The U.K. government is adopting an overly “narrow” perspective on AI safety, risking its position in the burgeoning AI landscape, according to a report published today.
This document, released by the House of Lords’ Communications and Digital Committee, is the result of several months of evidence gathering, engaging a diverse range of stakeholders, including technology giants, universities, venture capitalists, media representatives, and government officials.
Key findings of the report emphasize the need for the government to shift its focus towards immediate security and societal challenges linked to large language models (LLMs), such as copyright violations and misinformation, rather than fixating on exaggerated apocalyptic scenarios and hypothetical threats.
“The swift evolution of AI large language models could significantly impact society, similar to the internet’s emergence. It’s essential that the government approaches this with care to harness opportunities, especially without letting unfounded fears about distant risks hold it back,” stated Baroness Stowell, chairman of the Communications and Digital Committee. “To leverage opportunities, we must address risks, yet do so with a balanced and practical view to avoid lagging in the upcoming AI gold rush.”
These insights emerge as global conversations about AI intensify, driven by the transformative capabilities of tools like OpenAI’s ChatGPT, which has significantly raised public awareness of LLMs over the past year. This surge of interest elicits both enthusiasm and concern, fostering vibrant discussions on AI governance. For instance, President Biden has recently issued an executive order focusing on establishing standards for AI safety, while the U.K. aims to lead in AI governance through initiatives like the AI Safety Summit, which convened influential political and corporate leaders at Bletchley Park last November.
Simultaneously, a debate is surfacing regarding the extent of necessary regulation for this emerging technology.
Regulatory Capture and Open AI Development
Yann LeCun, Meta’s chief AI scientist, recently joined numerous experts in an open letter advocating for transparency in AI development. This initiative seeks to counterbalance a growing trend among tech companies like OpenAI and Google, which are perceived to be pushing for “regulatory capture” of the AI sector through lobbying efforts that hinder open AI research and development.
“Historical evidence demonstrates that hasty regulations can consolidate power and stifle competition and innovation,” the letter asserted. “Open models foster informed discussions and enhance policy formulation. If our goals are safety, security, and accountability, then openness and transparency must be prioritized.”
This tension underscores the central theme of the House of Lords’ “Large Language Models and Generative AI” report, which recommends that the government explicitly include market competition as a core AI policy goal to prevent regulatory capture by dominant players like OpenAI and Google.
The report highlights the critical issue of “closed” versus “open” ecosystems, concluding that competition dynamics will not only determine the leaders in the AI and LLM space but also influence effective regulatory approaches. According to the committee:
“This involves a contest between organizations that operate within closed systems and those that prioritize open access to foundational technology.”
In its conclusions, the committee considered if the government should take a definitive stance favoring either approach. Ultimately, it determined that a “nuanced and iterative” strategy will be crucial, although the evidence collected reflected the interests of various stakeholders.
For example, while Microsoft and Google expressed general support for open-access technologies, they warned that the security challenges posed by publicly available LLMs necessitate stricter regulations. In its submission, Microsoft noted that “not all parties are well-meaning or properly equipped to tackle the challenges posed by these advanced language models.”
The company emphasized:
“Some entities may weaponize AI instead of using it as a tool, while others may underestimate the safety challenges on the horizon. Immediate action is essential to harness AI for protecting democracy and fundamental rights, expanding access to AI skills for inclusive growth, and advancing sustainability initiatives.”
Regulatory frameworks must also address the risk of intentionally misusing advanced models for harmful purposes, such as exploiting cyber vulnerabilities or developing hazardous materials. Additionally, there are unintentional risks that may arise, like using AI to oversee critical infrastructure without adequate safeguards.
Conversely, open LLMs enhance accessibility, fostering a “virtuous circle” that encourages innovative experimentation and transparency. Irene Solaiman, global policy director at AI platform Hugging Face, stated in her testimony that disclosing information about training data and publishing technical documentation is crucial for effective risk assessment.
“What’s paramount in openness is transparency. At Hugging Face, we have been diligently working on enhancing clarity to ensure researchers, consumers, and regulators can easily understand the various components of our systems. Often, deployment processes are not well-documented, leaving deployers with significant control over release methods without necessary insights into pre-deployment considerations.”
Ian Hogarth, chair of the U.K. government’s newly established AI Safety Institute, pointed out that private companies currently shape the LLM and generative AI landscape while often “marking their own homework” regarding risk assessments. He remarked:
“This presents structural challenges. Relying on companies to evaluate their systems’ safety is untenable going forward. For instance, when OpenAI launched its GPT-4 model, the team conducted extensive safety evaluations and released a GPT-4 system card outlining their findings. Similarly, when DeepMind introduced AlphaFold, it provided insights on potential dual-use applications of the technology.”
This scenario has led to a somewhat unusual dynamic where private firms drive advancements, simultaneously conducting their own safety assessments, which may not be viable given the significant implications of this technology.
The challenge of avoiding or striving for regulatory capture remains central to these discussions. Several leading LLM developers also advocate for regulation, which many perceive as a strategy to impede competitors striving to catch up. The report highlights apprehensions about the lobbying from these industries and the risk of government officials becoming overly dependent on a limited pool of private sector expertise when shaping policies and standards.
Accordingly, the committee suggests implementing “enhanced governance measures within the Department for Science, Innovation and Technology (DSIT) and regulators to reduce the risks of unintentional regulatory capture and groupthink.”
These measures, as recommended in the report, should encompass:
- Applying metrics to assess the impact of new policies and standards on competition.
- Incorporating red teaming, systematic challenges, and external critiques into policy development.
- Providing additional training for officials to boost their technical proficiency.
- Ensuring proposals for technical standards or benchmarks undergo public consultation.
Key Takeaways
However, the report reiterates one of its primary conclusions: the AI safety discourse is overly dominated by a limited narrative focused on catastrophic risks, particularly originating from those who developed these models.
While the report advocates for mandatory safety assessments for “high-risk, high-impact models” — assessments that extend beyond voluntary commitments from a select few companies — it also contends that discussions surrounding existential risks are often overstated. This alarmist rhetoric detracts from the more immediate challenges present with LLMs today.
“It’s highly unlikely that existential threats will emerge within the next three years and even less likely within a decade,” the report concludes. “As our understanding of this technology improves and responsible development practices increase, we anticipate a reduction in existential risk concerns. The government must monitor all potential issues but should not be distracted from seizing opportunities and addressing more pressing risks.”
To capitalize on these “opportunities,” the report acknowledges the need to tackle immediate concerns effectively, including the proliferation of misinformation through text and the creation of audio and visual “deepfakes,” which experts increasingly find challenging to discern. This is especially relevant as the U.K. approaches a general election.
“The National Cyber Security Centre predicts that large language models will almost certainly be utilized to produce fabricated content; that hyper-realistic bots will facilitate the spread of disinformation; and that deepfake tactics will become increasingly sophisticated as the forthcoming national election approaches in January 2025,” it stated.
Additionally, the committee firmly expressed its stance against the use of copyrighted materials for training LLMs, a practice by OpenAI and other major tech companies, who argue that AI training falls under fair use. This has led artists and media organizations, including The New York Times, to initiate legal actions against AI companies exploiting web content for training purposes.
“One aspect of AI disruption that urgently requires attention is the unauthorized use of copyrighted content to train LLMs,” the report asserts. “While these models depend on comprehensive datasets for optimal functionality, they should not exploit all available materials without securing permissions or compensating rights holders appropriately. This is a matter that the government can swiftly address, and it should act on it.”
It’s essential to note that the Lords’ Communications and Digital Committee does not entirely dismiss catastrophic scenarios. In fact, the report recommends that the government’s AI Safety Institute conduct and disclose an “assessment of engineering pathways to catastrophic risk and warning indicators as an immediate priority.”
Moreover, it recognizes a “credible security risk” linked to the rapid proliferation of potent AI models, which can be easily misused or can malfunction. Nevertheless, despite these realizations, the committee believes that an outright ban on such models is inappropriate, given the likelihood that the worst-case scenarios may not materialize and the significant challenges involved in enforcing such a ban.
The report suggests that the government’s AI Safety Institute should develop “new methodologies” for identifying and monitoring models once deployed in practical settings.
“Completely banning them would be excessive and likely ineffective,” the report concludes. “However, coordinated efforts are necessary to oversee and mitigate the cumulative effects.”
Overall, while the report acknowledges that LLMs and the broader AI movement do present real risks, it advocates for a strategy shift that focuses less on “sci-fi end-of-the-world scenarios” and more on the positive impacts AI can potentially offer.
“The government’s approach has veered too far towards a narrow interpretation of AI safety,” the report states. “It must recalibrate, lest it misses out on the opportunities presented by LLMs, falls behind its international counterparts, and becomes overly dependent on overseas tech companies for this critical technology.”
Ranking
