Is AI dumb? Sundar Pichai on language and intelligence

In a recent interview with Sundar Pichai, CEO of Google, on the “Decoder” podcast from The Verge, Nilay Patel asked a pointed question:

Do you think language is the same as intelligence?

Sundar Pichai answered (in short):

I think language encodes a lot of intelligence.

The sentence stuck with me for a few days. It builds a bridge between the AI doubters and the AI enthusiasts, and it describes both at once: the reach and the limits of what these systems can do today.

AI: chattering parrots or creative geniuses?

A system like ChatGPT writes text that often feels strikingly human. But is that intelligence? Skeptics see only “stochastic parrots” that string words together without real understanding. The other side points to “emergent” abilities: skills nobody programmed in that suddenly appear, such as a model writing a passable poem or carrying an idea further.

Language: a mirror of cognition

Language is more than a means of communication. It is a picture of how we think. Whoever talks about love, justice or freedom is squeezing highly complex thoughts into words. That very ability, making the abstract tangible, is the intelligence held inside language. And it is exactly that which a language model has been served, in the form of billions of sentences.

Collective intelligence: language as cultural memory

Language also carries the knowledge of those before us. Fairy tales, myths, historical texts: they hold the gathered experience of past generations. When we speak and read, we draw on that memory. A language model does nothing fundamentally different, just on a scale no human could ever read. What it pulls from it is the residue of a great deal of human thought.

Sapir-Whorf: does language shape our thinking?

The Sapir-Whorf hypothesis turns the question around: language doesn’t just mirror our thinking, it shapes it too. The famous example about the Inuit and their many words for snow is now seen as too neat. The underlying idea still holds: what a language has words for is easier for its speakers to think. Whether that effect is strong or weak is contested. But it raises an uncomfortable question: if a model learns mainly from language, does it also learn the thinking that runs through that language?

Emergent intelligence: surprises from the machine

Again and again these systems show abilities nobody built in on purpose: they write poems or propose scientific hypotheses. Proponents call this “emergence”, ability that appears on its own as you scale. Others counter that a good share of these jumps is more a matter of how you measure than genuine new skill. Both can be true. And both are good reasons neither to freeze in awe nor to wave it away too quickly.

The limits of AI: hallucinations and missteps

So far, so impressive. And yet none of this reaches human intelligence. The models trip over things that are obvious to us, such as doing arithmetic cleanly. And they “hallucinate”: they invent facts with full confidence. Anyone who has once missed one of those in a report knows that “sounds right” and “is right” are two different things. That is exactly where the difference lies: mastering language is not the same as understanding the world.

Conclusion: a look ahead

So is AI dumb? It’s the wrong question. Pichai’s sentence puts it better: language holds a lot of intelligence, and a model that masters language draws more from it than the parrot critique will admit. But not everything. It moves across the surface of our thinking, often surprisingly sure-footed, sometimes spectacularly off.

For practice, that means neither awe nor mockery, but judgement. We get a tool that sounds almost human and yet is not. Using it well means calling on its strength and knowing its limit. That distinction will stay with us for years. And, as so often with this technology, it runs through language.