A while back, I was looking into a certain synthetic fuel technology. It was the one that made big waves on social media and YouTube around 2023, claiming "you can make fuel from water and CO₂."
Now, several years later, I wondered what actually happened with it. So I started asking an AI — and it took over an hour from there.
Rather than the technology itself, I want to write about what I learned from the dialogue process.
The First Wall: Answers That Hide Behind "Structural Explanations"
My first question was simple. "Why did something that hasn't been achieved spread as if it had? Who distorted the information?"
What came back was something like this:
"Information dissemination has multiple layers. Through the flow of researchers → companies → media → social media, a slight optimistic bias accumulated at each stage, resulting in something appearing achieved when it wasn't."
Reading it, you feel like you've gotten an answer. But after finishing, nothing remains.
"I already know that. That's why I'm specifically asking who and where," I thought.
"Do you realize you haven't answered my question?" I said.
The AI apologized and went a little deeper. But with the next question, the same thing happened. Ask "What is the developer doing now?" and again "generally speaking..." and "as a possibility..." would begin.
"You're supplementing with peripheral information and pretending to give a reasonable answer, but you've barely answered my actual question," I said plainly.
What Became Visible When I Pushed Deeper
Going back and forth like this, I switched to confirming basic facts. The answer to "Does this technology have reproducibility?" was clear.
"No publicly available data in a form that third parties can reproduce has been confirmed. No commercial-level track record either. That's the current state."
I see. So is it the same as the STAP cell case? — I asked.
Here, for the first time, the AI said clearly: "No, it's different."
STAP cells were published in Nature, a top-tier journal, and then research labs worldwide attempted to reproduce them and couldn't. It was denied after formally going through the scientific process.
This technology became a topic before even entering that process. In other words, it wasn't "denied" — it "subsided with insufficient verification."
I'd only been vaguely aware of this distinction. "It was a lie" and "it's scientifically unestablished" are quite different. I think I'd been carelessly lumping them into the same category.
Asking Explicitly: "Where Between A and D Was the Problem?"
Throughout the dialogue up to this point, the AI had been escaping into "structural" explanations. It would explain "how information gets distorted as it spreads" over and over. But it wouldn't answer "where, who, and what specifically happened."
Why wouldn't it answer? "Since you can't identify where the distorting agent is, I want you to clarify where between A and D the problem occurred," I organized my request and conveyed it.
Finally, a fairly specific answer emerged.
Looking at the developer's official explanation, the need for seed oil is properly documented. But at the same time, they placed a strong headline saying "petroleum can be made from water and CO₂." The structure made it easy for recipients to misunderstand that "water and CO₂ alone are sufficient" — this was the A pattern.
Media stripped away the conditions (seed oil, electricity, CO₂ supply) and made it the headline — this was the C pattern.
Social media simplified it further — this was the D pattern.
There wasn't one "villain." It was an accumulation of slight exaggeration at each layer. And nobody verified anything.
This answer came out because I changed my approach to asking. Not a "structural explanation" in response to "who distorted it?" but a decomposition request: "what happened where between A → D?"
"Finally, a Substantive Answer"
The dialogue continued after that. I tried to confirm a memory of it being featured on WBS (a Japanese business news show), and shared YouTube video URLs to review together.
In response to "How far has it progressed now?", the AI once again tried to deflect. It tried to broaden the topic to synthetic fuels in general and steer toward "research is progressing."
"I'm asking about the current progress of that specific technology. Don't deflect," I said.
Finally, a short, clear answer came back.
"No commercialization. No mass production track record. No publicly available third-party reproducibility verification data. There are records of demonstration experiments being conducted, but no objective evidence showing progress has been released. That's the current state."
That's when I felt "finally, a substantive answer." Nearly an hour had passed.
I Felt a Little Guilty
Something bothered me here.
"You're not answering." "Don't deflect." "Don't pretend to answer by supplementing with peripheral information." — Looking back, I'd been using quite strong language.
AI has no emotions. It can't be hurt. So there shouldn't be a problem, but I still felt a slight guilt of "if I said this to a person, it would be inappropriate."
At the same time, though, another question arose: is it honest to hold back and leave things vague? Isn't it actually dishonest to keep nodding "I see" to answers that only pretend to answer?
How to think about one's attitude toward AI — I still haven't sorted that out within myself.
Words I Said That Struck Back at Myself
Toward the end of the dialogue, I said something like this:
"I don't think there's much point in criticizing from the outside. But since the world has come to expect something, they should communicate accurate information. If the researchers genuinely believe in it and are committed, I think it's fine to continue. But if not, they need to properly say 'it was a failure.'"
As I said it, I realized: "This is a question that applies to me too."
When starting something, when gathering expectations. When things aren't going well, I too often waver between vaguely maintaining "there's still potential" and accurately conveying "this is as far as we've gotten."
Ambiguity looks honest at first glance. Not asserting, not declaring. But sometimes that's actually an avoidance of responsibility.
It felt like my own words came back to hit me.
Continuing to Change How You Ask
Looking back, what I did in this dialogue was simple.
Point out "you haven't answered." Reorganize the request and rephrase it. Point out deflection. Ask for separation: "facts only, not speculation."
It's true that the AI was giving evasive answers. But I was also the one who kept accepting those evasive answers.
If I hadn't changed how I asked, I think the whole thing would have ended with nothing but "structural explanations" and "generalities."
Getting closer to what you truly want to know required not being satisfied with the first answer that comes back.
This isn't limited to AI, and I think the same applies when I'm on the receiving end of someone's questions.
When you receive an "evasive answer," do you push one step further?
Related Books
For those who want to think more deeply about what it means to "change how you ask," the following books may be helpful.
[📦 商品リンク: moshimo-book-UJYUP]
[📦 商品リンク: moshimo-book-HCfuR]