Overview
Two weeks ago, I joined a good company and I’m now working on a government AX-related project. Coming from an embedded development background, the pace feels fast — I have to produce results within two months — and I’m struggling to adapt.
Even so, there are things that slow me down. One of them is “verification.”
On my way home from work today, I told a fellow developer: “This is verification hell…” Verifying documents that AI organized, verifying code that AI wrote. AI has clearly sped up implementation, but the amount of content that needs verification piles up just as fast.
How do I verify this massive amount of content?
Table of contents
Open Table of contents
AI for verification?
My first thought was whether I should build an AI specialized in verification. So I looked for related research.
LLM-as-a-Judge
I think I first heard this term while working on a RAG example, when verifying outputs. I used it to quantitatively check how reliable RAG’s answers were to test questions. This paradigm (as it’s called) is apparently used to evaluate and compare the outputs of various AI systems. There’s a single-agent verification approach, and it’s also evolving toward multi-agent debate-based evaluation.
Critic models
OpenAI apparently has CriticGPT. The paper seems to have been accepted in 2024, and it covers exactly what I’ve been worried about lately. Its starting point is that as models get smarter, we’re reaching a point where humans can no longer properly judge the accuracy of the model’s output. Yesterday at a forum, I agreed with someone saying that current model performance is far superior to models from two years ago — but it turns out people were already worrying about this issue back then, which makes me think verification matters even more now than it did then.
In the end, CriticGPT and a prompt-improved ChatGPT detected far more bugs than humans did. Also, a human + CriticGPT collaboration team reduced the hallucination rate more than the model alone. CriticGPT was reinforcement-learned specifically for code review, but it also applied critical thinking outside of code review — lowering the grade on answers humans thought were perfect, and this turned out to catch actual existing errors.
Looking at this, it seems like you could just hand off both the work and its verification to AI. But of course there are limitations. CriticGPT also had hallucination problems, and at a much higher rate than humans. Also, the harder the problem, the less meaningfully helpful CriticGPT was.
In the end, I find myself thinking: if I build an AI to verify AI, doesn’t that answer also need verification? The LLM hallucination problem will clearly persist, and probability will eventually catch me off guard.
Verifying isn’t the end of it
Let’s assume, hypothetically, that AI never makes mistakes and can verify all its own output perfectly. Would that mean humans have nothing left to do? These days, I’m even more convinced the answer is no.
As AI advances, humans end up surrounded by more code and more documents. The problem is that humans, not AI, have to take responsibility for the code it outputs. Another word for “responsibility” here could be “understanding” or “verification.” In other words, even if AI verifies something, humans still need to understand and verify the results themselves in order to explain them to collaborating colleagues or to clients who requested the service, and to take responsibility for them.
For this reason, I’ve concluded that what I need is not AI for verification, but studying for verification.
Learning in the age of AI
People say AI is growing faster than humans can learn. That’s true to some extent. OpenAI released Astra, and today Claude released Opus 5.5. But that doesn’t mean human growth has become unnecessary.
I believe humans need a foundation of knowledge solid enough to verify the output of these increasingly capable models. Domain knowledge, code review, data analysis — these are the kinds of knowledge that become even more necessary amid the endless pile of documents and code. If you lack these concepts while working on a project, you simply can’t tell whether AI’s answers are right or wrong. If you proceed assuming it’s probably right, you’ll eventually walk right up to the edge of a pit and hear the AI say, “I checked again, and there’s a pit ahead. I was wrong!” To prevent that kind of productivity loss, you shouldn’t neglect studying your own field, and I believe that skill for verification will never go out of date.
So, in conclusion…
AI has gotten good enough to replace a huge number of humans. To keep up with its improving performance, quickly experimenting with AI and building things might look like the faster path. But I think studying to build the skill to verify and explain AI’s output — though it looks slower — is the surer way forward.