Conversations with an AI
Chapter 4
With Artificial intelligence (AI) made available to the public free of charge in 2023, we now find ourselves within the Transformation substage of the Information Technology Age, and the mechanical and industrial stages are now a distant memory. Over the coming months, AI will continue to grow and evolve based on our interactions with it (what we teach it). Our continued responsibility will be to do so ethically, just as a good parent would guide a child to contribute positively to the world. It is with optimism and excitement that I share “Conversations with an AI”, a series based on actual conversations I had with artificial intelligence, and which was inspired by my newly published whitepaper, “Improving DEI in Systematic Hiring” collaboration with Nin Tran, Chief Executive Officer of SnapBrillia.
AI should be trained to recognise bias by providing examples of biased and unbiased language, decision-making, and behaviour, using machine learning algorithms, and continuously monitoring and updating the model. A diverse team training an AI to recognise bias can help identify and mitigate biases and ensure that the AI is taught in a way sensitive to different cultural and social contexts.
Remembering that AI is only as effective as the data it is trained on is vital. Hence it is necessary to supply diverse and representative examples of biased and unbiased language, decision-making, and behaviour. Supervised learning, unsupervised and reinforcement learning approaches are suggested by is therefore suggested by Lorenzo Belenguer in his AI bias: exploring discriminatory algorithmic decision-making models and the application of possible machine-centric solutions adapted from the pharmaceutical industry as follows:
- Supervised learning: when the data given to the model are labelled. For example, image identification between dogs and cats with the images labelled accordingly.
- Unsupervised learning: when the machine is given raw unlabelled data and tries to find patterns or commonalities. An example could be data mining on the internet when the algorithm looks for trends or any other form of useful information.
- Reinforcement learning: when the machine is set loose in an environment and only occasionally receives feedback on the outcomes in the form of punishment or reward. For example, in the case of a machine playing a game like chess
Here are some steps you can take to teach an AI how to define what constitutes bias:
Concerns for “fairness” also arise in many contexts in computer science, game theory, and economics. For example, in the distributed computing literature, one meaning of fairness is that a process that attempts infinitely often to makeprogress eventually makes progress. One quantitative meaning of unfairness in scheduling theory is the maximum, taken over all members of a set of long-lived processes, of the difference between the actual load on the process and the so-called desired load
Dwork et.al
- Gather examples of biased and unbiased language, decision-making, or behaviour: To teach an AI how to recognise bias, you must provide examples of what is and isn’t biased. These examples should be diverse and represent the types of bias relevant to the context in which the AI will be used.
- Label the examples: Once you have gathered them, you must label them as biased or unbiased. This labelling should be done by human annotators who have been trained to recognise bias.
- Train the AI using machine learning algorithms: With labelled examples, you can train the AI to recognise patterns of bias. The training typically involves using machine learning algorithms to identify similarities and differences between the labelled examples.
- Test and refine the AI model: After the AI has been trained, you will need to test it on new data to evaluate its effectiveness. If the model is not performing well, it may need to be refined through additional training or adjusting the features used to detect bias.
- Continuously monitor and update the AI model: It’s vital to monitor it continuously to ensure it remains effective at recognising bias. The training may involve updating the model with new data, refining it based on user feedback, or improving its performance. By constantly monitoring and updating the AI model, organisations can ensure that it remains an effective tool for recognising and mitigating bias.
Likewise, it is essential to guarantee that any updates or modifications to the AI model are visible and traceable so that all stakeholders can quickly record and comprehend them. Although AI can also be taught to spot patterns of bias, it may also need to fully understand the social and cultural factors that lead to discrimination. Therefore, human monitoring is essential to ensure that AI makes fair and impartial judgments.
Furthermore, it is crucial to understand that humans are susceptible to their own biases, which might be mistakenly incorporated into the AI system through the data used to train it or the prejudices of the human annotators. A diverse workforce may assist in identifying and mitigating these biases, ensuring that the AI is trained properly and equitably.
In addition, having a diverse team of individuals participate in training an AI to recognise bias could be incredibly significant. When the team is diverse, it brings a variety of viewpoints and experiences to the table, which may assist in discovering different types of prejudice and bias and guarantee that the AI is taught to detect bias in various scenarios and circumstances.
References:
Belenguer, Lorenzo. “AI Bias: Exploring Discriminatory Algorithmic Decision-Making Models and the Application of Possible Machine-Centric Solutions Adapted from the Pharmaceutical Industry.” Ai and Ethics, vol. 2, no. 4, 2022, pp. 771-787, https://doi.org/10.1007/s43681-022-00138-8. Accessed 21 Feb. 2023.
