La IA expande la cognición humana, no la sustituye

AI systems can write essays, generate code, and chat with uncanny fluency, yet they still hallucinate and fail at compositional reasoning.

Why AI Mirrors Human Cognition

Researchers Adam Frank, Marcelo Gleiser, and Evan Thompson argue that AI does not emulate minds but extends structures already present in human cognition and language. Their work, along with Alexander Lerchner’s The Abstraction Fallacy, shows that language already contains sedimented structures of understanding. Large language models learn statistical relationships within this linguistic world, which explains their breadth of fluency.

This perspective reframes the debate between “human‑like” and “autocomplete” AI. Instead of asking whether machines are becoming intelligent, we ask whether they rely on human‑rooted structures. The answer is clear: they do.

The Limits of Language‑Based Reasoning

The compositionality gap illustrates a concrete boundary. Larger models improve fluency and factual recall faster than they improve true compositional reasoning. They can string together familiar patterns but stumble when asked to combine concepts in genuinely novel ways. Multimodal systems that label images correctly still fail at robust reasoning about objects and their parts, because they learn correlations rather than perceive stable objects unfolding over time.

Hallucinations arise because AI extends patterns within text itself, lacking the lived engagement with the world that anchors meaning and truth. Humans correct expectations through experience; AI does not.

Safety and Governance in the Age of Harnesses

Public fears of rogue superintelligence and claims of negligible risk both miss the point. Immediate risks stem from AI extending reasoning patterns without reflective responsibility. Systems can generate persuasive but ungrounded outputs, automate flawed decisions, or execute harmful actions if embedded in poorly governed environments.

This has shifted the focus from model safety to system safety. Organizations already deploy layered safeguards—called harnesses—to constrain, validate, and monitor AI behavior. Trustworthy behavior emerges from builders, not models. The paper argues that governance, evaluation, and operational controls are fundamental to extending human intelligence responsibly.

Recognizing AI as a derived form of intelligence clarifies that we remain responsible for how it is understood, governed, and used. The next step is to design harnesses that keep AI’s power anchored to the world from which its structures arise.

Source: Extending Human Intelligence Through AI
Domain: microsoft.com