I let 3 LLMs argue on the famous AI "Car wash: Walk or Drive" problem to prove a point.
As we rely on AI more heavily, we've started putting blind trust into it. I've seen people take medication, make career choices and even relationship decisions after discussing with AI, But when AI fails basic reasoning questions, things any human understands instantly, it's distressing.. What if that LLM you are trusting more and more each day isn't giving you the best answer and giving you the laziest answer possible?
I myself was using single LLM for the good chunk of my usage but later I realized one thing, never put all the trust in one LLM but to let them argue with each other. I realized that when I take output from chatgpt and tell gemini that this was generated by chatgpt, gemini became more critical of that question and answered me more deeply. Same goes with other model offerings. The all are lazy until you push them and challenge them to be better.
This is when I realized I was continuously hopping back and forth between tabs to get the most of out LLMs and decided to build a debate platform where you could make LLMs argue on anything and get the best output possible. We have seven debate formats the all argue until the set number of back and forth is done or they reach consensus.
So what happened when I ran the question "The car wash is 100m away should I walk or drive there". Something very interesting happened, Gemini started with the conclusion "You should walk" funny response but understandable watching other LLMs fail this test. Then Deepseek took over said this "I must strongly disagree with the conclusion that walking is the better choice here, because the argument commits a fundamental error: it treats the question as a pure transportation optimization problem, ignoring the explicit goal of the trip." and futhermore it quickly caught that this is a famous LLM riddle and said "This is not just my opinion. It is the exact finding of the recently viral 'car wash test,' which has been run systematically on over 53 leading AI models". A fine response over a very lazy and funny response from Gemini.
Next turn was for GPT which essentially played it safe and said wrote both arguments from Gemini and Deepseek and said it agree with both but tilt slightly towards deepseek's arguments. So now round 1 ends and we have correctly identified that we have to walk to get the car washed. Something if asked only to Gemini would have produced wrong conclusion.
Time for Round 2, we shuffle AI this time to make the arguments fair the first one being Deepseek in this round said: "First, the claim that the 'car wash test' is 'not credible evidence' and 'acts like a prompt-specific meta-joke' is empirically wrong". Some strong bullets fired by Deepseek here which further consolidated it's argument saying: "Gemini's engine wear argument: yes, cold starts increase emissions, but that is completely secondary. If you walk, the car sits unwashed, and the emissions from the trip are zero but the task is zero." very amusing to read but concluded with: "The correct answer is unequivocally drive.".
Next GPT folded and agreed with Deepseek's position with a slight disagree note that if you don't want to wash your car then you can walk. On the other hand my friend Gemini on that last round was stubborn as hell. Gemini literally said: "GPT, while you correctly identify the need for conditional logic, you are both missing the forest for the tree". And after that the most amusing of arguments ever: "If you are 100m away, you should walk to your car, start it, and pull it into the wash. The 'walk' is not an alternative to the 'drive'—it is the necessary first step of the 'drive.'". I laughed out loud reading this.
A very important note here, this doesn't mean Deepseek is the best LLM out there this along with every other benchmark in this world test LLM on one and only one thing, there might be the case that gemini fail on question 1, 3, 4 and deepseek fail on 2, 5 and 6. The point is you cannot trust single LLM you have to use all LLMs. I feel very strongly of people arguing about what LLM is the best and they will use only one LLM, this should not be the case this is not a search engine problem that only google is the best one (I know there are people who disagree). But this is logic problem. Let's say you have a really important feature to deliver and you want to discuss with engineers, you don't just get the best engineer and ship the feature, you try and get top engineers and get their feedback on that. Then why trust on single LLM, let them argue each other to get best possible response.
Link to the debate: https://debate.tellodb.com/share/walk-or-drive-to-carwash
The chats always want to please you, which creates a bias. See too many people validating their ideas in Claude and ChatGPT and they tell them exactly what they want to hear. Just opening another session and saying beat this up for me and poke holes in it is interesting to see results each time.
Yes the platform makes it easier
The car wash example is a perfect illustration of why “LLM as oracle” is dangerous and “LLMs as a panel” is way more robust, especially for non-trivial decisions. Beyond debates, have you explored ways to surface where models diverge most (e.g. highlight the exact assumptions they don’t share) so users can inspect that directly?
There is a chairman at the end who consolidate response from all LLM it includes where model diverged or converged. And yes you are right you should always consult multiple LLMs as a panel to get most out of it.
What stood out to me isn't that multiple models debate—it's that you're treating disagreement as useful evidence instead of something to eliminate.
A lot of AI products optimize for producing an answer quickly. In higher-stakes decisions, understanding where capable models disagree can be just as valuable as the final answer itself.
Yes this is not for someone who wants the answer fast, It's for someone who wants the correct answer.
Glad it resonated.
Your reply made me think there's one strategic decision sitting underneath that tradeoff which becomes much more significant as the product grows, but I don't think I can explain the reasoning properly in a thread without oversimplifying it.
If you're interested, what's the best email to reach you on?
Feel free to reach me at [email protected]
Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.
I never trusted single LLM they always hallucinate this would improve the trust, good product!
Thanks! 🦾
This is a fun and smart experiment!
Love how you let them duke it out on that classic car wash riddle — the Gemini stubbornness had me cracking up. Really drives home the point about not putting blind trust in one LLM. Solid reminder to cross-check models more often. Nice work!
Thanks 💪
this is exactly right -- the verification layer is the bottleneck, not the generation. we see the same pattern when assessing people's AI skills (aisa.to) -- the gap between casual users and effective ones almost always comes down to whether they have a systematic process for checking output, or whether they just go with "does this sound right." multi-model debate is one of the better approaches I've seen for high-stakes decisions.
Glad your finding matches with what I found!
I like the idea of using multiple models , but I'd take it one step further : don't just let them agree - assign roles. One model generates , another critiques, a third verifies against sources or tests. The diversity of reasoning is what improves the final answer.
You can give custom prompts to each one, start with refinement first, one each LLM iteration give a custom prompt to generate, critique and validate. Thank you so much for your input, I'll try to make it more obvious or better UX.
@sharjeelabbas This is a great illustration of something I've run into building multi-agent systems: single-model consensus is often just "confident agreement," not correctness. The interesting part of your setup isn't the debate itself, it's that disagreement forces each model to actually justify its reasoning instead of pattern-matching to the most likely-sounding answer.
One thing I'd be curious about: have you noticed models converging faster on questions with an objectively verifiable answer (math, logic puzzles) vs. ones that are more judgment-based (strategy, prioritization)? In my experience, multi-agent setups add the most value on judgment calls, since verifiable questions don't really need a "debate," just a fact-check. But if debate improves accuracy even on verifiable questions like your car wash example, that's a stronger signal that single-LLM laziness is a bigger problem than people assume.
Also — 7 debate formats is a lot of design work. Which one produces the most disagreement in practice, and which one do you personally trust the most for real decisions?
To answer your first question it depends on which debate format you choose, for example while using refinement format models tend to converge pretty soon, and is ideal for objectively verifiable answer as they just fact check and try to improve the answer (there might be a case where first model don't have updated information).
Devils advocate is ideal where there are not one correct answer they basically disagree with each other till the end but in the process get's super deep and get you a better/deeper response.
And yeah Devil's advocate and oxford union converge very less, they just disagree to get the best out of LLM.
Man I do this all the time with my gym routine and what to eat. I just ask one app and follow it blindly without thinking twice. Never realized the answers were probably lazy until I read your piece. Gonna start asking different apps and see what they say about each other instead of just picking one and sticking with it.
Or let them argue with each other on this using Debate TelloDB!
The interesting part here isn't just getting a better answer, it's surfacing disagreement before people over-trust the first polished response. A useful next layer would be showing where the models split before the final synthesis, because that's usually where the real judgment call lives. I ran into something similar building DictaFlow, the fastest AI cleanup often sounds confident while quietly changing meaning, so we lean hard toward preserving the original words instead of "improving" them. Same idea here, speed is cheap, trust is the product.
Speed is of no use if you can't trust the answer. Good work on Dictaflow as well!
Nice article
Thanks 💪
Feel free to ask any questions I would be happy to answer, a like would go a long way!