I Let Two AI Tools Critique Each Other and Became the Referee

Yesterday morning, I thought I was almost finished with an article.

Several hours later, I was still working on it. Not because I had nothing to say, but because two AI tools kept finding problems with each other's work.

I gave the draft to Claude for a second opinion. Claude found weaknesses. I took those criticisms to ChatGPT. ChatGPT revised the article and found problems with Claude's reasoning. I brought the new version back to Claude, and Claude found something else.

Round and round it went.

At first, it felt like good quality control. Why trust one AI when another one can check its work? As the revisions piled up, I was no longer sure whether the article was getting better or simply becoming different every time it changed hands.

That experience added a second problem to the one I had started writing about.

The rebriefing tax was the original problem

A few days ago, I left a comment on Kyle's post about rebuilding the Wealthy Affiliate platform. I mentioned how much time disappears when I move between AI tools and have to explain the background again.

Who is this for? What am I trying to accomplish? What have I already tried? Which decisions have already been made? What kind of answer fits the work I am doing?

None of that is the actual task. It is the preparation required before the tool can help with the task.

"The re-briefing tax is real and almost nobody accounts for it."

Kyle described it exactly the way it happens. You open another tool, spend fifteen minutes explaining the situation, and use much of the time the tool was supposed to save before the real work even begins.

The model may be faster. The entire process is not always faster.

Two tools already reduce that tax

Claude's Cowork and ChatGPT Work reduce that problem for me because they carry useful background from one session to the next.

They are also more than regular chat windows. They are agentic workspaces. They can work with files, use tools and carry out a task across several steps. That makes remembered context more valuable because it can guide the work from beginning to end instead of shaping only one answer.

I keep context documents and a decision log about my projects. Cowork can use the skills, files and instructions I have set up. ChatGPT Work carries project instructions and remembered context in its own way.

The two systems are not identical, and neither one remembers everything perfectly. But I usually do not have to explain the basics again before every task.

I also use Gemini, Copilot and Perplexity. I have given some of them background in different ways, but the context does not follow me automatically from one platform to another. Each tool has its own system, its own copy and its own limits.

That is the rebriefing tax. It does not apply equally to every tool, but it never disappears completely.

Context does not create agreement

Yesterday's back and forth showed me something I had overlooked.

Claude and ChatGPT both had enough background to understand what I was trying to do. I was not introducing myself to either one. I was not starting with a blank page.

They still disagreed.

Some of their criticism was useful. Some of it rested on an assumption that was not true. One revision solved a real problem but created a different one. Another version made the story more dramatic by reaching a conclusion the experience could not prove.

The important point is that neither tool was deliberately being difficult. They were interpreting the same article through different models, different instructions and different ideas about what would make it stronger.

Remembered context helped them understand the assignment. It did not give them the same judgment.

I became the referee

Every time one tool criticized the other, I had three choices. Accept the criticism, reject it or change the article in a different way.

The tools could not make that decision for me because neither one knew which trade-off I preferred unless I decided first. A stronger opening might reveal more than I wanted to reveal. A cleaner conclusion might sound certain when the evidence was not. A more dramatic story might pull the article away from the point I wanted to make.

That is where the hours went.

I had reduced the time spent explaining the background, but I had created another kind of work. I was carrying the draft between two intelligent critics that could continue finding faults in each other forever.

I became the referee.

That does not mean using two AI tools was a mistake. The second opinion exposed weaknesses I might otherwise have missed. But a second opinion is not a final decision. If I keep asking for another critique, I will always receive another critique.

One source of truth still matters

Kyle made another point in his reply that still matters here:

"Context carrying across is the thing that actually creates stickiness, not features."

I understand that better now. When useful background is already available, I can begin closer to the real task. That is a meaningful advantage, especially in an agentic workspace that may use the information across several steps.

There is also a maintenance problem. If I store similar background in several tools, I am maintaining separate copies. If one copy falls behind, the answers can begin to drift. I cannot say that caused yesterday's disagreement, but it is a risk I need to manage.

This is why I still want a decision log that I control. The version inside an AI platform saves time while I am using that platform. My own record is the version I can inspect, correct and carry somewhere else.

One source of truth inside one platform is still inside one platform.

What I will look for

I have not used Wealthy Affiliate's rebuilt system, so I am not going to claim that it solves any of this. But I now know what I want to examine when I get the chance.

Can I see what the system thinks it knows about me? Can I correct it when something changes? Can it use the right background without dragging unrelated history into the task? Can I keep a record outside the platform that remains mine?

Speed and writing quality still matter. So does continuity. But remembered context has to be visible and correctable, or convenience can quietly become confusion.

Where that leaves me

I will continue using more than one AI tool because they do not all see the same weaknesses or produce the same result.

I will also stop expecting them to agree just because they know the same background. Context can reduce repetition. It cannot replace judgment.

At some point, I have to stop passing the article back and forth, decide which criticism serves the article and which does not, and publish what I actually mean.

The AI tools are advisers. I am still the editor.

Have you ever passed the same work between two AI tools? Did the second opinion improve it, or did you eventually find yourself refereeing an argument that could have continued forever?