What's the most important question we're missing?
- Is there a dangerous assumption in our hypothesis?

Claude will point out things like "3 of those 5 competitors actually went bust" or "Korean market entry barriers are much higher than you think."

Step 5. Decide on the next validation (4 min)

After this workflow, you already know:

• Major players in the market
• Pricing positioning
• Which of your hypotheses are solid vs risky
• Who you need to interview next

Filtering hallucinations: spotting AI lies

AI doesn't say "I don't know." It plausibly makes things up.

Another sketched scenario: someone asks Claude "what's the market share of the top Korean payment gateway?" Answer: "NHN Payco ~65%, Inicis ~28%." The number sounded credible. They checked. Wrong. There's barely any external disclosure data on this. AI presented nonexistent data as if it existed.

Academic research shows 25–35% of AI-generated citations and statistics are entirely false or contain errors. Legal is even worse — LLM accuracy on legal queries shows 69–88% error rates.

5 ways to filter hallucinations

1. Always demand the source. If AI gives a statistic, immediately ask "source?" If it says "according to the 2023 Korean Payments Association report," try to find that report yourself. 99% of the time it doesn't exist.
2. Suspect too-clean numbers. "82%", "exactly 5", "about $50,000" — round numbers are mostly hallucinations. Real-world statistics are messier.
3. Watch how often "I'm not sure" appears. Claude admits "I don't have precise information" more often than ChatGPT. That's why it's more reliable.
4. Cross-check across AIs. Ask the same question of Claude, ChatGPT, and Perplexity. If all three agree, higher confidence. If one disagrees, that one likely hallucinated.
5. Be extra cautious with Korean information. AI has far less Korean data than English. Korean startup, payments, and marketing tool info is more often wrong.

5 ready-to-use market research prompts

1. Competitor pricing comparison

My SaaS target: [target]
Competitors: [3 competitors]

Compare as a table:
(1) Base plan pricing (KRW)
(2) 5 main features included
(3) Primary customer segment
(4) Korean language support & Korean payment methods

Constraints:
- Post-2024 disclosure data only
- Mark unclear info as "needs verification"

2. Market size and trends

I want to understand the [market segment] market.

(1) Korean market size 2024–2025
(2) YoY growth rate
(3) Main growth drivers
(4) Key trends for next 6 months

Cite sources for each.

3. Finding user pain points

I'm building a service for [target customer].
They currently use [existing solution].

What are the top 3 frustrations with the existing tool?

(Based on actual user communities or reviews — not your guess.)

4. Validating Korean-specific strategy

Why do overseas SaaS products fail in the Korean B2B market?

Focus on:
- Payment systems
- Terms of service differences
- UI/UX localization failure cases
- Customer support

5. Stress-test your hypothesis

My hypothesis: [your hypothesis]

How could this be wrong?

(1) What has to be true for this hypothesis to hold?
(2) Are those assumptions actually true?
(3) If this is wrong, do we waste 6 months?

Wrapping up

AI is not "the partner that gives you answers" — it's "the mirror that makes you ask better questions." In 30 minutes you can map the market and find the holes in your hypothesis. But something is still missing. Statistics and guesses aren't enough — you need actual human voices.

Next post: take this map into the field, and learn to ask one person deeply for an hour before you ever ask 100.

References

1. ChatGPT vs Perplexity vs Claude: A Complete Guide for Marketing Leaders in 2026 — Genesys Growth
2. AI Hallucination Statistics & Research Report 2026 — SuprMind
3. AI Citation Hallucination — Citely
4. Korean AI SaaS Market — KoreaDeep
5. Prompt Engineering Guide — Google Cloud

Previous: 5 Questions Before Building (A Sustainability Check)
Next: One Deep Interview Beats 100 Surveys (Mom Test)

About the scenarios in this post
The "82% case" and "payment gateway case" are scenarios fabricated by Fail School to illustrate AI hallucination patterns. The hallucination rate statistics (25–35% error, 69–88% on legal queries) are based on real research — UNC Charlotte's AI Hallucinated Citations Research Guide and Stanford RegLab Legal LLM Hallucination Study (2025).

Minchul Kim, CEO of Freeive, Fail School