Large Language Models (LLMs): Getting personal
Personalisation has the potential to democratise who decides how LLMs behave, but comes with risks for individuals and society, say Oxford researchers
A new paper from researchers at Oxford Internet Institute, University of Oxford, highlights the benefits and risks of personalising Large Language Models (LLMS) to their users.
LLMs are artificial intelligence systems that generate written responses to text prompts. Due to their stratospheric growth in the past two years, hundreds of millions of people now interact with LLMs. Yet, the initial design and development decisions behind LLMs mean that small groups of developers, researchers or human annotators are providing the technology with the information it needs to respond to queries. This influences the conversational norms, values or political beliefs embedded in a model. Invariably, without wider participation during training, there is a risk that the diverse worldviews of those who use LLMs are excluded or misrepresented in their text responses.
In their paper, published today in Nature Machine Intelligence, the Oxford researchers present personalisation as a potential solution to sustaining different worldviews in language technologies. Like any new technology, they argue the responsible adoption of personalisation requires balancing the new benefits it can bring while managing potential risks for individual users and society as a whole. These benefits and risks are not purely theoretical: it has recently become possible to personalise ChatGPT, a widely-known LLM developed by OpenAI.
For individuals, the benefits of personalisation include increased ease in finding information, in a format tailored to their communication preferences. The user may also enjoy a technology that better adapts to their diverse beliefs or memorises information about their needs. Personalisation may result in a more empathetic connection and a sense of ownership of it being “my technology”.
However, this greater usefulness and deeper connection to the technology may fuel over-dependence and addiction. As with other types of artificial intelligence (AI), there is the risk of people anthropomorphising the technology and becoming attached to it. Personalisation is not possible without personal data; so, there is also an increased risk of users’ privacy being compromised.
Personalising LLMs also impacts society. Personalisation can bring better inclusivity and democratisation by diversifying which members of society have influence on how LLMs behave. If answers are more in tune to individuals’ needs, labour forces using LLMs could also become more productive.
However, not everyone has equal access to technology and those excluded risk becoming more disadvantaged by a widening digital divide. A further concern is that personalisation could contribute to societal polarisation and echo chambers when individuals less frequently encounter beliefs different from their own. The technology also has the potential to become a powerful instrument in generating persuasive and targeted disinformation, which is already problematic in the online world.
Commenting on the findings Hannah Rose Kirk, lead author and DPhil Student at Oxford Internet Institute, University of Oxford, said: “It’s vital we start the conversation now on what responsible personalisation looks like, as the technology is being developed. That way we have the best chance of enabling individuals and society to reap its benefits, without a lag in understanding or regulating the risks.”
Professor Scott A. Hale, Oxford Internet Institute, University of Oxford added: “Examining the risks and benefits of personalisation now while approaches are still being developed is the best way to create more inclusive and responsible technologies.”
Dr Bertie Vidgen, Visiting researcher at the OII and The Alan Turing Institute, and co-supervisor and author, added, “Personalised AI models feel like an obvious win—but that’s true only up to a point! If we aren’t attuned to the risks as well as the benefits, the consequences could be huge. This paper brings some much-needed clarity to this important debate.”
Download the full paper, ‘The benefits, risks and bounds of personalizing the alignment of large language models to individuals | Nature Machine Intelligence‘ authored by Hannah Rose Kirk, Dr Bertie Vidgen, Paul Röttger & Professor Scott A. Hale.
Editors notes
Notes for Editors:
For more information and briefings, please contact: Sara Spinks/Roz Pacey, Media and Communications Manager. T: 01865 280528 E: press@oii.ox.ac.uk
About the Research
Hannah Rose Kirk’s PhD is supported by the Economic and Social Research Council grant ES/P000649/1.
About the OII
The Oxford Internet Institute (OII) is a multidisciplinary research and teaching department of the University of Oxford, dedicated to the social science of the Internet. Drawing from many different disciplines, the OII works to understand how individual and collective behaviour online shapes our social, economic and political world. Since its founding in 2001, research from the OII has had a significant impact on policy debate, formulation and implementation around the globe, as well as a secondary impact on people’s wellbeing, safety and understanding. Drawing on many different disciplines, the OII takes a combined approach to tackling society’s big questions, with the aim of positively shaping the development of the digital world for the public good. https://www.oii.ox.ac.uk/
About the University of Oxford Oxford University has been placed number one in the Times Higher Education World University Rankings for the seventh year running, and number two in the QS World Rankings 2022. At the heart of this success are the twin-pillars of our ground-breaking research and innovation and our distinctive educational offer. Oxford is world-famous for research and teaching excellence and home to some of the most talented people from across the globe.
