I've noticed how over the past year, the claims that we've already achieved artificial general intelligence are becoming louder and louder. The latest buzz in Nature, of course, has added fuel to the fire. But there is a fundamental problem that many are overlooking.

Here's the point: people confuse two completely different things. On one hand, we have language models that demonstrate impressive results on tests and handle a variety of tasks. On the other hand — that does not mean we have created true general intelligence. This is precisely a mix-up between increasingly complex pattern recognition and actual intelligence.

Looking at the historical definitions of AGI, they always emphasized different qualities: reliability across different contexts, the ability to generalize when faced with novelty, flexibility. Not just high scores on tests in artificial conditions.

What’s interesting is that recent research shows that systems which excel at test tasks often fail with the slightest change in conditions. Medical models, for example, give correct answers even when key data is missing, but become unstable with small shifts in distribution. That’s not intelligence; that’s training for specific scenarios.

On an economic level, the picture is even more telling. Even the most advanced systems can reliably perform only a small part of real work tasks, despite high test scores. Recent data indicates that most companies have yet to see significant returns from AI implementation. This does not look like true general intelligence.

There’s another point often ignored. When language models and humans give the same answers, it doesn’t mean they reason the same way. I’ve seen examples where a model confidently drew a conclusion in an uncertain situation, while an expert human refrained from judgment precisely due to lack of information. Superficial agreement hides deep differences in the reasoning process.

Current systems remain fragile. They depend on how the query is phrased, lack stable goals, and cannot reliably reason in the long term. Even stories about models solving open mathematical problems are mostly about combining and iterating existing methods, not creating new strategies.

The problem isn’t just terminology. When these systems start being integrated into real decision-making processes in science and government, overestimating their capabilities can lead to serious errors in trust and responsibility allocation. Therefore, mixing advanced statistical approximation with general intelligence is not only a conceptual mistake but also a practical risk.

The models we have are powerful tools, yes. But they remain tools, not agents with true flexible competence. That’s an important distinction.