Biases are caused by… well it's 2 things. This is how the algorithm is designed and then the data was trained on if it's a trained algorithm. If it's not a trained algorithm, if it's not trained on data. If it's programmed, well then someone programmed that bias into it. It could be that someone has a rule that says, if someone applies for a position and this person has the wrong colour or the wrong gender, then don't even invite this person for an interview. That would be a discriminatory bias that was built directly into the rule. But I guess this is not what you think about and this is also usually not what we talk about when we talk about bias. And we talk about biases that we're learned from the data that these systems were trained on and this is well again it's caused by a combination of the training algorithm and the data but a lot of it comes from data. So let's take this again, getting a job or like a screening job applications. If somehow the way we train a system is to try to predict who might get a job and so we only invite those points of you who are like you to get a job and if it's then a workplace where let's say 90% off of the employees are seeing you, or 90% are men or 90% are white or whatever. Then there's a lot of somehow then the data is kind of skewed so it's particular people and then the system will learn these patterns and it will reproduce the patterns. The big example is this compass model that they developed in the US where they looked at the recidivision risk, like what is the probability that a convict will commit new crime during parole, and this system ended up highly overrating black people and highly underestimating the risk of white people committing new crime. Of course this is also based on statistics. I think it's 6 times more likely to be in prison or convicted for crime if you are a person of colour and this again has to do with some structural problems in society and some cultural historical things and this creates a bias because it means that if you use this system then to make future predictions or even take decisions about what to do with people and whether they should be permitted parole or not. Then that will be a biased position. So the bias comes from capturing and learning patterns and maybe somehow overrating the value of those patterns when it then comes with decisions or recommendations. And of course bias is natural, we humans are also biased. I think for humans this is kind of an availability bias, that we are biased by what we see and a model would usually assume that the world it's only what it has seen there's nothing else in the world than what the model has seen. If it has seen that all the people that were hired and accompanied are men, it will think that this is kind of the right, that this is a correct reproduction of how things are or how things should be.
The proper functioning of any artificial intelligence system depends on the quality of the information or data used to train the model. This also means that the AI system may reproduce or amplify biases present in that information, which can significantly influence the responses it provides to user queries. For example, if the information used to train the AI model incorrectly states that America was discovered in 1600, the system will reproduce this incorrect information every time a user asks a question about it. Furthermore, the interests of those who develop or finance these systems also come into play, since the selection and treatment of training data may be conditioned by ideological, economic, or political criteria. It cannot be explicitly stated that biases can be completely eliminated in an artificial intelligence system; however, models can be trained to be as objective as possible within the constraints of the system.
From my technical point of view, bias in AI is inevitable. To train artificial intelligence – that is, to get it to go from processing simple numerical data to generating something that a human can interpret, such as words or classifications- it is necessary to use large volumes of data. The root problem is that this data must be selected in advance, and the simple act of choosing to use some data and discarding others already introduces, by definition, a bias into the model. A very graphic example would be trying to create an AI that estimates a person's income based solely on their age. If we use only data from the 1,000 richest people in the world to train that model, the AI will learn a distorted reality. If we then go out into the street and apply that model to any average citizen, it will fail miserably, because its ‘worldview’ is biased toward billionaires. Therefore, eliminating bias 100%, because the source data is generated by humans who already have biases. However, we can reduce it. Users have a key role to play here. Especially in LLMs: tools as a simple as the “like” or “dislike” button on a response help to retrain the model and correct these deviations little by little.
Bias from AI systems comes straight from the data it uses—if the AI gets correct, accurate, and credible data, the output won't be biased at all. That data comes from humans, like when you give it a prompt with wrong info that sneaks in your own assumptions or mistakes, or when the AI searches the web and grabs junk from shady sites full of rumors or one-sided views. These sources taint the results because AI just mirrors whatever patterns it finds, good or bad. Eliminating bias completely? Impossible, because even humans are biased—we all carry opinions, experiences, and blind spots from our backgrounds—so expecting perfect AI is a pipe dream that ignores reality. No matter how clean you try to make the data, human input or imperfect web info will always creep in somewhere down the line. The best we can do is get smarter about checking sources upfront and questioning outputs, but total perfection just isn't happening. It's like trying to make humans unbiased; it won't work, so we adapt instead.
It has been proven that AI systems are increasingly causing distortions. It can be observed that false assumptions, misguided conclusions and prejudices are occurring more and more frequently and are intensifying. The background that is assumed to be the cause of this development and distortion of the content presented is that AI systems are based to a large extent on probabilities in their way of working. This results in more and more false statements. In addition, texts that are used by AI-based systems are used as a source for them. This vicious circle, which is quite problematic, leads to an ever-faster development, which promotes the distortions just mentioned and makes meaningful research is more difficult. Something similar happens when it comes to prejudices. In general, prejudices are to be classified as critical when it comes to social developments. Technical innovations, such as social media platforms in recent years and now increasingly AI systems, have already resulted in a development that promotes prejudices and platitudes instead of eliminating or refuting them. This can lead to serious problems for social developments. In particular, the education system, whose task it is to educate about generalized opinions, which are ultimately the basis for prejudices, faces the challenge that the frequent use of AI systems reinforces prejudices through the above-mentioned mechanism and does not eliminate them. The development of regulations that counteract this is urgently needed.
I think that is hard to answer, based on the fact that “bias” tends to pair with one’s subjectivity. Who will judge what bias is? Most likely a human, which, in turn, has an empirical bias of their own. To eliminate bias would be like expecting 100% of humans to agree on one topic in particular, how likely does that sound? That said, if one narrows down a scope enough, we can “bias” (or de-bias) the models to behave in a particular manner that we happen to agree on, but we can’t expect it to behave with our same level of approval in other areas.
If there is a lack of impartiality, none of the AI users will have confidence in it.
