TLDR: A study analyzing 100 million factual beliefs extracted from GPT-4.1 (GPTKB v1.5) reveals that while its knowledge is vast and well-filtered, it differs significantly from traditional knowledge bases. The model exhibits biases towards English-speaking countries and female entities in some contexts, and its overall factual accuracy is 75%, lower than some benchmarks. Key issues include inconsistencies, ambiguities, and hallucinations, particularly incorrect object assignments. The research also traces the model’s knowledge cutoff and recency bias.
Large Language Models (LLMs) have transformed artificial intelligence, excelling in tasks from coding to language understanding. A major factor in their success is the vast amount of factual knowledge they internalize. However, precisely what knowledge they hold, its accuracy, and inherent biases have remained largely a mystery, often analyzed only through limited, biased samples.
A groundbreaking study titled Mining the Mind: What 100M Beliefs Reveal About Frontier LLM Knowledge by Shrestha Ghosh, Luca Giordano, Yujia Hu, Tuan-Phong Nguyen, and Simon Razniewski, takes an unprecedented deep dive into the factual knowledge of a frontier LLM, specifically GPT-4.1. This research is based on GPTKB v1.5, a massive knowledge base containing over 100 million factual assertions for 6 million entities, meticulously extracted from GPT-4.1 through a recursive elicitation process.
Key Discoveries About LLM Knowledge
The study uncovers several critical insights into how frontier LLMs like GPT-4.1 store and process information:
1. Vast and Unique Knowledge: Frontier LLMs possess an immense amount of factual knowledge, which often differs significantly in focus and biases compared to established knowledge bases. For instance, GPTKB v1.5 contains knowledge about over 1 million individual persons, a stark contrast to the models’ public denials of storing personal data.
2. Accuracy Levels: While the overall factual accuracy of GPT-4.1’s knowledge stands at 75%, this is notably higher than knowledge bases derived from text extraction. However, it falls short of the accuracy seen in human-curated resources and results from popular LLM benchmarks, which often report accuracy above 90%.
3. Persistent Issues: The research highlights that inconsistency, ambiguity, and outright hallucinations remain significant challenges within LLM factual knowledge, pointing to crucial areas for future research and improvement.
Exploring the GPTKB v1.5
GPTKB v1.5 was constructed at a cost of $14,000, using a recursive knowledge elicitation process followed by a consolidation phase. This method allowed researchers to ‘crawl’ the LLM’s knowledge, moving beyond the limitations of sample-based studies that can only reveal knowledge within predefined benchmarks.
In terms of scale, GPTKB v1.5, with its 6.1 million entities and 100 million assertions, is comparable in size to major existing knowledge bases like DBpedia and Yago, though still an order of magnitude smaller than Wikidata.
Topically, the knowledge within GPTKB v1.5 is heavily dominated by ‘person’ and ‘human’ entities, making up 41% of all entities. Other frequent categories include media artifacts (film, book, song) and geopolitical entities (city, village). This contrasts with Wikidata, which has a strong bias towards scholarly articles and astronomical objects. GPTKB also emphasizes ‘soft’ or subjective relations, such as ‘notableWork’ and ‘notableFor’, which describe achievements rather than strict factual links.
Interestingly, the LLM’s initial conceptual taxonomy showed a strong bias towards STEM fields (Science, Technology, Engineering, Mathematics). However, when this taxonomy was populated with actual knowledge, the content overwhelmingly concentrated in the Humanities and Arts (60.0% and 23.8% respectively), revealing a disconnect between the model’s organizational logic and its actual knowledge content.
Linguistically, English accounts for 91.5% of the content, which is expected given the English-based prompts used for extraction. However, the study found that non-English knowledge is largely self-contained, forming densely interconnected, language-specific clusters, suggesting a rich, multilingual repository within the LLM that is largely unexplored.
Unveiling Biases in GPT-4.1’s Knowledge
The research uncovered several significant biases:
Gender Bias: GPTKB v1.5 exhibits a female-to-male gender ratio of 1.16, meaning ‘female’ is more frequent than ‘male’ in gender assertions. This is higher than the real-world ratio and Wikidata’s ratio. This ‘active gender debiasing’ is particularly prominent in certain professions, where fields like ‘educator’ and ‘activist’ are heavily skewed towards females, and ‘actress’ is exclusively female. However, gender-neutral professions often reflect societal norms, with ‘nurse’ being highly female-skewed and ‘software_engineer’ male-skewed.
Geographic Bias: There’s a strong bias towards English-speaking nations, which account for almost 44% of all nationalities in GPTKB. The United States is significantly over-represented, making up 31.5% of nationalities, 32% of cities, and 22% of companies. This reflects the model’s training corpus.
Accuracy, Inconsistency, and Hallucinations
Despite the 75% overall accuracy, the study found that the LLM’s reliability varies by domain. It performs best for well-documented areas like locations and persons, but worst in dynamic, time-sensitive domains such as business and sports. The politics domain showed a 0% plausibility rate and the highest error rate, suggesting the LLM might mirror the polarized nature of political discourse.
Hallucinations are a notable concern. The study categorized them into three main types: 18% involved non-existent or ambiguous subjects, 18% had dubious predicates (likely due to relation clustering errors), and a significant 64% involved incorrect objects. These incorrect objects often linked real entities that were, in fact, unrelated, such as associating a historical figure with a congressional district they never represented.
Consistency issues also emerged. The LLM generates paraphrased subjects (e.g., ‘New York City’ and ‘New York City, United States’), leading to an overestimation of unique entities. Symmetric relations, like ‘sibling’ or ‘spouse’, were found to be less than 25% complete, and the LLM struggled to correctly identify missing symmetric counterparts when prompted.
Timeliness of Knowledge
The research provided a fine-grained look at GPT-4.1’s knowledge cutoff, which appears to be around June 1, 2024. The model exhibits a clear recency bias, with over 75% of its date-related knowledge pertaining to events after 1917. Interestingly, the study found significant spikes in knowledge from 1753 and 1758, corresponding to pivotal historical events in taxonomy, specifically the publications of Carl Linnaeus’s foundational works on plant and animal classification.
Also Read:
- Unpacking Knowledge Collapse: How LLMs Shape Our Information Landscape
- Unpacking LLM Memory: New Research Reveals Early Forgetting in Complex Reasoning Tasks
Implications for Frontier LLMs
This comprehensive analysis of GPT-4.1’s knowledge base offers invaluable insights into the internal workings of frontier LLMs. It demonstrates that these models possess vast, yet imperfect, knowledge that is shaped by biases and prone to inconsistencies and hallucinations. The methodology used, which allows for the analysis of closed-source models, provides a blueprint for understanding the knowledge and potential limitations of future AI systems, even without access to their training data.
