How AI and Computational Methods Are Transforming the Study and Teaching of AI Ethics

What Can Computational Methods Tell Us About AI Ethics?

Researchers at Northeastern University London have explored how computational methods, especially artificial intelligence (AI), can contribute to philosophical studies—and how philosophy can inform AI. Their focus is on AI ethics, which is a key part of their MA Philosophy and Artificial Intelligence and MSc AI and Ethics programmes. Students come from varied backgrounds, mainly philosophy or computer/data science. The goal is to combine philosophical insight with computational skills, preparing students to engage responsibly with AI technology.

In line with this approach, the researchers applied computational techniques to their own teaching materials, analyzing philosophy course texts to better understand their pedagogical strategies. This self-reflective project, involving alumni and current researchers, has yielded surprising and valuable results.

Canon Analysis: Exploring Course Texts with NLP

The team selected the reading lists from two courses—AI and Data Ethics, and Advanced Topics in Responsible AI—and performed computational analyses on these texts, calling this effort a ‘canon analysis’. Though not a formal canon, these texts have been curated over years to cover important topics in AI ethics.

Absolute Word-Frequency Analysis

Using a simple word-frequency counter, they identified the most common unique words across all texts. The word “human” appeared most frequently, which may seem obvious given the humanities perspective, but it’s notable given the technological focus of the material. The texts often contrast humans with data (the second most frequent word) and machines (eighth in frequency).

However, raw frequency alone doesn’t reveal the uniqueness of terms within this specific corpus. To dig deeper, a relative frequency analysis was necessary.

Relative Word Frequency Analysis (TF-IDF)

The researchers applied TF-IDF (Term Frequency–Inverse Document Frequency) to measure how important words are in the AI ethics corpus compared to another body of texts. For this, they used a contrasting “Wittgenstein Corpus”—papers discussing Wittgenstein’s work.

• Top 10 words in AI Ethics Canon: human, philosophy, ethic, moral, philosophical, robot, language, data, theory, system
• Top 10 words in Wittgenstein Corpus: political, technology, social, design, review, agent, science, develop, knowledge

This comparison highlighted that terms like “human” and “ethic” are particularly significant in the AI ethics texts. The Wittgenstein corpus, meanwhile, centers more on political and social themes. The prominence of “Wittgenstein” in its corpus confirmed the analysis was effective.

Using AI: Semantic Clustering of Texts

To apply more advanced AI, the texts were converted into numerical vectors using SciBERT, a transformer model specialized for scientific literature. This enabled measuring the similarity between papers via cosine similarity, where 1 indicates nearly identical content and values near 0 indicate low similarity.

K-Means clustering, an unsupervised machine learning method, was then used to group papers based on similarity. To optimize cluster numbers, the Elbow Method and Silhouette Score helped identify six as the best choice.

Because SciBERT embeddings are high-dimensional, Principal Component Analysis (PCA) reduced the data to 112 components explaining 95% of variance, making clustering feasible. The clustering showed strong internal consistency, with every paper most similar to its cluster center.

However, interpreting the themes behind these clusters proved difficult, possibly due to the small dataset and high dimensionality. This demonstrated the challenges of applying AI techniques effectively without large amounts of data.

Using AI: Topic Analysis with LDA

As an alternative, Latent Dirichlet Allocation (LDA) was applied. LDA is a topic modeling technique that identifies groups of words that tend to appear together, assigning probabilities of topics to each document. This soft clustering approach makes thematic interpretation easier.

The analysis revealed six coherent topic clusters:

• Topic 0: Social issues, social media, gender, culture
• Topic 1: Superintelligence
• Topic 2: Applied concerns such as sustainability, health, and the arts
• Topic 3: Robots, personhood, and artificial agency
• Topic 4: Design and responsibility
• Topic 5: Privacy and risk

The topics showed logical overlaps, such as superintelligence and artificial agency, which shared over half of their papers. The results provided clearer thematic insights than the previous clustering method.

What AI Can Reveal About Philosophy

While AI-powered knowledge discovery is common in STEM fields like biology and genomics, it has seen limited use in humanities and social sciences, including philosophy. However, natural language processing (NLP) techniques and pre-trained large language models are opening new possibilities.

This research demonstrates how AI tools can help map and understand philosophical literature in AI ethics. The team plans to continue this work, sharing insights with students to enrich their learning experience and refine teaching methods.

For those interested in further advancing skills in AI and ethics, exploring specialized courses can provide practical expertise. More information on relevant programmes and training is available at Complete AI Training.