Which AI is best? They all converge, none is neutral

No AI is the best: not the freest, not the cheapest

Asking which is the best artificial intelligence today is like asking which is the best tire brand for a car you're not even sure won't get repossessed. It depends on the use, the bill, and how much censorship you're willing to swallow. After weeks of cross-comparisons, the verdict is uncomfortable for every vendor: Gemini, ChatGPT and Claude are the same product in different packaging, and the model that wins on Monday loses on Friday. The only honest answer remains "it depends".

They all predict the same, none is neutral

The diagnosis that prevails among the most technical users has nothing to do with technology: it's about marketing. ChatGPT, Gemini, Copilot or Claude work as word predictors with pretensions, and their comprehension in complex contexts has narrowed to a dead heat. What sets them apart is not the engine, but the vendor's bias and the legal framework each model was built under. Google drags its own, OpenAI another, and for certain queries —tax, political, gender-related— the answer always lands on the same side of the table. Some call it bias, others call it indoctrination. Either way, it never switches off.

The myth of the local model without a muzzle

The first instinct is to run a model at home with Ollama and feel safe from everything. It half works: running on your own machine, nobody audits your prompt and you don't depend on a service that can change its terms. But the freedom is a mirage. Qwen, the Chinese reference for local use, doesn't come without an agenda: it swaps the Western sarracena finger for love of the Party, and there are topics it simply won't tolerate. Mistral in its base version and the "uncensored" models on Hugging Face promise more than they deliver, because the advertised neutrality hides a bias of its own. As a bonus, a model installed months ago is so outdated it doesn't know about events from the very year it was downloaded.

Numbers that sound good and don't exist

The real risk isn't ideological, it's reliability. These tools produce false data with dangerous conviction: they cite non-existent sources, insert plausible figures, and write with a confidence that makes you doubt yourself. For a draft or an outline they're fine; for anything with consequences —money, health, a command that deletes files— verifying is mandatory. It's the only ground where all camps agree, even if they fight over everything else.

The bill is ruinous and data is the moat

Playing is free. Working is not. Those who use them daily agree that without opening your wallet they're good for tinkering and little else: subscriptions, API credits, rising rates. The message is blunt —AI is here to change everything, but it's a money pit—. Meanwhile, the race is blurring: the sharpest analysis holds that models are converging and will end up competing for access to exclusive data, with a social network's accounts or a search engine's ecosystem as the real moat, rather than engine power. The arrival of China's Kimi K3, with 2.8 trillion parameters, tightens that idea.

And amid all this, the anecdote that captures the moment: someone asks whether military intelligence servers are already based on biological computing. The only real thing is lab-grown neuron cultures, a dish of cells that plays Pong and dies within weeks. The rest is marketing with a bigger number.

Which AI do you stick with when the monthly bill arrives and the answer doesn't add up?