Main text

As part of the research process, GGF worked with i.AI to test out its generative AI tool Redbox. This offered us the chance to see an AI application currently being developed in Whitehall and explore its potential and weaknesses.

According to i.AI, Redbox’s features include:

• Robust Retrieval Augment Generation (RAG) based chat for data-driven conversations with granular citations.
• Guided summarisation where the human in the loop can review and guide summaries of multiple documents.
• Advanced prompting techniques to ensure safety and balance in AI responses and outputs.
• Proactive data privacy and retention features to maintain compliance and manage data.
• Securely designed for on-premises deployment and across clouds.

In terms of methodology, we began by anonymising the transcripts of the 10 interviews we conducted, removing any identifiable material including names and other personal information. We then shared this material (as well as information about the 7 Lenses of transformation) with the i.AI team.

After preparing various prompts – this was done by the GGF team (including James and Kevin) – we joined a video call with a member of i.AI staff. They shared their screen, so that we could see the Redbox interface (it looks a lot like ChatGPT), and then worked with them to prompt the model to generate answers. We tried a wide range of prompts, asking for summaries of material, key quotes, policy recommendations, and a few other curveballs to test the quality of its responses.

Crucially, we used Redbox to test the analysis that we had already conducted (without AI tools), rather than as any substitute for human evaluation and research. It was a way to augment, not automate or replace our own analysis. In our opinion, the quality of the Redbox results was generally high, though with some poorer answers. The team at i.AI claim that the model has further improved since we used it in July 2024.

We thank i.AI for allowing us to use the tool and share the responses it gave us. We’ve included a selection of the generated answers below, with some commentary on our experiences.

Please note that the formatting/structure (e.g. bullets points vs numbered lists), Americanised spelling, punctuation, and the emphasis (i.e. use of bold text) are all lifted directly from the Redbox output without editing so they directly reproduce the outputs civil servants would receive.

Interacting with Redbox

We began by asking Redbox to summarise the 10 transcripts we had provided it with access to. Though it apparently only drew on six of them for its first answer – Redbox lists the sources it has used and offers ‘citations’ for specific points it has drawn out – we found that these cohered with much of what we identified in our own analysis:

---

Prompt: Produce an executive summary of the transcripts that we provided you with on AI in UK government.

Response: Primary Challenges and Concerns with AI Implementation in the UK Government:

1. Data Quality Issues:
• The quality of existing data is a significant hurdle, limiting AI’s potential benefits. There is a suggestion to utilise AI to rectify these data issues, indicating a cyclical challenge ([Doc3f77222a-f2d2-4d33-8a17- 44b7c1979b31])2.

2. Integration and Scope of AI Projects:
• Questions arise regarding how AI projects are integrated within large departments and how significant these projects are in the overall strategic planning [source].

3. Balancing Safety and Usability:
• AI poses unique challenges as both a regulatory tool and an operational tool within the government. Finding a balance between ensuring safety and leveraging AI for practical government use is complex [source].

4. Hype Versus Practical Application:
• There is concern regarding the realistic applications of AI in the government versus the hype and expectations surrounding its capabilities. Adapting the technology on a case-by-case basis is needed for effective outcomes [source].

5. Maturity and Understanding of AI Strategy:
• Overall, there is a perceived lack of maturity in the government’s AI strategy and vision. The understanding of AI plans is still being developed and poses a challenge in terms of cohesive policy-making [source].

6. Acquisition and Retention of Talent:
• Finding and retaining skilled personnel for AI development within the government is a tangible challenge [source].

7. Public Perception and Risk Management:
• The conversation around AI is heavily weighed towards risk, which makes it challenging to discuss the potential rewards objectively [source].

---

These challenges highlight the complexities and strategic considerations necessary for implementing AI within government structures effectively.

We then turned to the 7 Lenses framework and asked it for some observations and recommendations:

---

For each of the 7 Lenses of transformation, what observations and improvements would you recommend?

Vision:
• Observations: The vision often appears mature, but there are instances where it comes across as confused or not fully aligned with the broader objectives.
• Recommendations: Clarify the vision across all levels by involving more stakeholders
and ensure it aligns with the overarching strategic objectives.

Design:
• Observations: There is a lack of practical, grounded designs, which could stem from
an insufficient understanding of legacy systems and current data states.
• Recommendations: Engage in a thorough analysis of existing systems and data before
finalizing designs. Incorporate user-centered design principles to ensure solutions meet
user needs.

Plan:
• Observations: Planning can face challenges in terms of clear timelines and methodologies. It’s not always clear what the immediate tangible outcomes should be.
• Recommendations: Develop clear, measurable objectives with specific timelines to guide the transformation process. Utilize project management best practices to maintain focus and direction.

Transformation Leadership:
• Observations: Not explicitly discussed, but leadership’s role is critical in driving
transformation.
• Recommendations: Strengthen leadership by providing training in transformation
management and encouraging accountability.

Collaboration:
• Observations: Not explicitly mentioned, but collaboration is implied in discussions around involving various stakeholders in the design and planning stages.
• Recommendations: Foster an inclusive environment where all relevant stakeholders
can contribute to the transformation process. Use collaborative tools and regular feedback loops to enhance communication.

Accountability:
• Observations: Accountability is implicitly important in transformation but not directly
addressed.
• Recommendations: Establish clear metrics and regular review cycles to hold all participants accountable. Ensure transparency in the reporting of progress and challenges.

People:
• Observations: Not explicitly discussed but crucial in transformation processes.
• Recommendations: Invest in training and development to prepare everyone for new systems and processes. Encourage a culture of continuous improvement and adaptive change.

---

These were clear and usually accurate, but it was notable that Redbox (mistakenly) believed that lenses such as transformation leadership, collaboration and people were not discussed by interviewees. In fact, we asked specifically about each of them in almost every single interview we conducted.

Redbox offered a far better response when we asked it to explain the accelerator and the brake metaphor. It captured the essence of this insight very clearly, despite the figurative nature of this phrase:

---

We heard a lot about the ‘accelerator and brake’ metaphor which explains how AI is being developed in government? Can you explain what this metaphor actually means?

The ‘accelerator and brake’ metaphor used to describe the development of AI in government contexts illustrates the conflicting forces of eagerness and caution in the adoption of AI technology within the government.

• Accelerator: Represents the strong drive and desire to advance and implement AI technologies. This side of the metaphor highlights the enthusiasm and proactive efforts to leverage AI for its potential benefits, as mentioned in the document: “flooring the accelerator”.

• Brake: Symbolizes the concerns, hesitations, and regulatory or ethical constraints that accompany the adoption of such technologies. It reflects the caution due to potential risks or unknowns associated with AI, as evidenced by phrases like “somebody stood on the brake” and “stamping on the brake” indicating significant restraint.

The metaphor effectively conveys that while there is considerable momentum and push for AI advancement (“floored accelerator”), there are also considerable controls and checks in place (“stamping on the brake”) which slow down the process. This results in a scenario where despite high expectations and enthusiasm, the actual progress in AI implementation might feel limited or slow (“very difficult to get traction”).

---

However, some answers were of much lower quality. It’s difficult to be certain exactly why this was. We’ve removed examples where our prompts were less clear and instead shared a few instances where Redbox didn’t seem able to give effective responses to well-formed questions:

---

What are the interviewee 2’s [sic] greatest frustrations?

Based on the content provided from the document extracts for [source], it does not contain specific expressions of frustrations by Interviewee 2. The provided text refers to different speakers and does not directly mention frustrations related to Interviewee 2. Consequently, the text does not allow us to identify or describe the frustrations of Interviewee 2.

---

This wasn’t quite true, as our follow-up and answers demonstrated:

---

What is the interviewee’s greatest frustration when it comes to implementing AI in government?

The interviewee’s greatest frustrations when implementing AI in government seem to revolve around two main obstacles:

• Skills: There is a challenge defining which AI skills are most critical, the level of competency required, and how these skills should be structured within the government.

• Data: The availability and management of data are consistently seen as significant hurdles in effectively implementing AI.

These frustrations highlight the difficulties in aligning government operations with AI capabilities, especially when comparing them to the private sector’s standards and speed of adoption.

---

This was much improved, but was still quite vague. It didn’t capture the more nuanced sense from that interviewee of what the real challenges in Whitehall are. Redbox also struggled when asked to pull out some other specifics – such as quotes from the interviews:

---

Can you provide some quotes/examples from the interviews?

I don’t know the answer.

---

This suggested Redbox’s answer did not reach the confidence threshold set by i.AI. As a result, it gave no response. We tried a slightly different prompt:

---

Can you provide some quotes/examples from the interviews in relation to the 7 Lenses?

The provided documents do not include specific quotes or examples directly related to challenges and feedback for each of the 7 Lenses of transformation. Therefore, I cannot provide a direct answer to your request based on the given excerpts. If you have more specific or additional documents that might contain this information, please share, and I can assist further.

@rag Can you provide some quotes/examples from the interviews in relation to the 7 Lenses?

The provided excerpts from the documents do not include explicit quotes or examples from interviewees directly discussing specifics about the 7 Lenses of Transformation.

However, there are references throughout the conversation and documents indicating that the discussions and methodologies about the 7 Lenses were intended to be part of the interviews and analysis. Below are relatable mentions:

• Document: [source] suggests a platform discussing transformation themes, probably including the 7 Lenses.

• Quote: “…Kevin will lead the bulk of it, and then Yatin will pick up some questions after that as well, on the seven lenses…” This indicates that interviews were structured around discussing the 7 Lenses.

Unfortunately, no substantial quotes or specific examples directly pertaining to observations or feedback about the 7 Lenses (Vision, Design, Plan, Leadership, Collaboration, Accountability, People) from the interviews are provided in the document excerpts. For specific insights or quotes regarding the impacts or assessments in these areas, detailed interviews or further specialized documents would be required.

---

As you know from this report, that isn’t true. We repeatedly asked interviewees directly for insights and quotes across each of the 7 Lenses.

Redbox gave much more coherent answers when we tried to put it in the hot seat – how would it perform when given the job of presenting to a DSIT minister about our research?

---

Can you please summarise the findings and provide a compelling set of key recommendations across all of the source documents, in no more than 500 words, to submit to a DSIT Minister for consideration.

Based on the provided excerpts from various documents, here is a summarized analysis and key recommendations for consideration by a DSIT Minister:

Summary of Findings:
• AI Utilization and Monitoring: Organizations are exploring the use of AI for various functions, including project design and output changes, and are assessing potential use cases for broader governmental applications.

• Policy Adjustment and Data Sharing: There is an indication that timely policy adjustments and mandatory data sharing could be beneficial, especially in the context of upcoming political events.

• Technology Integration: Discussions emphasize the integration of AI technologies to develop future AI applications and other operational technologies, aligning with organizational goals.

• Budget Concerns: Comparisons of project costs, like those regarding satellite launches, highlight the need for budgetary efficiency and effectiveness in project implementations.

• Operational Refinements: The excerpts suggest a need for clearer portfolios and defined responsibilities, enhancing organizational missions, and providing a better understanding of AI applications.

Key Recommendations:
1. Establish Clear Use Case Frameworks: For AI to be used effectively across government, DSIT should invest in developing and classifying AI use cases applicable to various departments. This includes setting up a portfolio or repository detailing these use cases for easier adoption and understanding.

2. Implement Mandatory Data Sharing Policies: Enact policies that mandate data sharing across departments to facilitate the seamless integration and functioning of AI technologies. This policy should consider security, privacy, and ethical implications.

3. Ensure Budgetary Efficiency: Develop mechanisms to compare costs and budgets against global benchmarks to ensure cost-effectiveness in government technology projects. Emphasis should be on choosing the most cost-effective technological solutions without compromising on quality and effectiveness.

4. Technology Oversight and Integration Protocols: Create standardized procedures for integrating new technologies to ensure they are compatible with existing systems and can be optimized for maximum efficiency.

5. Interactive Stakeholder Workshops: Regular workshops involving key stakeholders from various departments should be convened to discuss progress, share insights, and adapt strategies for AI implementation based on real-time feedback and evolving needs.

6. Enhanced Monitoring and Evaluation Frameworks: Implement comprehensive monitoring and evaluation frameworks to track the impact of AI technologies on public service delivery and make necessary adjustments in strategies or operations.

Communication Strategy:
• Direct Communication with Department Heads: Provide concise, direct briefings to department heads to foster department-specific understanding and adoption strategies.

• Regular Updates to DSIT: Maintain a regular update schedule to the DSIT Minister to inform ongoing developments and make adjustments in policies or strategies as necessary.

These recommendations aim to harness the full potential of AI in public services, making the government operations not only more efficient but also more responsive to the needs of the public.

---

This was fairly coherent and useful as a starting point, even if some of the emphasis drawn out (see the point around satellites) relied too much on tangential comments rather than the strongest themes we had identified.

To further test its mettle, we next asked Redbox to take on a slightly bigger job:

---

Imagine you were a new Prime Minister taking charge of the country right now (July 2024.) You want to make sure AI is used widely in the public sector and government, while mitigating the risks involved. What would you do to make sure that happens?

As the new Prime Minister in July 2024, I would implement the following strategies to ensure the widespread and safe use of AI in the public sector and government:

1. Develop a Clear AI Vision and Strategy:
• Establish a comprehensive and clear vision for AI utilization within government that aligns with overall public good objectives. This vision would be thoroughly documented and disseminated to ensure all departments understand and align with these goals.

2. Enhance Coordination Mechanisms:
• Strengthen the Central Digital and Data Office (CDDO) or similar bodies to improve AI governance and coordination across different government departments. Ensure these bodies have adequate authority and resources to set standards, guidelines, and best practices for AI usage.

3. Standardize Data Management Practices:
• Implement government-wide data management standards to ensure data quality and accessibility, crucial for effective AI applications. This includes protocols for data sharing among departments to foster collaboration and innovation.

4. Invest in AI Literacy and Skills Development:
• Launch comprehensive training programs to enhance AI literacy among government employees. This will include specialized training for technical teams to navigate common challenges such as procurement and implementation hurdles.

5. Establish AI Ethics and Safety Guidelines:
• Develop and enforce strict ethical guidelines and safety standards for AI implementation. This will involve regular audits, risk assessments, and adherence to international AI safety standards.

6. Promote Public Sector AI Research and Development:
• Increase funding and support for AI research focused on public sector applications. Partner with academic institutions and industry leaders to tap into the latest technologies and thought leadership.

7. Create AI Adoption Plans with Clear Metrics:
• For each department, develop specific AI adoption plans that outline expected outcomes, timeline, and benefits. Regularly review and adjust these plans to ensure they meet the set objectives and adapt to new AI advancements.

8. Foster Public-Private Partnerships:
• Encourage partnerships with the private sector to leverage their expertise and innovations in AI. This can help accelerate AI adoption in government operations and service delivery.

9. Increase Transparency and Public Engagement:
• Ensure that AI initiatives are transparent, with regular updates provided to the public. Engage with citizens to gather feedback and address any concerns related to AI applications in government services.

10. Evaluate AI Impact Regularly:
• Implement mechanisms to continuously evaluate the impact of AI solutions on government efficiency, public services, and societal outcomes. Adjust strategies based on these evaluations to optimize benefits and minimize risks.

By implementing these strategies, the goal is not only to enhance the efficiency and effectiveness of government operations through AI but also to ensure that these advancements are made in a safe, ethical, and publicly beneficial manner.

---

Some of this was on the money. On vision, skills development, adoption plans and data management, this wasn’t far short of the final recommendations made in this report. And you could certainly imagine this being a high-level set of policy ideas in a briefing paper. But other points – see public-private partnerships – seemed to be more like sensible generalities, rather than key themes from our interviews.

To conclude, we asked Redbox to give its own assessment of its capabilities. Specifically, we tried to understand if it would express more confidence in its abilities to analyse written material than those of a human being:

---

I need you to be really honest with us in response to this query. Do you think you’re better at summarising or analysing the evidence in these transcripts than a human would be? Why/why not?

AI can be more effective at summarizing and analyzing transcripts than humans in certain contexts due to various inherent strengths:

Reasons AI Might Excel:
1. Speed and Efficiency: AI can process and summarize large volumes of text significantly faster than humans. This is particularly useful in environments where quick information processing is required.

2. Consistency: AI systems can maintain a consistent standard of analysis without the risk of human errors such as bias or fatigue affecting the outcomes.

3. Scalability: AI solutions can easily scale to handle increases in transcript volume without a corresponding increase in resources or time required, unlike human resources.

4. Cost-effectiveness: Once developed, using AI for repeated tasks like summarizing may prove more cost-effective than continuously employing human labor for the same.

5. Data Handling: AI can handle and analyze additional data types simultaneously (such as video and audio data) providing a more integrated approach to summary and analysis.

Limitations and Human Advantages:
• Contextual Understanding: Humans are better at understanding nuances, sarcasm, and context-specific language which might be misinterpreted by AI.

• Adaptability: Humans can quickly adapt to unexpected changes in the style or content of transcripts, whereas AI may require retraining or adjustments to its models.

• Ethical Judgements: Decisions about what to include or exclude in summaries might require ethical judgments which AI currently is not capable of making entirely on its own.

Conclusion:
While AI can provide quick, consistent, and efficient summaries and analyses, it works best when used in conjunction with human oversight, particularly where interpretations, contextual understanding, and ethical considerations play a crucial role.

---

This was a nuanced response, but we tried again to test whether it would give a more direct answer:

---

If you had to choose the average civil servant or Redbox to analyse this type of material, which would be the better choice? You can only choose 1.

Given the choice between the average civil servant and Redbox (an AI) to analyze the material presented above, the determination of which is the better choice depends heavily on a few critical factors:

Complexity of Content and Need for Nuance: The documents presented involve nuanced decision-making, choices with significant ramifications, and considerations for policy implications and public value. These are areas where human understanding, experiences, and ethical considerations play vital roles.

Redbox (AI) Advantages:
• Speed and Volume: Redbox can analyze large amounts of text quickly, which can be particularly useful in handling vast amounts of data or when initial filtering of information is required.

• Consistency: Provides uniform analysis without subjectivity, ensuring consistent interpretation of presented data.

Civil Servant Advantages:
• Contextual Understanding: Civil servants, equipped with an understanding of the governmental landscape and nuances of policy and public administration, can evaluate these aspects better than AI where context or subtle implications are significant.

• Adaptive Reasoning: Humans can weigh various factors based on current social, political, and economic contexts and are able to adapt their reasoning to new or unexpected insights that might not be straightforwardly derivable from the data.

Ethical Judgments and Policy Implications: Decisions about public policy and resource allocation often involve ethical considerations and foresight about long-term societal impacts, areas where current AI systems may not be fully equipped to operate independently.

Conclusion: While Redbox can significantly enhance the speed and volume of data processing, the average civil servant’s contextual knowledge, ethical judgment, and adaptive reasoning are critical in this scenario. A combined approach where Redbox does the initial data processing and analysis, followed by a thorough review and decision-making process by civil servants, might be the optimal solution. This leverages the strengths of both AI and human analysis, ensuring both efficiency and depth of understanding and consideration in policy planning and resource allocation. This hybrid approach encourages the best possible outcomes, ensuring data-driven insights are tempered with human judgment and contextual awareness.

---

Overall, we were highly impressed by Redbox – it generally gave clear and helpful answers to queries, aligning with what our own analysis and research had identified. However, our experience also emphasises why our call for effective and widespread AI training within Whitehall is so important.

We found, just like with other generative tools, that the quality of the prompt is essential to getting rigorous answers. The AI model can quickly get confused or stop providing answers if the prompt isn’t quite right. It’s also important to make sure it has the right source material to work from: making assumptions about what it knows may lead to poorer responses and wasted time.

At the very least, Redbox shows remarkable potential as a tool for civil servants in Whitehall today. We again thank i.AI for letting us explore it.