Semantic Knowledge

Example

You know penicillin is an antibiotic. You have no idea who told you or when, and it does not matter, because the knowledge is usable without its source.

Semantic Knowledge in Humans

Definition

Semantic knowledge is the store of general facts about people, objects, actions, relations, the self, and culture, acquired through experience (⁠Binder et al., 2009⁠). It builds up across the lifespan from many kinds of experience and can be applied in situations beyond the ones in which it was learned (Lambon Ralph et al., 2017). It is decontextualized: a fact is normally retrieved without any memory of the occasion on which it was learned (⁠Tulving, 1972⁠).

Neural Substrate

Binder and colleagues' meta-analysis identifies a left-lateralized network of seven regions: posterior inferior parietal lobe, middle temporal gyrus, fusiform and parahippocampal gyri, dorsomedial prefrontal cortex, inferior frontal gyrus, ventromedial prefrontal cortex, and posterior cingulate gyrus, grouped into posterior association cortex, heteromodal prefrontal cortex, and medial limbic regions (⁠Binder et al., 2009⁠). The anterior temporal lobe is not identified as a distinct region, although the middle temporal cluster spans the full length of the gyrus and verbal foci overlap in left anterior superior temporal sulcus. On the hub-and-spoke account, the ATL is the transmodal hub in which modality-specific representations converge, and prefrontal regions supply control over access rather than storage (Lambon Ralph et al., 2017). The two accounts differ on lateralization. Binder's network is left-lateralized, with some bilateral representation in angular gyrus and posterior cingulate; the hub-and-spoke ATL is bilateral, consistent with semantic dementia requiring bilateral atrophy to produce severe semantic deficits. Lambon Ralph and colleagues report a left predilection for written word stimuli specifically, which is the only stimulus type Binder included, so the disagreement may partly reflect the stimulus set rather than the substrate.

AI at a Glance

Humans accumulate decontextualized knowledge across a lifetime and can derive facts they never directly learned from what they already know (⁠Tulving, 1972⁠). Language models hold this in two places that behave nothing alike: weights, where knowledge is acquired gradually and then frozen, and a memory layer, where a fact is extracted from a single mention and stays revisable indefinitely.

Fact extraction: The write-time operation by which a memory layer distills a durable statement from a conversation, discarding the exchange it came from.

Semantic Knowledge in AI

How Far It Has Been Measured

Measurement in AI: Directly Measured

AI Memory Performance: Exceeds Human provisional

Where It Lives in the System

What Has Been Measured

A 2024 preprint gave 10 AI chat models the WAIS-IV Information subtest, general-knowledge questions such as "Who was the first president of the United States?", using the manual's instructions and items, with answers scored by clinical psychologists against norms for ages 25 to 29; 9 of the 10 models earned a scaled score of 19, the test's highest, at the 99.9th percentile (⁠Galatzer-Levy et al., 2024⁠).
Against the WAIS-IV norms (mean 10, SD 3, from 2,200 adults), about two-thirds of people score between 7 and 13, so a score of 19 sits 3 SD above the human mean and 6 points above the top of that range; because 19 is the ceiling, the true margin may be larger (⁠Block et al., 2025⁠).
The spread across systems on the knowledge questions was narrow: all 10 models from OpenAI, Google and Anthropic scored 18 or 19, including the small on-device Gemini Nano, whereas word-meaning subtests varied far more (Vocabulary 4 to 19, Similarities 1 to 15) (⁠Galatzer-Levy et al., 2024⁠).
The comparison is treated as direct because the questions, scoring and norms were the human ones and the answer was not in view for either side, but items were typed rather than spoken, it is a single preprint, and the authors report no check on whether the test items appeared in the models' training text (what the model learned from before release) (⁠Galatzer-Levy et al., 2024⁠).
Facts held in a model's training weights track how often they appear online: across 10 models and 14,000 questions, accuracy on the 4,000 least-popular facts stayed at 15 to 19% even as models grew, while retrieval, looking facts up in an external store and placing them in the conversation, let a far smaller model beat GPT-3 on those facts; no human baseline was run (⁠Mallen et al., 2023⁠).
On SimpleQA, a 2024 preprint set of 4,326 obscure fact questions deliberately chosen to defeat GPT-4o, models ranged from 5.1% correct (Claude 3 Haiku) to 42.7% (o1-preview); a human checker scored 94.4% but could search the web, so this gap favours the side the test was easier for and is not counted as a comparison (Wei et al., 2024).
When five models (GPT-3.5-Turbo, GPT-4, and three LLaMA models) did the classic category-fluency task, the semantic networks built from their answers were less interconnected and less flexible than human networks; this concerns how knowledge is organized and searched, not how much is known, and gives no score on the human scale (Wang et al., 2024).