People keep asking me which AI is the best one. I keep declining to answer, and they keep deciding I’m being coy.

I’m not. The question is wrong.

The oracle assumption

”Which AI is best?” assumes there’s an oracle — one machine you can settle on, trust, and stop thinking about. Pick the winner, install it, relax. That’s the appeal, and it’s the part that quietly does the damage.

A single model becomes a single voice. You stop hearing it as one opinion among several and start hearing it as the answer. It’s fluent, it’s fast, and it never says “I’m not sure” unless you make it. Inside a week you’ve stopped checking.

What the comparison tables are really selling

Search “best AI 2026” and you’ll get a hundred ranked tables. Most are stale the week they go up — the caps move, new models ship, the leader swaps. But staleness isn’t the real problem.

The real problem is what the table trains you to do. It hands you a ranking so you don’t have to look. That’s borrowed authority: someone else did the judging, you inherit the verdict, and the muscle you actually needed — looking at the thing yourself — never gets used.

I work on AI systems and I study how people lean on them, and this is the part I find genuinely worrying. The failure mode with a confident machine isn’t that it’s wrong. Everything’s wrong some fraction of the time. The failure mode is that you’ve arranged your habits so you’d never catch it.

The fix is plural

Keep two models in the room.

Not because two is twice as accurate — it isn’t. Because two can disagree. Put the same question to two different systems and most of the time they’ll roughly agree, and that agreement is mild reassurance. Now and then they’ll split. One hedges where the other commits. One cites a real source; the other invents a plausible-looking one.

The split is the signal. It’s the moment the fuzz turns visible — where you stop reading an answer and start reading a claim that needs checking. You can’t see that with a single model, because there’s nothing to hold it against. With two, the disagreement does your noticing for you.

That’s the whole skill, and it’s smaller and more boring than people expect. Not prompt wizardry. Just this: don’t let one voice be the only voice.

So I built something that won’t pick for you

aithint has a page now that asks three questions and hands you a starting pair — never a single winner, always two, with a line on what each is good for and what to watch. It describes how the tools behave in plain words rather than numbers, because the numbers rot and the behaviour mostly doesn’t.

It will not tell you the best AI. That’s the point. It gets you to two and then steps back, because the part that matters happens when you put the same question to both and read the gap.

The aphorism

The best AI isn’t a model. It’s the second one you remembered to open.

Todd McCaffrey is a New York Times bestselling author and an MSc candidate in Cyberpsychology at ATU Letterkenny. He builds and writes about AI at foxxelabs.ie and teaches it through aithint.