Bayesian inference explains why people accept nonsensical sentences

Language comprehension, much like vision, can fool us. A striking example occurs when people encounter sentences that compare things in logically impossible ways—such as "More students have been to Russia than I have"—yet still judge them as grammatically acceptable. A team of researchers has now explained why these "comparative illusions" occur and, more importantly, why some feel stronger than others.

The illusion phenomenon#

Comparative illusions arise when a sentence contains a grammatically embedded comparison that doesn't quite make sense. Despite the underlying logical flaw, readers routinely perceive such sentences as acceptable rather than nonsensical. For decades, researchers have puzzled over this quirk of language processing.

A noisy-channel framework#

The researchers built on prior work suggesting that the brain treats language comprehension like signal recovery through noise. The theory proposes that when you encounter a sentence, your mind calculates the probability that various interpretations could have been corrupted into what you actually read. An interpretation becomes more likely if it's both plausible and would naturally corrupt into the observed sentence.

To test this idea, the team developed a hybrid model that married statistical language models—algorithms trained on large text corpora—with real behavioral data from humans. This approach allowed them to quantify how strong each interpretation should feel, moving beyond merely identifying which ones people prefer.

New predictions confirmed#

The model successfully predicted not only the fine-grained gradations in illusion strength—why some comparative illusions feel more compelling than others—but also explained a previously mysterious pattern: sentences with pronoun-based subjects (like "I") triggered different illusion strengths than those with full noun phrases. These novel predictions aligned with empirical findings, strengthening the noisy-channel account.

Broader implications#

The findings extend beyond a quirk of sentence structure. They add to a growing body of evidence suggesting that noisy-channel inference—the brain's tendency to infer what was meant despite degraded input—may be a unified principle underlying both illusions and normal language comprehension. This framework potentially explains diverse language phenomena through a single computational lens.