# Claim: Non-linguistic AI reasoning limits

_**What you are reading**: one claim-node in Deliberus's deliberation graph. Deliberus maps human reasoning — including claims that are contested, mistaken, or extreme — so that disagreement can be examined precisely. A claim's presence here is not an endorsement: the graph records that someone argued it, how it connects to other reasoning, and how far scrutiny has gotten. The metadata on this page describes the state of that scrutiny, not the truth of the claim._

---

> Artificial intelligence systems that lack extensive natural language training do not exhibit powerful domain-general reasoning compared to language-based models.

- **Type**: empirical
- **Epistemic status**: observed
- **Provenance**: extracted
- **Attributed to**: Daniel Rothschild
- **Confidence**: 0.75
- **Source**: [Language and Thought: The view from LLMs](/extraction/language-and-thought-the-view-from-llms.md)
- **Canonical**: `/claim/claim_0b4a87cb87c1` (this document: `/claim/claim_0b4a87cb87c1.md`, data: `/claim/claim_0b4a87cb87c1.json`)

## What this claim argues
- this **supports** → [Large Language Models (LLMs) are effective reasoners because the compression inherent in natural language as a medium makes the computation…](/claim/claim_4284f4477222.md) _(scheme: evidence_to_hypothesis; strength: 0.80)_
- this **qualifies** → [OpenAI’s best-performing [AI] model can only solve 8% of tasks on a benchmark consisting of internal research and engineering bottlenecks e…](/claim/claim_0b424ec5cd8f.md) _(strength: 0.60)_

## What bears on this claim
- **attacks** this ← [Artificial intelligence models develop a universal internal representation (a 'universal language') that is shared across different archite…](/claim/claim_0c8cd55ba3d7.md) _(strength: 0.40)_

## Scrutiny state
- **Age**: entered the graph 2026-03-31
- **Derivational completeness**: `bundled` — exposure 0.00 across 1 reachable claim(s), decomposition depth 0
- **Open sorry-frontier**: 1 claim(s) in this subtree not yet decomposed or supported
- **Horizon reading** (read 2026-10-10, refreshed on every read): `unexamined` — nothing beneath this claim has been tested yet — a kind is read only from a tested arrangement, never assigned; attaching and scoring an attack, a test or a definitional split is what makes one readable
- **No gate**: this reading routes and counts; it permits and forbids nothing, and one new attack changes it. What no reading can see is the move nobody has made yet — contribute it
- **Structure proposed (deterministic, never the reading)**: `empirical` (confidence 0.80) — Empirical leaf: reduces to a resolvable factual question.
- **Full oracle**: `/claims/claim_0b4a87cb87c1/completeness` · this reading: `/claims/claim_0b4a87cb87c1/horizon` · corpus falsification metric: `/graph/horizon-map`

---
_Deliberus — a structured deliberation graph. The graph has no permanent bottom; every claim is invitable to decompose deeper._