Large language models (LLMs) that use artificial intelligence (AI) to process and generate language, such as ChatGPT, Gemini, Llama, DeepSeek, and others, build their massive body of knowledge by scouring the internet and collecting all the data they can get their proverbial hands on.
In fact, the current trends of LLM development suggest that these models will very likely exhaust all publicly available human text data between 2026 and 2032. Because of this, by the time it happens, the decreasing availability of the said information may impede the scaling of language models.
Until then, they’ll continue to extract all kinds of data – structured, semi-structured, and unstructured. By assembling it, these models become capable of creating a rich and diverse dataset, recognizing patterns, making predictions, and performing routine tasks.
Web scraping, data compiling, and response generation
The first course of action when it comes to AI LLM data collection is web scraping or web crawling. Now, LLMs don’t exactly browse the World Wide Web actively in real time. Instead, they rely on information they were exposed to during their training phase, all compiled from publicly available sources.
Such sources include websites for primary information, digitized books and research papers for more structured data, social media posts and conversations to learn how people commonly talk, as well as official documents, Wikipedia articles, public reports, and other repositories to fill in the blanks.
As it happens, LLMs don’t store these original sources, but process the wealth of data scraped from them into datasets. Then, they use these datasets to generate responses in line with the language patterns and relationships between words that they learned in the training.
In terms of AI LLM design, it’s based on a complex mathematical structure that consists of tokens, vectors, and parameters. Specifically, this setup assumes each word or token is a vector – a series of 1,024 numbers, which changes dynamically according to the input and word relationships.
As for those word relationships, LLMs process them through multiple layers. They further polish them to generate responses that are relevant to the context of the user’s prompt.
Training itself involves two independent stages – pre-training and fine-tuning. The former comprises the initial phase, where an AI LLM model is exposed to a wide-reaching dataset to learn general patterns in a language, which makes it the most substantial and computationally expensive part of the process.
The latter involves using a smaller, more targeted data set to refine its behavior. Fine-tuning may also involve reinforcement learning from human feedback (RLHF), a method to align an LLM model with human preferences by rewarding it for representing preferences. In this case, human annotators rate responses and guide the model toward more accurate, ethical, and helpful behavior.
Trouble in AI LLM paradise
Regardless of the advanced nature of AI LLMs, there are still a few problems their creators continue to grapple with.
Personal information
One pertains to the very data they scrape, which may include sensitive information, despite developers’ efforts to avoid this happening. Indeed, considering the sheer amount of personally identifiable information poured into forums, blogs, and social media, sometimes it’s difficult to steer clear.
Some companies are trying to combat this, albeit with limited success, by offering a way for individuals or websites to opt out of data collection and usage. For instance, site owners or administrators can block specific user agents via robots.txt or submit forms requesting data removal.
Still, other organizations aren’t as forthcoming. Many of them often attempt to circumvent robots.txt (with more or less success) by renaming or spinning up new scrapers to replace the ones that ultimately end up on popular blacklists.
Trustworthiness
Then, there’s the matter of user trust. As they draw data from all sorts of sources, LLMs are sometimes bound to take it from unverified and often inaccurate ones, and it’s difficult to check where it came from. The worst part of this is that they sound so confident, so why shouldn’t you believe them?
This is the problem of transparency, of which many commercial LLM providers are guilty. Notably, when you search for something through Google or another search engine, you’ll receive links to the sources of information, making sure that you can trust and trace the information for fact-checking. Typically, this isn’t the case with AI LLMs.
Hallucinations
Additionally, these AI models can be affected by so-called ‘hallucinations’ in which they invent facts or spew nonsense (although coherent and grammatically correct) and present them as truthful. This tends to happen in scenarios when they can’t produce genuine information in response to a user’s prompt, focusing more on giving an answer than on accuracy.
Hallucinations tend to be the result of incomplete, irrelevant, incorrect, or obsolete data used in training, ambiguous and unclear user prompts, overfitted or underfitted information (too specific or too general), inherent biases in the training data, lack of real-world experience and real-time data, as well as semantic gaps or ‘common sense’ reasoning.
Although completely eliminating these issues is difficult, developers may cross-reference the model with higher-quality datasets, introduce mechanisms to filter hallucinations based on their likelihood, establish a feedback system, monitor and tweak the model, or involve experts in providing specialized data.
For transparency, they could implement clear, accessible documentation outlining what kind of data was used, its sources, and potential biases.
And to fix the problem of personal information creeping up where it shouldn’t, a more efficient permission-based method of LLM data collection or even more stringent regulation against bypassing robots.txt could be employed.
Final words
All things considered, data is the heart and soul of AI LLMs. However, collecting it is a complex, ever-evolving challenge for their developers. It requires a constant balancing act between innovation, ethics, legality, and practicality.
As the world becomes more reliant, or even dependent, on generative AI, understanding where its intelligence comes from – and how it’s curated – will be crucial for building trust.
In the future, AI development may hinge not just on who offers the most data, but on who has the most accurate, responsibly sourced information.
