Are Current Large Language Models Already AGI? Scholars Debate

Recent academic publications have examined whether today’s large language models (LLMs) already meet established definitions of artificial general intelligence (AGI). These discussions draw on experimental results, theoretical frameworks, and comparative studies of model performance.

In early 2026, Nature published an essay stating that current LLMs fulfill the first two levels of a proposed intelligence cascade, which covers Turing-test and expert-level performance. The essay asserts that, based on these criteria, current systems display a high degree of general intelligence.

Four faculty members from the University of California San Diego contributed a Comment to Nature, arguing that LLMs can be considered AGI by reasonable standards. Their backgrounds span philosophy, AI and data science, linguistics, computer science, and policy, reflecting a multidisciplinary approach to the question.

The authors outlined a tiered framework for general intelligence, including Turing-test, expert, and superhuman levels. They stated that leading LLMs meet the first two tiers, suggesting that these systems possess broad, adaptable capabilities across a range of tasks.

What the numbers show

• In March 2025, GPT-4.5 was identified as human 73% of the time in a Turing test
• A 2023 study analyzed 591 LLMs for general cognitive-like abilities
• A 2025 survey identified ongoing limitations in LLMs regarding causality and memory

The Nature essay also addressed the training and application differences between LLMs and humans. It stated that LLMs are trained on multimodal data and can assist with non-linguistic tasks, such as recommending steps in scientific research.

Objections to classifying LLMs as generally intelligent often cite their lack of embodiment or autonomy. However, the Nature essay argued that these factors do not prevent LLMs from being considered generally intelligent under the proposed framework.

Additional research in 2023 found evidence of a general factor of ability across hundreds of LLMs, indicating that these models exhibit interrelated cognitive-like skills similar to human general intelligence. However, a 2025 survey concluded that LLMs still face challenges in areas such as embodiment, symbol grounding, causality, and memory, which some researchers consider essential for human-level intelligence.

The Nature Comment further stated that general intelligence does not require flawless performance, universal mastery, or human-like architecture. The authors argued that occasional hallucinations in LLM outputs do not disqualify these systems from being classified as generally intelligent.

* This article is based on publicly available information at the time of writing.