Rapidly-advancing capabilities in Large Language Models (LLMs) have caught the attention of researchers, policymakers, and publics, reinvigorating debate about the impact of AI on society. These are complex systems, based on advanced AI methods, and whose impact on society is shaped by how, where, and by whom they are deployed. Ensuring they deliver for all in society will require action across ELISE’s research themes:
Trustworthiness and certification
LLMs generate highly-plausible content. However, this content is often incorrect: it can be fictitious, biased, or out-of-date, in ways that might not be immediately clear to the user. There is a pressing need to identify how each of the EU’s characteristics for trustworthy AI can be embedded in further developments in this field. To enable their wider use, new mechanisms for certification may be needed, creating a proving ground in which users can test the performance of models – and their adherence to regulatory requirements – before implementing them in practice.
Security and privacy
Progress in LLMs has generated a range of security and privacy concerns. It is often not clear what data has been used, or how, in developing the model Queries submitted to LLMs are visible to the organisation providing the model. To avoid such information being misused or made public, emerging user guidance suggests not submitting such information in search queries; further work is needed to understand how data used by foundation models aligns with current regulations. They also influence the cybersecurity environment: injection attacks can undermine the integrity of the system in operation, and LLMs can be used to generate new malware. Wider use of these models also creates new risks for the information environment in which they operate. Their ability to generate misinformation exacerbates issues associated with AI-generated fake content; the result of wider use could be convincing phishing campaigns, for example. In response to these new AI-enabled threats, AI-enabled tools will be needed to enable rapid response to new security concerns.
Explainability, accountability, and decision-making
Concerns have already been raised about the transparency of LLMs and foundation models. These concerns relate to the ability to scrutinise what data the systems have been trained on, how the models work, or what their performance limitations might be when deployed. Technical methods to enhance their interpretability along some axes are in development, but the size of these models and the size of the datasets on which they are trained make the development of such methods highly challenging.
Integration of LLMs into existing systems requires a mix of technical and operational interventions. Practices for scrutinising the strengths and limitations of these models are in flux. Users need to be able to understand the limits of these systems – how they should or should not be used in practice – and be able to work effectively alongside them in deployment. Addressing these issues will require technical strategies for AI integration, for example through new interfaces, alongside efforts to build organisational capability and skills.
Embedding AI ethics
The development of large foundation models highlights the power asymmetries that shape AI development. Progress in recent years has been driven by a small number of companies with large-scale resources. While their use could bring a range of social and economic benefits, these new capabilities require careful stewardship to ensure that they do not also result in harm, for example through misuse of data, misuse in deployment, disruptions to the online information environment, or intensive use of natural resources in development. Human-centric AI methods – based in responsible research and innovation practices – can help direct the development of these models towards more socially beneficial outcomes. Action to demystify and de-hype advances in AI is also needed, to enable publics and policymakers to understand the limits, risks, and potential benefits of further developments in this domain.
Tackling these issues requires advances in core underlying methods to increase the power of AI techniques and to embed human-centric perspectives in their development. Progress under each of the cross-cutting themes described in this document can help steward the development of LLMs and other foundation models towards more socially beneficial outcomes, by combining progress in AI methods with action to ensure these systems are more robust and aligned with the needs of society. Alongside these technical advances, Europe’s wider innovation ecosystem will also influence the extent to which the benefits of these powerful systems are realised in practice. Capabilities in core technologies need to be connected to an environment where organisations have the skills or know-how to adopt new AI tools, and where start-ups and scale-ups can translate this know-how to new products and services. Creating this innovation ecosystem is at the core of ELISE’s work.