Algorithmic fairness is perhaps one of the most cited values when people talk about AI ethics. However, there is also deep disagreement about what it means, and what it requires in different contexts.
Using generative AI as an example, I aim to explore the senses in which ‘algorithmic fairness’ can be understood. In particular, I draw a distinction between two conceptions of algorithmic fairness: the procedural conception and the integrated conception.
According to the procedural conception, algorithmic fairness is satisfied to the extent that some morally relevant biases in the design of AI systems are removed. The integrated conception, in contrast, says that algorithmic fairness to the extent that AI systems refrain from producing consequences threatening a fair society. I argue that the integrated conception is a better approach to algorithmic fairness.
For the purposes of this article, generative AI is defined as AI systems that ‘can take inputs such as text, image, audio, video, and code and generate new content into any of the modalities mentioned. For example, it can turn text inputs into an image, turn an image into a song, or turn video into text’[1].
Procedural Conception of Algorithmic Fairness
According to the procedural conception, algorithmic fairness is achieved to the extent that AI systems are designed in a way that minimise certain morally relevant biases. This is a dominant conception, I believe, of algorithmic fairness in much of the recent discussion of AI ethics.
What biases, then, are morally relevant? Emilio Ferrara, for instance, has helpfully distinguished several major types of biases in AI:[2]
| Type of Bias | Description | Examples |
| Sampling Bias | The training data are not representative of the population they serve, leading to poor performance and biased predictions for certain groups. | A facial recognition algorithm trained mostly on white individuals that performs poorly on people of other races. |
| Algorithmic Bias | The design and implementation of the algorithm may prioritize certain attributes and lead to unfair outcomes. | An algorithm that prioritises age or gender, leading to unfair outcomes in hiring decisions. |
| Representation Bias | A dataset does not accurately represent the population it is meant to model, leading to inaccurate predictions. | A medical dataset that under-represents women, leading to less accurate diagnosis for female patients. |
| Confirmation Bias | An AI system is used to confirm pre-existing biases or beliefs held by its creators or users. | An AI system that predicts job candidates’ success based on biases held by the hiring manager. |
| Measurement Bias | Data collection or measurement systematically over- or under-represents certain groups. | A survey collecting more responses from urban residents, leading to an under-representation of rural opinions. |
| Interaction Bias | An AI system interacts with humans in a biased manner, resulting in unfair treatment. | A chatbot that responds differently to men and women, resulting in biased communication. |
| Generative Bias | The model’s outputs disproportionately reflect specific attributes, perspectives, or patterns present in the training data, leading to skewed or unbalanced representations in generated content. | A text generation model trained predominantly on the literature from Western authors may over-represent Western cultural norms and idioms, under-representing or misrepresenting other cultures. Similarly, an image generation model trained on datasets with limited diversity in human portraits may struggle to accurately represent a broad range of ethnicities. |
Generative AI is particularly vulnerable to biases. As Ferrara pointed out, ‘text-to-image models like StableDiffusion, OpenAI’s DALL-E, and Midjourney…When prompted to generate images of CEOs, these models predominantly produced images of men, reflecting gender bias.’[3].
There are, however, a variety of approaches to the mitigation of such biases.The first approach is data preprocessing, which involves preprocessing the training data of AI such that they are representative of the entire population, including historically marginalised groups.
This can be achieved, for instance, by data augmentation, which involves ‘creating synthetic data points to increase the representation of underrepresented groups…[or by] adversarial debiasing, which involves training the model to be resilient to specific types of biases’[4].
In the context of generative AI, this might involve, for example, the inclusion of diverse data sources that reflect the diversity of human experience, to avoid overrepresenting any demographic group in the final output of the system.
The second approach is better model selection, which entails endorsing techniques that prioritise the mitigation of biases. In the context of generative AI, for example, the relevant AI system can penalise the discriminatory outputs it produces, making it less likely to generate similarly biased outputs over the long run.
The third approach is to post-process AI decisions. In generative AI, for instance, this might include ‘using additional filters or transfer learning techniques to refine the models further. Regular audits, continuous monitoring’[5], and so on.
While these are some viable ways to mitigate biases by AI systems, they alone cannot adequately capture AI’s wider social consequences that can be morally concerning from the standpoint of fairness. I argue that we need a more integrated conception of algorithmic fairness that does not simply focus on the minimisation of biases in the design phase of AI.
Integrated Conception of Algorithmic Fairness
According to this conception, algorithmic fairness is achieved to the extent that AI systems not just fulfil some morally relevant design requirements, but also address other areas of moral concern in relation to fairness. We care about fairness in employment, fairness in education, fairness in criminal proceedings, fairness in political participation, fairness in the allocation of public resources, and so on. The conception recognises the intimate relationship between AI and other areas of moral concern from the standpoint of fairness.
The key reason why the integrated conception is more appealing is that AI systems, however well-designed they are, can be misapplied in ways that undermine fairness in other domains. Consider the following cases in generative AI.
Scenario 1: Fairness in Employment
Imagine an organisation implements a generative AI system to assess job candidates based on their responses to assessment questions. The creators of this AI system ensured that the training dataset included a diverse range of ethnicities, genders, and accents to uphold fairness. Every feasible method to diversify the dataset has been explored, and the generative AI model has been meticulously selected.
However, despite these precautions, potential fairness issues persist. For instance, candidates from higher socioeconomic backgrounds may have better access to resources such as coaching for AI interviews. Consequently, they are better positioned to provide responses that closely match what the AI system is programmed to prefer.
Scenario 2: Fairness in Educational Attainment
A school system introduces a generative AI system to analyse student performance data and generate suggestions for individualised learning pathways. This AI recommends specific courses, resources, and pacing customised to each student’s unique needs. The developers ensured that the AI was trained on a diverse dataset, encompassing a broad spectrum of abilities, backgrounds, and learning styles.
However, students from lower socioeconomic backgrounds may face limitations in accessing the technology necessary to fully engage with AI-driven learning tools. If the AI system fails to consider these discrepancies in access, it could exacerbate the achievement gap by providing more affluent students with tailored and effective learning experiences, while leaving the students from modest financial backgrounds at a disadvantage.
Scenario 3: Electoral Fairness
In a national election campaign, a generative AI system is used to produce a series of deepfake videos. These videos purportedly depict a candidate making inflammatory and derogatory remarks during private meetings or public speeches. Rendered with high realism, they pose a challenge for the average viewer to discern from genuine footage.
Many voters who encounter these deepfake videos come to believe that the candidate indeed uttered these statements, significantly altering their perception of the candidate’s character and policies. Consequently, this misinformation dissuades individuals from voting for the candidate, founded on erroneous premises. Additionally, less financially endowed campaigns lack the technical capabilities or swift response teams required to refute false claims before they disseminate widely.
Scenario 4: Fairness of Competitions
An international film festival extends an invitation to filmmakers to submit short films judged on creativity, storytelling, and technical execution. Renowned for showcasing original works by emerging artists, the festival often serves as a platform for winners to secure significant career opportunities, including contracts with major studios.
Among the participants, one filmmaker opts to use advanced generative AI tools to craft a short film. This AI possesses the capability to write scripts, generate realistic human-like CGI characters, and even compose original scores, all based on input parameters or existing cinematic styles. The filmmaker inputs a theme and some stylistic preferences, and the AI autonomously produces a polished short film with minimal human intervention. Filmmakers lacking access to such advanced AI tools complain that they are significantly disadvantaged, and the competitive landscape places undue emphasis on technology rather than artistic talent.
Certainly, there are significant debates regarding whether there is any morally objectionable ‘unfairness’ in these scenarios. For instance, it could be argued that in scenario 4, there is no issue if the film festival has not explicitly excluded AI-generated films from consideration. Providing an informed moral response to all these cases is beyond the scope of this article. However, the point remains that fairness-related concerns arising from such scenarios are not uncommon.
The procedural conception of fairness does not account for these scenarios. It simply emphasises which biases in the design of AI should be mitigated and why. However, here are a few questions about the integrated conception that seem to make the procedural conception more favourable. The first question is, if we do not pay attention to the design of AI, then what else? Regardless of the moral considerations we have in mind, at the end of the day, it all comes down to how we ought to design AI. Yet, this thinking mistakenly places the burden of upholding algorithmic fairness exclusively on AI designers. We cannot disregard questions about how we should deploy AI to avoid incurring fairness-related concerns. It should not be assumed that there are no constraints on what purposes we use AI to serve.
The second question concerns the considerable moral disagreement over what fairness requires in different contexts. Since the integrated conception of algorithmic fairness encompasses too many other potential areas of moral concern, one might worry, it offers little guidance for AI developers and scientists.
However, the distinction between the procedural and integrated conceptions does not primarily aim to guide how exactly we should design AI. It merely highlights what algorithmic fairness should entail: whether it should solely focus on minimising biases or whether it should encompass the numerous other domains in which our interest in fairness is relevant.
Conclusion
Finally, you might wonder if there are viable ways to address the numerous potential issues regarding the fairness of your AI design and deployment. Recognising the challenges you may face, at IGS, we offer a series of training modules in AI ethics, such as
Get in touch with us, and we will be able to provide advice on the best AI ethics training we can offer for you.
[1] “What Is Generative AI?” n.d. NVIDIA. https://www.nvidia.com/en-us/glossary/generative-ai/#:~:text=Generative%20AI%20models%20can%20take.
[2] Ferrara, Emilio. 2023. “Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies.” Sci 6 (1). https://doi.org/10.3390/sci6010003.
[3] Ibid.
[4] Ibid.
[5] Ibid.
