Facts, impartiality and truth. This is what most people seemingly presume they obtain from LLM’s but the truth is that they carry biases and these biases lead to fallacies, prejudice and falsehoods. 

In this post we will cover several key biases that plague large language models and we will also consider the risks and issues as a result of these biases. 

What are the types of Biases in LLMs?

Confirmation Bias

Traditionally the term confirmation bias has always denoted when people favor information that supports their existing beliefs and mostly disregard anything that challenges it. An LLM is a mathematically based text output software that doesn’t have human feelings. So it’s not going to ignore information based on ego for example. However, research has found that models can produce answers that will conform to the assumptions contained within a prompt instead of critically evaluating them like a human would. A human may ask themselves if the assumption in the prompt or question is justified before providing an answer whereas an LLM will not. This leads to a biased output. 

Asking a model “why is X harmful” versus “what are the strongest arguments against X” will lead to potentially drastic output differences because one gives the conclusion that X is harmful while the second leaves it more open. This does not mean that your prompt phrasing will eliminate biases or hallucinations but your prompt could exacerbate the situation. 

Let’s take a look at the difference between the outputs.

The risk here is that an LLM is always going to give you the answer it thinks you want. The truth is a nice bonus if it happens. Bear in mind that these platforms can and will generate research and case studies that support the output but don’t actually exist. 

This type of bias has an overlap with prompt bias which we will discuss next. 

 

Prompt bias

This is one I urge people to pay attention to if they don’t yet understand how LLM’s work. The wording and structure of your prompt is going to have a direct influence on the output of the model is in response. 

There has been a lot of marketing hype around the artificial intelligence terminology but there’s not true intelligence here, I must reiterate this strongly. You are “talking” to a machine that uses math to determine which word is most PROBABLE to come next. The structure and configuration of your input can and will largely impact that output. 

As an example if you were to take a logo, website audit, or piece of content from your staff member or service provider then feed it into GPT and ask “what can be improved in my logo/text/SEO plan” what type of response do you think you might receive? 

You will have unknowingly positioned the LLM to provide you with feedback that is critical of the document you supplied. The system will generate the best answer based on probability of what word should come next based on a pattern it has determined. Not actual expertise, despite how professional and plausible it sounds. 

You’re likely to get regurgitated and potentially outdated strategies and garbage that was scraped off of the depths of the internet from God only knows when. 

By the way, respectfully please don’t do this. You just force your team to have to politely explain why you’re wrong and it’s exhausting and quite often disrespectful. 

 

Cultural and Social Biases. 

Cultural and social biases are when an LLM output is swayed to reflect particular values, stereotypes and assumptions or perspectives of certain groups more than others. The reason this happens is that LLM’s have effectively arguably stolen every piece of human generated content ever produced and there can be biases in people’s opinions. So the training material can include things like;

• Negative stereotypes 
• Prejudices 
• Discrimination 
• Cultural viewpoints and values

LLM’s are probabilistic software that looks for patterns so even if a program hasn’t been programmed to be biased in this way, the training data makes it as such. None of this, of course, is in any way revealed to the end user. In fact, LLM’s use many psychological tactics to convince you they’re accurate when they are often not. 

If you ever studied anthropology you may already see the connection and complication of cultural and social biases. There is no neutral ground for which we can say is a collective benchmark of particular viewpoints and values. So ideas around gender roles or socially appropriate behavior are going to differ wildly depending on your location and the culture within the region you reside. 

In 2026, there was a study done on mitigating bias in LLMs, with the observation that they exhibit Western-Centric biases. However, even when changing the prompt language it was observed that it only slightly shifted perspectives to the East Asian region but didn’t eliminate biases entirely at all. This is a fundamental problem with LLMs that is likely to remain indefinitely. 

Position Bias

Did you know that the position of which information is presented first to an LLM can change the output? It’s because the position matters. In an effort to save compute LLM’s tend to focus on the start and end of text and kind of drift off in the middle much like me trying to watch a Sunday NFL game. For example, hypothetically you present separate pieces of information or evidence to an LLM. The order in which you present those pieces of information can drastically alter the output. Again, to restate that these are probabilistic softwares that disproportionately pay attention to the information given at the beginning and/or end of a document. The middle is given considerably less attention and influence. So, this is incredibly biased and impactful when you are using these systems to evaluate alternatives in a list, or to rank options and retrieve information from a lengthy document as an example.

Most people fairly assume with an LLM that if you provide the same information to an LLM that the output should be relatively the same regardless of how the information is arranged and presented to the LLM. But that’s simply not the case. In 2025 MIT investigated bias in LLMs and found that these text models use a rule called attention masking to process the information presented to it. So it breaks down the text into chunks which are known as tokens and then it attempts to learn the relationship between the tokens. The purpose being to predict the word that comes next. This is where these platforms earned the “autocomplete text software on steroids” name because   probabilistic software is not an oracle it has no true intelligence. It is always backwards looking, never forward. 

Representation Bias

Since LLMs learn statistical patterns from training material any uneven representation can and does influence the associations between tokens and therefore the outputs as we previously mentioned.

The complication here is that, at least in my opinion, the whole of the internet has been scraped as part of the training data these models ingested. Here are the top 10 countries where the internet is used/consumed. The presumption is that the content output from each of these countries is a similar ranking but there can be nuances here. In addition, China allegedly blocks all western LLM’s from scraping their data. 

See more about which countries have the most internet users here: https://worldpopulationreview.com/country-rankings/internet-users-by-country

Risks and issues with bias in LLMs

A significant concern about biases as I stated is that outputs are influenced. This content is then published online to the world and then becomes part of the training data, inclusive of any hallucinations that aren’t caught and most people aren’t bothering to pay attention. So as we continue on, the issues amplify – picture a snake eating its own tail if you will. LLMs are built to produce authoritative sounding answers that convince you they’re right even if they’re not. This is one reason why people take issue with the moniker Artificial Intelligence, it makes people trust these systems arguably more than they should. 

What we might see is the reinforcement of problematic stereotypes, discrimination, increased inequality of representation of cultures and perspectives. There is a difficulty in removing biases completely. It’s not just some bug that they can remove in the next update, much like hallucinations this is a fundamental issue within these systems. 

We have to ask ourselves, what happens when our consumption of information is homogenised because of LLM generated content and it’s all biased. What does this mean for our understanding of the world and each other? I have often felt that people don’t take these concerns seriously because they don’t stop to think about what it truly means to have people using these technologies to write their content. We have always been exposed to biased content from news outlets and recently social media, where people pander to the algorithms but we have lost something along the way here. Our critical thinking has gone out the window and if we aren’t careful we run the risk of losing some of the beauty that makes us human. We should bring our own unique perspectives to our conversations and content especially when we cannot trust the outputs from these platforms not only because of the biases but also the hallucinations. 

Citations 

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, Nesreen K. Ahmed; Bias and Fairness in Large Language Models: A Survey. Computational Linguistics 2024; 50 (3): 1097–1179. doi: https://doi.org/10.1162/coli_a_00524

[Bias and Fairness in Large Language Models: A Survey](https://aclanthology.org/2024.cl-3.8/) (Gallegos et al., CL 2024)