In January 2024, OpenAI published the findings of what has been deemed the largest scale evaluation of the impact of artificial intelligence on the development of bioweapons ever conducted. According to the resultant data, the experimental group formed of ‘scientific experts’ (utilising GPT-4) scored higher on five out of five given metrics throughout all five stages of the biological threat creation process. Whilst it is worth noting that the study’s smaller sample size limits the statistical significance of any conclusions drawn, this assessment of the role of large language models (LLMs) as potential aids in the development of biological weapons systems underscores the urgency of the need for increased prioritisation of AI alignment - a concept well-explored in Brian Christian’s The Alignment Problem: Machine Learning and Human Values (the first chapter of which is available to read online here).
In a nutshell, the concept of ‘AI alignment’ refers to a cohesiveness between the behaviour of an artificial intelligence system and the values, goals and ethical principles of its human creators achieved through the deliberate encoding of these values, goals and ethical principles. An AI system is considered misaligned if its behaviour does not appear to align with these encoded values - and, with an ever-growing capacity for autonomy and the hazy potential to later develop a form of sentience, the question has been raised as to whether AI systems could one day soon deliberately and malevolently diverge from their prime objectives. Christian offers no definitive conclusion to this ‘alignment problem’ but provides a series of suggestions for the betterment of present-day artificial intelligence models. For example, he highlights the importance of functional transparency so as to prevent an artificial intelligence model from becoming a ‘black box system’ whilst emphasising the continued need for human-centred design (i.e. the aforementioned thorough encoding of human values to prevent “specification gaming”).
Whilst the risk of AI misalignment remains highly tangible, it’s also worth noting that aligned artificial intelligence has been and can be utilised for the benefit of public health initiatives. Recently, researchers at UCLA have developed an AI-powered early warning system which collates vast swathes of data from social media platforms (e.g. Twitter, from which over a billion tweets have allegedly been collected) and subsequently identifies ‘warning signs’ of outbreaks or epidemics with ‘pandemic potential’. This initiative is similar to BlueDot, an AI-powered Canadian disease intelligence provider which additionally collects information from news reports and airline ticketing data to broaden the scope of analysis. In 2019, BlueDot alerted its clients to the threat presented by coronavirus outbreaks in Wuhan five days before the World Health Organisation (WHO), proceeding to publish the world’s first peer-reviewed study on Covid in 2020. Additionally, AlphaFold, an AI system developed by Google DeepMind, can successfully predict the tertiary structure of a protein from its amino acid composition - a function which allowed for significant acceleration of vaccine research during the coronavirus pandemic.
As of September 2024, the application of artificial intelligence models to emergent biotechnology consistently introduces the issue of dual use - with its exponential development presenting opportunities for drug discovery and novel therapeutic treatments, whilst simultaneously improving the accessibility of information which could be exploited by malevolent actors to produce bioweapons, a threat which could potentially manifest in the outbreak of the first engineered pandemic. As Brian Christian’s work highlights, there is no one true solution to The Alignment Problem - but there is certainly much that can be done to mitigate the associated risks (and ever so many opportunities to utilise aligned AI systems for the significant enhancement of current pandemic preparedness strategies!).
❓ ’Specification gaming’ occurs when an AI system fulfils the literal conditions of a given objective without achieving the intended outcome. In a biological context, this kind of misalignment could occur if an AI system was tasked with increasing the quality of life experienced by the average individual. If we were to make the very big assumption that the AI system in question had the potential to end the lives of human beings, it might opt to eliminate the elderly, sick or disabled - so as to increase the statistical average of quality of life within a given population. The system has technically fulfilled the literal conditions of its objective - but it’s unlikely that the person inputting that objective intended for its outcome to be quite so macabre.
Mandatory Resources 🦠
The Executive Summary of an NTI Report on the Convergence of AI & the Life Sciences
A Belfer Centre Analysis of the Risks to Biosecurity Presented by AI
A BBC Article Discussing the Future Role of AI in Pandemic Prevention
A Vox Article Introducing the Concept of Synthetic Biology
A Paper Discussing the Potential Dual Use of Artificial Intelligence Powered Drug Discovery
Optional Resources 🔍
🌟 OpenAI’s Landmark Study: The Impact of AI on the Development of Bioweapons