• As artificial intelligence (AI) technologies evolve, balancing innovation with safety has become essential.
• Trust and safety professionals working with members of the AI community will be key to addressing the complexities of online and AI safety.
• Through open discussions and implementing robust safety measures, we can ensure AI technologies are harnessed responsibly for the benefit of all.

As artificial intelligence (AI) technologies, particularly large language models, rapidly evolve, balancing innovation with safety has become crucial.

Trust and safety (T&S) professionals have long been dedicated to protecting online communities and platforms from various harms, and the intersections between T&S experts and the AI community are therefore key to addressing the complexities of online and AI safety.

To achieve this balance, open discussions on trust and safety, especially regarding harmful content, are essential.

The growth of the digital trust and safety community

In the last 10 years, the T&S community has made great progress, gaining valuable practical experience and knowledge. This advancement has been closely connected to the rise of social media, which has rapidly expanded the digital landscape and brought new challenges for managing online safety and user data.

The recent third edition of the TrustCon24 Conference well acknowledged this growth. The event, held in San Francisco from July 22-24, was organized by the Trust and Safety Professional Association and aimed to foster a global community of practice among T&S professionals.

Throughout the conference, the potential for AI systems to generate or amplify harmful content was widely discussed. For instance, AI technologies can be misused to create realistic fake media, which can contribute to misinformation, defamation, and exploitation.

Moreover, AI can facilitate cyberbullying and harassment by, for example, using automated bots to send abusive messages at scale and creating deepfakes or synthetic media for defamation.

The proliferation of AI-generated child sexual abuse material (CSAM) was identified as another significant challenge. The advanced quality and photorealism of such material complicate law enforcement efforts to assess and address risks effectively.

With more than 15 billion AI-generated images circulating only in 2023 and growing, the scale of this issue is daunting. The Internet Watch Foundation, for example, identified 20,254 AI-generated images posted to a dark web CSAM site within a single month, highlighting the urgent need for action.

Bridging the gap between AI safety and T&S

While the challenges of artificial intelligence safety required strong collaboration, the T&S community and the AI and machine learning (ML) communities have developed somewhat independently.

The TrustCon24 session ‘Trust & Safety x Responsible AI: Bridging the Gap’ focused on this separation. Dave Willner, a Fellow at the Stanford Cyber Policy Center, highlighted how connections between the communities have grown over the last years and months, but that there are still significant misalignment and gaps in safety practices.

Meanwhile, Humane Intelligence CEO Rumman Chowdhury referred not only to the amount of research around AI and AI safety, but also to the still minor practical experience in how artificial intelligence products could be exploited to create harm.

For example, AI teams working on datasets might not be familiar with established T&S protocols for handling illegal content. Similarly, ML engineering teams may lack experience with policy-making and regulatory procedures. This disconnect impedes the effective implementation of safety measures and the development of cohesive strategies to address AI-related issues.

• Common language: Effective communication between the AI and T&S communities is essential, as miscommunications have sometimes led to misunderstandings. Developing a unified terminology is crucial for ensuring that both communities can work together effectively. The Typology of Online Harms by the Global Coalition for Digital Safety is an attempt to build foundational language around online harms.
• Risk assessments: Digital safety regulations, such as the previously mentioned EU DSA and the UK Online Safety Act, have begun requesting risk assessments to actively manage potential harms. However, gaps remain, particularly in the nuanced understanding of AI-specific risks. The EU AI Act aims to fill these gaps by setting comprehensive standards for artificial intelligence systems, ensuring they are developed and deployed safely. T&S professionals’ expertise in risk assessment can inform comprehensive evaluation frameworks for AI systems, providing invaluable skills in identifying threats and understanding human-AI interactions to integrate safety considerations throughout AI development.
• Safety by design: The concept of ‘safety by design’ involves integrating safety measures into every stage of product and service development. This proactive approach, used by T&S professionals, can be applied to AI systems to anticipate and mitigate risks early. In practice, as discussed during the above-mentioned panel, this could mean that an AI system declines to perform a task that could affect children.
• Measuring safety: Accurately measuring AI safety remains a critical challenge. Despite the inherent complexities, striving for effective measurement is essential. Existing efforts like the National Institute for Standards and Technology (NIST) AI Risk Management Framework play a crucial role in measuring and managing AI risks. In addition, the Global Coalition for Digital Safety’s report How to Measure Digital Safety Effectively to Reduce Risks Online provides a categorization of metrics. For example, the category of process metrics covers the approach, implementation, and outcomes of systems relating to digital safety, which can also inform AI safety assessments and foster a common understanding among experts.