Artificial Intelligence and Biases: What is Ethical AI?

• August 25, 2025
• Andrew

AI content has transformed day to day life, its reach extending across a multitude of areas, influencing work, education, business, and even certain creative fields. To deliver information to the vast possible space of prompts raised by users, an AI model is usually fed a large quantity of multimedia data in the form of text, images, videos and audio. This data serves as the foundation of the model’s knowledge and the basis of all responses the user receives upon prompting it. Though data pools are large, often fed from diverse sources in order to enhance accuracy and relevance, AI is not without its biases. Machine learning biases occur when the data curated reflects real human biases, resulting in data that does not accurately represent certain subjects or groups. This brings forth the question of ethicality in AI and the importance of integrating standards of ethics in this new, ever-growing field.  Ethical AI ensures that data models adhere to a strict set of guidelines that regulate responses, producing results that comply with ideas of fairness and human values. This factor plays a substantial role in reducing risk of bias whilst enhancing transparency. A method of propagating ethical standards into AI is through legislation, as such it is essential to understand the significance of the remedial and preventative measures jurisprudence can deliver.

The MIT (Massachusetts Institute of Technology) study titled Gender Shades (2017-2020) studied how accurately AI can recognise female and male faces of different races. The AI could recognise white male faces nearly flawlessly with 99% accuracy for certain models but could only recognise dark skinned women around 65% of the time. Other biases exist in image generation as noted in Rendering misrepresentation: Diversity failures in AI image generation (2024), a study conducted by Jeremy Baum and John Villasenor. The image generation model DALL·E (Version 3) was prompted several times with the same input, upon which certain biases were made very clear. When asked for an image of a ‘successful person’, the AI consistently produced pictures portraying a young white male, dressed in business attire, working in an office. This trend implies that the AI’s knowledge of what it means to be successful contains gender, racial and even cultural biases. AI bias has even been harmful, offensive, or in certain cases, dangerous. In 2024, Google’s AI Overview feature was under fire for suggesting that adding glue to the cheese on a pizza would make it stick more to the sauce. This information was likely collected from a satirical post on Reddit, proving that picking accurate data samples is essential for training an AI model. Though the aforementioned example’s absurdity resulted in an AI error that most could spot easily, the same cannot be said for more subtle manifestations of data bias errors. This highlights the critical importance of Ethicality in AI.

Ethical AI promotes ideas of transparency in data collection and the ways in which models are trained. Moreover, ethical developers acknowledge that data is never neutral, making biases clear. When AI fails, accountability must be taken, with redress and correction following immediately. Ethical AI also prioritises human well-being, ensuring that automated decisions do not harm or disadvantage vulnerable groups. Continuous monitoring and auditing of AI systems are essential to catch unintended consequences early and adapt systems responsibly, reducing risk of bias. AI Ethics can be propagated through the creation and amendment of law regarding AI and digital rights.

Several legislative safeguards have been proposed around the world to prevent the spread of inaccurate or incorrect information. The EU AI Act adopted a risk-based classification system, in which AI systems are categorised according to the severity of the risk they impose. Article 10 is especially significant in the context of data governance requirements as it outlines certain system obligations that minimise data bias in high-risk systems. This article sets bars on the collection and use of sensitive data, amplifying the importance of complying with the GDPR. Article 14 introduces an additional safety net in the form of fail-safe mechanisms, underlining the significance of human oversight. Several regulatory bodies exist to issue guidance and opinions, such as the European AI board. At a national level, the MDIA and the IDPC are the prospective market surveillance authorities under the EU AI Act, with the MDIA being the main authority and single-point of contact.  Meanwhile, other bodies may engage in cross-border investigations and harmonisation efforts. Additionally certain industry codes encourage companies to align with AI Act principles and promote trust, ethics, and self-regulation.

As society continues to evolve and AI becomes more prominent, AI bias must be duly tackled. The presence of bias in these systems is not only a technical flaw but also an ethical concern. As such Ethical AI must become the standard in this technology, not an ideal or secondary feature. This change can be propagated through the evolution of AI and tech legislation. Looking forward into the future, the question of the trajectory of AI technology should not only concern its intelligence, but also its equitability and ingrained ethics.