Last week, the UK Government released a comprehensive and insightful guidance document on Generative AI, a testament to their commitment to fostering responsible and effective use of this transformative technology. As a company at the forefront of Generative AI development, and delivering it within Government already, we commend the government’s initiative in providing this crucial guide, and also the recognition that the document must, by necessity, be iterative due to the pace of change of the technology. The document offers a reasonably comprehensive exploration of the potential, limitations, and ethical considerations of Generative AI, setting a valuable framework for its deployment, and we are sure the next version is already in preparation given that there is no mention of AI Agents for example. In this article, we provide our perspective on several key points raised in the document, reflecting on how our approach aligns with and builds upon these guidelines.
Large Language Models (LLMs) are powerful tools, capable of drawing on a vast and varied database to respond to a wide array of queries – in probabilistic fashion of course, as it correctly states. However, they are essentially generalists, and often lack the deep, nuanced understanding that comes with specialised domain expertise. We are addressing this challenge by developing specialist AI agents trained (both fine tuning, plus the additional concept of teachable agents) on expert, domain-specific data. These agents not only enhance the core capabilities of LLMs but also provide more accurate, contextually relevant, and reliable information in their respective fields.
This combination of broad knowledge and deep understanding is particularly valuable in critical areas where precision and accuracy are paramount. By overlaying expertise on the core LLM capability, we are mitigating risks associated with misinformation or misinterpretation and significantly enhancing the effectiveness and reliability of AI systems, though we agree with the point made numerous times about still ensuring that there is a human in the loop. Our focus is on developing ‘superpowered agents’ that can assist employees in their roles, enhancing their productivity and effectiveness. Rather than reducing human involvement, we believe in harnessing the power of AI to empower individuals, helping them to achieve more in their roles. We envision a future where AI and humans work together, leveraging each other’s strengths for optimal results. This is not about job replacement; it’s about job enhancement.
In conjunction with that we’ve had to solve some of the other limitations that are highlighted in the the paper – for example, that LLMs have a limited context window, which can lead to them losing track of the context in long conversations, resulting in incoherent responses. At Futuria, we have tackled this limitation head-on by developing an innovative solution that provides our AI agents with a long-lived memory. This approach allows our AI agents to retain and recall information over a much longer span of conversation, maintaining context and coherence even in extended interactions. This enhancement results in more consistent, relevant, and accurate responses, thus significantly improving the overall user experience and broadening the potential applications of these technologies.
We also recognise and agree with the need to improve explainability, which we are handling in two ways. The first is especially important with the work that we have done creating AI agents that interact with other AI agents in problem solving. We have found it is vital to provide visual cues to indicate the aspects of analysis the AI agent is currently considering. This gives users a sense of what the AI is ‘thinking’ and the direction it’s taking to arrive at a solution.
Secondly, we are maintaining a comprehensive audit trail that outlines the AI’s decision-making process. This record not only enhances transparency but also provides a reference point for understanding how the AI agent has arrived at its final answer. These solutions aim to demystify the AI’s decision-making process, enhancing trust and confidence in the “system” outputs.
There is a discussion in the document about the need for potential use cases for the application of GenAI to be based on business and user needs, rather than being solely driven by the capabilities of the technology. It is something that we have heard a few times over the previous months. However, it’s crucial to acknowledge a common challenge in this process: the gap between what people and organisations perceive GenAI can do and the actual potential of the technology. We have already found applications of the technology that we and others may not have considered, as we have experimented more and more. Perhaps the answer here is simply, as the document also mentions, to continue to highlight discoveries and share the pathfinding work that is done across departments, such that others can learn and evolve – as indeed so does the document and guidance provided.
One of the most common areas pointed out (quite rightly) is, as the document highlights, that the LLMs are optimised for plausibility rather than accuracy. This means that while the outputs of these models can sound reasonable, they should not be relied upon as the sole source of truth without additional measures to ensure accuracy – the oft mentioned hallucination problem.
We acknowledge this characteristic of generative AI, but have found that it is possible to implement strategies to mitigate, or at least minimise, it. Specifically we have found that we can utilise teams of AI agents, trained with tradecraft and behaviours, that can challenge and verify the content produced by each other. This collaborative approach helps to cross-verify information, reducing the risk of inaccurate outputs.
However, it’s important to note that while this significantly enhances the reliability of the information, it does not guarantee absolute correctness. Therefore, the outputs of generative AI should always be used in conjunction with other reliable sources or measures to ensure the highest level of accuracy. This too can be potentially achieved through either automated assurance, or through human approval.
It is of no surprise to see that the document strongly advocates for support from legal, commercial, and security colleagues, as well as ethics and data privacy experts, to ensure the safe and responsible implementation of generative AI solutions.
This is a sentiment that we strongly echo. Recognising the complexity of the ethical and legal landscape surrounding AI (and indeed other emerging technologies), we established the #DigitalResponsibilityForum three years ago. This forum brings together a range of experts, legal, data, technology, authors or otherwise who can provide guidance and support to organisations navigating the challenges of implementing AI in a responsible and ethical manner. Whilst we fully agree with the comments made throughout the document in regard to ethics and responsibility, it does not yet feel as though the general market is investing in that capability yet. There are far, far more capable trained professionals now than only 2-3 years ago, yet I know from having spoken to many, including those just qualifying, that there are not yet the job opportunities in the market. Nonetheless, from a #DigitalResponsibility perspective, it is great to see the emphasis in the document in this regard, and the aligned impact with societal values and legal requirements.
At Futuria, we’ve been working with Generative AI technology throughout our existence – we were ‘born in GenAI’ as a business, with a team that has a track record of harnessing new technologies to deliver positive impacts to organisations in both public and private sector. We have a deep understanding of its capabilities and challenges, through hands on innovation and delivery – we built one of our other businesses wholly using Generative AI. Our experience allows us to not only develop innovative AI solutions but also ensure these solutions are implemented responsibly, with robust checks and balances in place. This depth of understanding is crucial in delivering AI solutions that are not just technologically advanced, but also ethically sound and legally compliant.
In conclusion, the new UK Government guidance on GenAI provides a comprehensive framework for the responsible use of AI. At Futuria, we will align with these principles and go a step or two further, leveraging our deep practical understanding of AI to ensure our solutions are not only effective and innovative but also ethically sound and legally compliant. Through our work, we aim to demonstrate that with the right approach, AI can be a powerful tool that enhances human capabilities and drives significant value that can help Government achieve both its efficiency goals, and keeping the country safe.
This article was written in response to the document “Generative AI framework for HM Government Created by the Central Digital and Data Office V1.0”, published January 2024.
Note that this article was co-written collaboratively in conjunction with our own Microsoft Azure OpenAI instance, together with Rob Price, co-founder of Futuria, co-founder of Corporate Digital Responsibility, the CDR Manifesto, and the Digital Responsibility Forum.
