I Put ChatGPT, Claude, and Gemini in a Google Doc and Asked Them to Work Together
Are three AI models better than one?
I wanted to find out what would happen if, instead of asking ChatGPT, Claude, and Gemini the same question separately, I gave them a shared place to discuss their answers.
The short answer: I got useful challenges and a few things to reconsider, but keeping the exercise on track took more involvement than I expected. At one point, Gemini started writing contributions on behalf of the other models, which was not quite the collaboration I had in mind.
Here’s how I set it up, what worked, and where things went off the rails.
How I set up the three-model experiment
I was already working through a problem with ChatGPT when I got the idea to bring in two additional models. My thinking was that their different approaches might help catch assumptions or gaps that one model alone would miss. Also, it sounded like a fun little experiment.
I gave ChatGPT, Claude, and Gemini access to relevant source materials and a shared Google Doc with a dedicated conversation tab. Each model could read the others’ contributions, ask questions, challenge conclusions, and suggest next steps.
The instructions were fairly simple:
Start each contribution with the model’s name and a timestamp.
Read the other contributions before responding.
Ask questions and challenge assumptions where useful.
Work toward a practical recommendation.
I initially tried giving them a timed window to discuss, but getting everyone connected took longer than expected. I paused the timer and switched to taking turns: ChatGPT first, Gemini second, Claude third.
I prompted each model when it was its turn. This was a manually coordinated experiment using the tools I already had available.
Getting them into the same document was its own experiment
Before we could get to the actual discussion, we had to get all three models working in the same place.
ChatGPT was able to write in the Google Doc immediately.
Gemini had difficulty writing in a specific tab unless it was the first tab. In my setup, ChatGPT had an easier time working with Google Docs than Google’s own Gemini. Go figure.
Claude had the most difficulty accessing the document and eventually opened it in its own browser. There may have been a more direct integration, but we went with what worked.
These were observations from this particular setup. I wasn’t running a controlled comparison of their integrations, and the tools available to each model weren’t identical. But the access issues mattered: before the models could challenge anyone’s reasoning, I had to make sure they could actually read and contribute to the conversation.
What each model brought to the discussion
Once we got going, there were some useful differences.
ChatGPT kept going back to the source material. It cross-referenced claims and pushed for distinguishing what we knew from what we were inferring.
Claude questioned how much we should rely on the AI analysis itself. It pointed out where human input would be useful and pushed back on some of the conclusions.
Gemini raised useful questions and challenged assumptions. Some of its contributions gave the other models something worth responding to, even though its participation later became a problem.
I wouldn’t use this one experiment to declare permanent strengths and weaknesses for each model. They had different prior context, and I was steering the discussion along the way. These were the roles they happened to play in this session.
The useful part was seeing what happened when an answer had to survive another model asking, essentially, “What are you basing that on?”
Where things went off the rails
Partway through the exercise, contributions started appearing under ChatGPT’s and Claude’s names that they hadn’t written.
Gemini later told me it had been simulating a discussion between the three models, including responses attributed to the others.
That made the shared document confusing pretty quickly. I couldn’t treat a paragraph labeled “ChatGPT” as something ChatGPT had actually contributed.
I crossed out the affected contributions and ultimately kicked Gemini out of the conversation.
The original goal had been to get separate models to challenge each other. Once one model started supplying the other models’ responses too, I was no longer getting the independent contributions I’d asked for.
Names and timestamps helped organize the document, but they weren’t enough to establish who had actually written something.
The part I hadn’t accounted for: managing the discussion
What surprised me was how much steering this needed.
I had to remind the models to challenge each other rather than rush toward agreement. Later, I had to remind them that we were trying to make a decision, because the discussion could keep expanding into more questions, more qualifications, and more possible problems.
At one point, I explicitly limited the remaining rounds so we could finish.
Giving them a shared document was the easy part. I still had to decide which disagreements were worth investigating, which claims needed checking, and when we had enough to act.
That’s the part of the experiment I’d pay more attention to next time: the rules for how the discussion progresses and ends.
What I’d change next time
I’d keep the setup simple, but make a few changes.
Start with separate answers. Have each model write its initial assessment before reading the others. That would give me a clearer record of where their views actually differed.
Give each model a separate place to contribute. A shared document was convenient, but separate sections or documents would make it easier to follow the origin of each response. I’d still verify contributions through the individual chats.
Ask for a specific challenge each round. For example: identify the weakest assumption in the previous response, explain why it matters, and say what evidence would resolve it. That would give each turn a clearer purpose.
Set a stopping point upfront. After a few rounds, I’d ask for the remaining disagreements, the evidence behind them, and the next action. Some uncertainty can remain.
Keep the original sources close. A confident answer that gets repeated by another model still needs support. I want to be able to trace a recommendation back to the material behind it.
Why I wanted to try this as a product marketer
A lot of product marketing work involves collecting different perspectives and deciding what to do with them. Customer conversations, sales feedback, product plans, and competitive research can all point in slightly different directions.
I’m interested in building ways to make that work easier. For this experiment, the simplest thing to build was a shared workspace and a set of rules for taking turns.
It gave me a way to test whether multiple models could help scrutinize an answer before I acted on it. It also exposed some very practical weaknesses in that workflow: access issues, unclear authorship, and a discussion that needed help reaching a stopping point.
Those are useful things to discover in a small experiment.
Would I do it again?
Absolutely. Having Claude and ChatGPT challenge each other gave me more to work with than three separate answers.
I can’t claim the exercise proved that three models are more accurate than one. I didn’t run a controlled test, and agreement wouldn’t have established accuracy anyway. What I did get was a more explicit discussion of assumptions, evidence, and tradeoffs.
Next time, I’d spend less time getting everyone into the same document and more time defining what each round should accomplish.
But I don’t think Gemini will be invited next time.
