Experts may have welcomed the UK government’s moves to increase its focus on frontier AI at its upcoming AI Safety Summit, but they still have a flurry of questions.

What are these advanced areas of AI, what risks do they pose, and what impact could frontier AI have on the UK’s current pro-innovation stance?

Frontier AI is set to take centre stage at the UK government’s first AI Safety Summit at Blechley Park in 1 and 2 November, where the topic will take three of the five objectives of the summit.

The five objectives of the Summit are:

• Develop a shared understanding of the risks posed by frontier AI and the need for action
• Put forward a process for international collaboration on frontier AI safety, including how best to support national and international frameworks
• Propose appropriate measures which individual organisations should take to improve frontier AI safety
• Identify areas for potential collaboration on AI safety research, including evaluating the potential capabilities of AI and the development of new standards to support governance
• Showcase how ensuring the safe development of AI will enable AI to be used for good globally, for example with new rapid medical and drug advancements.

Frontier AI Explained 

Katie Simmonds, managing associate and AI regulation expert from transatlantic law firm Womble Bond Dickinson explains: “Frontier AI applications are advanced foundation models that power systems such as OpenAI’s GPT-3 and GPT-4 (which underpin the widely known ‘ChatGPT’).

“Whilst these models have enormous potential, including efficiency and innovation benefits, they have potentially dangerous capabilities which should not be underestimated. Such models can pose risks to public safety and global security as by advancing towards human ability, they can be used to exploit vulnerabilities in software systems and spread persuasive disinformation at mass scale.”

Many of these models are generative AI models, which have inherent problems, such as producing unexplainable results, ‘hallucinating’ false outputs, and exacerbating existing biases.

New AI can be unpredictable due the nature of the algorithms, and dangerous capabilities can arise unexpectedly (regardless of intensive testing) and remain undetected until after harm has been caused.

Regulators and technology experts have found it continually difficult to specify how AI models should be applied, and what they should be allowed to do. The struggle to distinguish dangerous and useful outputs without knowing the context of their application can be detrimental to developers trying to advance the technology.

For instance, facial recognition can be useful for biometric security, but has also caused a flurry of human rights concerns if used by law enforcement authorities to track crowds.

Another overriding issue is that frontier models are much more difficult to train than to use, so end up being available to the public who will inevitably use the AI for tasks the developers never expected, which can result in misuse.

Open-source models can also result in dangerous capabilities being introduced at a later stage more easily by third parties, sometimes as an unintended consequence.

Recommended reading

• UK’s AI Safety Summit Ambitions Announced
• Scotland’s Innovation Minister Concerned Over UK AI Regulation Plans
• Scot-Secure West | The AI Genie is Out of the Bottle

The addition of tailored extensions to ChatGPT for instance can potentially introduce cybersecurity vulnerabilities or allow third parties to track the data inputted by users without informed consent.

Current regulation of AI in the UK: a ‘pro-innovation’ approach’

Amy Battinson, trainee solicitor at Womble Bond Dickinson added: “The UK government has so far promoted a pro-innovation approach and opted not to impose legislation in the area, yet there has been recognition that foundation models need more thought and intervention.”

In its White Paper presented on 29 March 2023, the Office for AI proposed a de-centralised and self-regulatory approach in order to encourage innovation.

Amy explained that the theory behind this stance was that regulators would use existing legal frameworks to implement the following five principles, on a non-statutory basis:

1. Safety, security and robustness

2. Appropriate transparency and explainability

3. Fairness

4. Accountability and governance

5. Contestability and redress

Katie said: “There is currently ongoing debate about whether the existing frameworks do enough to mitigate AI based risks, especially when confronted with frontier AI, and how best to balance this and the strive for innovation.”

“Since the White Paper was published, the government has been working with regulators to determine whether the principles are being applied in a proportionate and effective manner, or if statutory intervention is needed.”

“The UK government has recently launched the Frontier AI Taskforce (which was previously known as the Foundation Model Taskforce) to focus on the significant risks posed by these particular systems. This shows that the government is recognising the significant risks and threat posed by misunderstanding or underestimating frontier AI. It also shows the needs to be agile and responsive as we all begin to learn more about what AI can offer, how it can be harnessed for good, and how we can manage potential negative impacts”.