Artificial Intelligence in Postgraduate Research: Ethics of AI in Research

  • Past Events

On January 23rd 2025, Tom Stoneham (Professor of Philosophy at the University of York and Ethics Lead for the UKRI AI CDT in Safe AI Systems) presented a webinar on the subject of ethics of generative AI in research. Almost 200 people attended the talk, which outlined the training methods behind the models, their limitations and their impact on both cognitive function and the environment. 

UKCGE’s director, Dr Owen Gower, began by introducing Professor Stoneham and participants joined breakout rooms to participate in a discussion around the questions What words do you associate with AI?” and Is it a good idea to use genAI for these research activities?”

What words do you associate with AI? (349 responses)
Is it a good idea to use genAI for these research activities?

Once this exercise was completed, Professor Stoneham began by explaining that the word ethics’ has three meanings, but that his talk would address only one of these: the way of making choices in your life for which you take personal responsibility. For those seeking a set of rules about the application of AI, he suggested that the European Commission’s guidelines were a good place to start (he also said he would be happy to be contacted for help in this area). 

The EC guidelines have two important take-home messages which have a bearing on personal responsibility: 

  • Sometimes, being responsible means not using AI at all.
  • Transparency is more than owning up to using it – it also requires an understanding of its limitations.

Professor Stoneham then turned briefly to the workings of generative AI – by which he meant services like ChatGPT and Google Gemini as opposed to autonomous systems like DAISY or algorithms built for specific purposes. The Large Language Models behind genAI, he explained, are Stochastic parrots” in the sense that they try to work out the most likely response to a task or question. Although the predictions are often lifelike, the machine is still just guessing. 

He then turned to a series of examples, illustrating where generative AI has failed for a variety of reasons. Amongst these were the 2024 Google Gemini controversy, in which – when asked to generate an image of a 1943 German soldier – the software produced what one journalist termed racially diverse Nazis”. This, said Professor Stoneham, can be explained by the process of developing generative AI, which he outlined as follows:

1. Unsupervised deep’ learning, in which the software processes huge amounts of data drawn from the internet.

2. Supervised learning, in which humans go in and correct wrong outputs. Crucially, this work is done by people in low-income countries for poor rates of pay.

3. Reinforcement learning and fine-tuning, in which humans prompt and give feedback on the quality of the answers (embedding diversity, ensuring politeness etc). This is high-value work for which specialist prompt-engineers are paid in the region of $300,000 a year.

Professor Stoneham then turned to some further limitations of AI, citing a high-profile case in which a user of Google’s AI chatbot Gemini replied to a user asking a factual question about US households: You are a waste of time and resources … Please die.” Google told Sky News: Large language models can sometimes respond with nonsensical responses, and this is an example of that. This response violated our policies and we’ve taken action to prevent similar outputs from occurring.”

The Professor asked What’s an ethically acceptable error rate for this behaviour? One in a billion? This sounds rare but remember Google is pushing this out in systems used by billions of people, hundreds of times a day. We need to think carefully about these systems and what they’re doing.”

He then concluded his talk with a brief consideration of the impact of AI on cognitive function along with its environmental impact. In the latter case, our expectations of almost instant answers – he argued – has pushed companies to harness greater and greater processing power. Water is needed to cool systems as the processors generate considerable heat, and vast amounts of energy creates a large carbon footprint (although we don’t know how large because this information is kept private). By the end of 2023, he added, hyperscale data centres were using two billion litres of water a day, i.e. a glass of drinking water for every human on the planet. Many companies are also looking at using nuclear reactors to provide the necessary power because current power grids cannot keep up.

At the talk’s close, Professor Stoneham took questions from the audience and a discussion followed which included topics such as using generative AI to produce abstracts and whether it helps or hinders academic writing.

Related news