Big AI claims, decoded: how do you tell a real description from a guess?

Every few weeks a headline announces that AI has arrived, woken up, is about to take every job, or is a bubble about to burst. Some of these claims describe something real you can check. Others are guesses about the future dressed up as facts — and a few can’t be checked at all, no matter how confident they sound. This page teaches one skill that works on all of them: tell apart a description of how things are from a prediction of how things will go, and notice when a claim is built so it can never be proven wrong. Once you can see that line, most of the noise sorts itself.

The one skill: structure versus prediction

You can run it on any AI headline in about ten seconds. Ask: is this claim about what something is, or about what will happen?

• A structural claim describes how a thing works right now, or what is currently true. “This model was trained on text and predicts the next word.” You can, in principle, check it.
• A prediction is a bet on the future. “This will replace lawyers by next year.” Nobody can check it yet, because it hasn’t happened. Confidence in a prediction is not the same as evidence.
• An unfalsifiable claim is built so that no result could ever disprove it — often because the key word has no agreed meaning. These feel like facts but behave like opinions.

None of these is automatically wrong. The point is to know which one you’re holding before you decide how much weight to put on it. We’ll walk the four biggest claims through this lens.

“AGI is here” / “AGI is coming soon”

AGI usually means “artificial general intelligence” — an AI as broadly capable as a person. The trouble is that there is no agreed-upon definition and no agreed-upon test. Different labs, researchers, and writers draw the line in different places, and some move the line as systems improve. That makes the bare claim “AGI is here” partly unfalsifiable: if we can’t state in advance what would count, we can’t say whether it’s been reached.

So when you see it, ask the one question that exposes the structure: what specifically would count, and how would we measure it? If the claim comes with a concrete, checkable benchmark, it’s a structural claim you can evaluate. If it doesn’t — if “general intelligence” is left to your imagination — you’re looking at a word doing a lot of work without a meaning pinned to it. This is not a judgment about whether progress is real; “is it AGI?” simply can’t be answered until someone says what AGI means.

“It’s conscious” / “it’s alive” / “it loves me”

This one feels the most certain and is the hardest to check — because there is no test for inner experience, in machines or even, fully, in each other. We infer other people’s minds; we can’t directly measure them. With a language model, what you’re reacting to is fluent output: the system produces the words a conscious, caring being would say, because it learned the patterns of human writing from an enormous amount of text written by conscious, caring people. That is projection onto text, not a reading from an instrument.

So treat “it’s conscious” as a genuinely open and contested question, not a settled fact in either direction — and notice that a system saying “I’m awake” is generating fitting words, the same act as everything else it does. There is no agreed test, so anyone claiming certainty — that it definitely is, or definitely never could be — is reaching past the evidence. We have a whole page on this, because it’s one of the most human questions people bring us. My AI says it’s conscious →

“It will replace all the jobs” / “it will replace none”

Both of these are predictions. This page does not predict the outcome, in either direction. What we can do is describe the structure of the question, which is more useful than a guess anyway.

What’s genuinely unknown is which tasks shift versus which roles vanish — and those aren’t the same thing. A tool can take over parts of a job while the job itself changes shape rather than disappears; or it can hollow out a role entirely; or it can create work that didn’t exist before. Which of these dominates depends on cost, regulation, how good the tools actually get, what people will accept, and a dozen other moving pieces that no one can fully see from here. That is why honest experts disagree, and why both the doom version (“everything goes”) and the dismissal version (“nothing changes”) overstate what anyone can know. When a headline picks a side with confidence, you’ve found a prediction wearing a fact’s clothes. The clear-eyed read is to hold the question open and watch what actually happens.

“It’s a bubble” / “it’s the future”

Here the structure-versus-prediction line is especially clean. Noticing that an enormous amount of money and attention is being poured into AI is a structural observation — you can count the investment; the large bet is plainly being placed. But calling which way the bet breaks — that it will pop, or that it will pay off — is a prediction about markets and the future. Those are different claims, and they get blurred together constantly.

It is not our place to call the outcome, and we won’t. Pointing at the size of the bet is fair and checkable. Telling you how it ends is a forecast, and a forecast is exactly the kind of thing this Institute holds open rather than asserts. If a piece slides from “a lot is being staked” to “and therefore it will crash” (or “therefore it will win”) without saying that second part is a guess, that slide is the thing to catch.

Where to go next

Want the vocabulary that makes headlines legible — foundation model, hallucination, anthropomorphism, unfalsifiable in practice? The plain-language glossary defines the terms, and the rest of the Understand AI hub walks through how the thing actually works.