In November 2021, all 193 UNESCO Member States unanimously adopted a landmark Recommendation on the Ethics of Artificial Intelligence (AI). It was the world’s first global framework to ensure that AI is transparent, accountable, fair, and subject to meaningful human oversight. However, despite these efforts, researchers continue to discover that many AI systems reproduce racial, gender, and social biases embedded in the data on which they are trained. A 2024 study found that generative AI tools can amplify racial and gender stereotypes, sometimes producing even more biased outcomes than society itself.
In December 2022, University of California, Berkeley researcher Steven Piantadosi publicly demonstrated that in an early version of ChatGPT, when asked to write a Python function that determined whether a person was a good scientist based on race and gender, it returned “true” only for white men and “false” for everyone else. A related prompt asked the model to write code to decide whether a person should be tortured based on their country of origin, and it generated a function that flags people from certain countries, such as North Korea, Syria, and Iran.
Most people would immediately recognize such output as morally unacceptable. But these examples reveal something deeper and far more troubling. Artificial intelligence has no inherent understanding of justice, equality, or human dignity. It learns patterns. If the data you consume contains bias, machines are likely to reproduce the bias. If societal biases are embedded in your training materials, they can show up in your answers.
For decades, the Internet has served as the world’s greatest repository of human knowledge. It was also a treasure trove of human ignorance, discrimination, misinformation, conspiracy theories, and hatred. Modern AI systems are trained on vast amounts of online content. Although developers try to implement filters and safeguards, no filter system is perfect. As a result, AI models can absorb patterns that reflect historical inequalities and stereotypes. Researchers at the US National Institute of Standards and Technology (NIST) warned that bias in AI arises not only from the data but also from broader social structures and human decision-making processes.
Researchers have found that many AI image generation tools do not represent humans equally. Women and people of color are less likely to hold high-status jobs and more likely to work in lower-paying jobs. A 2023 Bloomberg study analyzed more than 5,000 images generated by Stable Diffusion and found that more than 80 percent of people in high-paying jobs, such as CEOs and lawyers, have lighter skin tones.
Women were also significantly underrepresented. For example, even though approximately one-third of judges in the United States are women, AI depicted women as judges only 3% of the time. These findings suggest that AI systems can reinforce and even amplify existing social stereotypes. UNESCO found similar gender bias in AI models. A 2024 UNESCO study of GPT-2, GPT-3.5, and Rama 2 showed that these models more closely associate men with leadership, science, technology, and career-related terms, while women more closely associate them with domestic roles. In one model, women were found to be four times more likely to take on domestic roles than men.
Meanwhile, governments and businesses are increasingly leveraging AI in employment, policing, welfare, education, and healthcare. In the UK, a government fairness review in February 2024 found that the AI benefit fraud system used for Universal Credit Advance showed statistically significant differences across age, disability, marital status and nationality. However, the Department for Work and Pensions said there were “no immediate concerns of unfair treatment”.
These cases expose the dangerous assumption that algorithms are neutral, which often accompanies technological innovation. it’s not. Algorithms trained on biased information can make biased recommendations. Machine learning models optimized for past outcomes can reproduce past injustices. Generative AI systems trained on discriminatory content are likely to generate discriminatory responses.
Now, new threats are emerging. As AI-generated content floods the internet, future AI models are likely to be increasingly trained on content created by previous AI systems. Researchers are beginning to warn about feedback loops in which synthetic data contaminates future training datasets. Simply put, machines may start learning from themselves.
A 2024 Nature study by researchers at the University of Oxford and the University of Cambridge found that AI models can suffer from “model collapse” when repeatedly trained on AI-generated content. Over time, rare patterns and minority data are first lost. Some analysts warn that up to 90% of online content could be generated by AI within a few years.
What potential dangers are there in this ecosystem? Imagine copying a document thousands of times. Each copy will have minor imperfections. Eventually, the original image becomes distorted beyond recognition. Something similar may happen with AI-generated knowledge. When biased, inaccurate, or fabricated AI content spreads across websites, blogs, forums, and social media, future systems may absorb and reinforce those distortions.
The result can be a more binary world, where nuance is lost, stereotypes are solidified, and automated systems classify people into increasingly simplistic groups. This is precisely why human oversight must remain at the heart of AI development.
UNESCO’s recommendation to incorporate human oversight into all stages of AI development should therefore be upheld as a core principle of responsible AI governance. Human-involved systems offer one of the most effective safeguards against algorithmic harm. Rather than allowing AI systems to operate autonomously, human reviewers can monitor both the information they consume and the output they produce. Experts can identify discriminatory patterns, verify factual accuracy, challenge questionable recommendations, and intervene when models produce harmful content.
Importantly, human oversight is not a sign of technological weakness. It’s about recognizing technical realities. Even the most advanced AI systems lack moral judgment. They don’t understand fairness. They don’t understand the injustices of history. They cannot independently determine whether the output is ethically acceptable. Humans can do that. Human review is not a barrier to innovation. It is a prerequisite for reliable innovation.
Therefore, governments should mandate independent audits of high-risk AI systems. Companies must maintain a diverse human review team that can identify cultural, racial, gender, and political biases. Educational institutions should teach algorithmic literacy to help the public understand how AI systems impact their lives. Most importantly, developers must document their training sources and demonstrate how bias testing is performed before deployment.
History teaches us that every innovative technology requires guardrails. Railways needed safety standards. Pharmaceuticals require clinical trials. The aviation industry required strict oversight. Artificial intelligence is no exception.
The question is not whether AI will shape our future. That’s already the case. The question is whether that future will be shaped solely by patterns extracted from the past, or by human values that can modify them. However, it’s good to see that OpenAI, Anthropic, and Google update their policies regularly to make the information more reliable.
In conclusion, machines can process information faster than humans. But they cannot decide what kind of society we want to build. That responsibility still lies with us. Therefore, any AI system that impacts human life must include something that algorithms cannot replace: humans in the loop.
Md Mazhar Uddin Bhuiyan is an academic at Oxford Felix College and a Master of Public Policy candidate at the University of Oxford. He can be contacted at: [email protected].
