In a sense, AI is like a child: its beliefs and perceptions are formed based on what and how it is taught. The information he receives determines his responses, and embedded algorithms and data influence his worldview, just as upbringing and education shape a child’s consciousness. AI is increasingly integrated into our daily lives, but along with its benefits come serious challenges. One of the most debated is the issue of AI bias, when algorithms can reproduce and even reinforce human stereotypes.
Journalist Remy Demichelis, author “Artificial Intelligence, its biases and ours” – books about bias in AI systems, assures, that “algorithms not only reproduce the bias inherent in society, but also reinforce it.”
AI algorithms often reproduce human stereotypes, which leads to discriminatory treatment due to belonging to a certain race, gender and other social groups. Attempts to correct these distortions have proven difficult, and the lack of uniform regulatory standards only exacerbates the problem.
There are plenty of examples of how AI can create distorted images. In 2023, the neural network Midjourney created negative images of the French suburbs, which quite naturally caused criticism and protests in the audience. Searches related to the bathhouse produced stereotypical results that contrasted greatly with common perceptions of France. Similar discrimination was also observed in AI chatbots, which show bias against women, people with disabilities and representatives of ethnic minorities.
The nature of bias in AI
The root of the problem is the data on which the algorithms are trained. For example, about half of them are in wedding photo databases comes from from the United Kingdom and the USA, which distorts the idea of wedding traditions in the world. This can lead to the AI identifying national costumes from other cultures as “exotic” or “folk costumes” rather than standard wedding attire.
The influence of bias in AI is recognized at the state level. The main challenges are not only the algorithms themselves, but their developers. If systems are created by white men, they may unwittingly ignore issues of discrimination.
Algorithmic Justice League stands out for testing algorithms to assess their bias. But such inspections do not yet have clear standards and are difficult to verify.
Attempts by technology companies to reduce bias are not always successful. For example, last year Google Gemini got into into a scandal due to the creation of historically incorrect images in an attempt to be “politically correct”. Critics, including Elon Musk, have accused AI of “distorting reality”.
Bias in AI occurs when algorithms are trained on data that already contains human stereotypes or imbalances. This can lead to discriminatory outcomes in, say, hiring, lending or justice.
Creating a completely neutral algorithm is challenging because AI models are trained on vast amounts of data that can contain cultural, racial, and gender biases. This means that even the most advanced systems can inadvertently reproduce these stereotypes.
The fact is that some open databases used for AI training have a disproportionate representation of different cultures, races, and genders. This can cause algorithms to be less accurate or biased against underrepresented groups.
Studies have shown that some facial recognition systems have a higher error rate when identifying dark-skinned people, especially women. This may be due to the fact that such systems were trained on data sets dominated by images of light-skinned men.
Some companies used AI to screen resumes, but these systems were found to favor men over women because they were trained on historical data where men were more likely to be hired for certain positions.
AI and social inequality
Automated systems are used for credit assessment, hiring, risk determination in courts and other important areas. But if algorithms are trained on biased data, they can make discriminatory decisions.
In the US, the system COMPASS assessed the likelihood of reoffending. The study found that the algorithm was more likely to falsely identify blacks as high-risk and whites as low-risk, even when this was not true.
Amazon created an AI to automate hiring, but the algorithm underrated women’s resumes because it was trained on male-dominated data, leading to gender discrimination.
These cases show how AI can reinforce stereotypes and inequalities. Women and minorities may be unfairly excluded from employment opportunities, credit, or access to education and medicine due to biased algorithms.
Regulation of AI in the world
European Union, USA and China have different AI strategies and regulations that reflect their priorities and values.
The European Union became the first jurisdiction in the world to adopt a comprehensive Law on artificial intelligence, which prohibits AI from using biometric data to classify people by race, political opinion or sexual orientation.
Europe lags behind the US and China in the development of AI. Its tough regulatory efforts could lead to big tech companies pulling their products out of the European market. For example, Meta (Facebook) said it would not launch its Llama AI model in the EU due to “unpredictability of regulatory policy”.
The European Union wants to become a leader in the regulation of AI, creating a law that will classify AI systems according to the level of risk: from very high to minimal. For example, the use of AI for real-time biometric identification in public spaces is prohibited due to the high risk of human rights violations. Low-risk systems are subject to minimal control. Violations of the law can lead to fines of up to 35 million euros or 7% of the company’s global turnover. The future regulation of AI will depend on whether Europe can find a balance between security and technological development in the face of global competition.
There is no single federal law on AI in the United States. Instead, regulation is carried out at the level of individual states and branch agencies. For example, the states of California and Colorado have implemented their own laws to ensure transparency and accountability in the use of AI. This approach gives companies more freedom to innovate, but at the same time can fragment the legal field.
For its part, China is strict controls development and use of artificial intelligence. In 2017, China announced its intention to become a world leader in AI by 2030, setting clear goals and standards. This approach provides centralized control, but can limit innovation and raise human rights concerns.
AI as a tool of authoritarian regimes, their influence on elections and manipulation of public opinion
Modern political campaigns use AI for analyzing voter behavior. Thanks to data from social networks, algorithms determine the preferences, interests and emotions of users. This allows you to personalize messages for different demographics. This approach – microtargeting – was used in election campaigns in the USA. Facebook and Instagram offer customizable ad campaigns that allow politicians to precisely target their messages to different segments of the electorate.
Generative AI is known to be able to create realistic images, videos and audio that are essentially “deep fakes” (deepfakes). These technologies are used to spread misinformation and manipulate people’s opinions. During the 2024 US presidential election, deepfakes were widely used to discredit the candidates. Yes, fake videos of politicians making statements that they did not actually make were circulated on social networks. This created confusion among voters and undermined confidence in the candidates.
There are not rare cases when authoritarian regimes use AI to strengthen control over the population. Algorithms analyze data from surveillance cameras, social media and other sources to track and suppress dissent. In China, the social credit system uses AI to monitor the behavior of citizens, granting or withdrawing privileges based on their “loyalty” to the regime.
Who should be responsible for AI errors?
When AI makes mistakes or makes discriminatory decisions, the question arises: who is responsible – the developers or the users? The European Parliament in its resolution notes, which, depending on the circumstances, may be the responsibility of the manufacturer, operator, owner or user of the AI. This shows how difficult it is to determine who is to blame for AI errors.
AI often is used to check content on the Internet. However, the question arises: does such a check become censorship? After all, removing “problematic” answers can prevent misinformation and hate speech, but excessive moderation can limit freedom of expression. It is important to make algorithms transparent and involve human oversight to maintain a balance between security and freedom of speech.
Obviously, the transparency of AI algorithms is important to prevent manipulation and discrimination. Companies must disclose details about how their algorithms work so that users can trust them. This includes information on data sources, training methods, and decision criteria. This approach will promote accountability and ethical use of AI.
Is it possible to create an ethically neutral AI?
That to create ethically neutral AI requires that models be trained on data that reflects the diversity of society. This will reduce the risk of bias due to one-sided or incomplete data. Adhering to moral principles and taking into account the possible social consequences of the use of technology will help the involvement of ethical committees in the development and implementation of AI.
Completely eliminating bias in AI is difficult due to various technological and social constraints. Algorithms can reproduce existing stereotypes, if they are present in the training data. Even efforts to eliminate bias can lead to new forms of discrimination. Therefore, it is important to constantly monitor the work of AI and adjust it, involving experts in ethics, sociology and other fields.
AI can become a powerful tool in the fight against discrimination. Algorithms they can analyze large volumes of data to reveal hidden biases in decisions, allowing for timely corrections. In employment, AI can help ensure fair selection of candidates by analyzing CVs without taking into account personal characteristics such as gender or ethnicity. But to achieve this, algorithms need to be carefully tuned and made transparent and accountable.
Tetyana Viktorova
