Can content moderation offer a model for AI oversight?

51 minutes ago

7 min read

By Hannah Robbins

August 10, 2026

As AI becomes more deeply embedded in online communication, questions about transparency, accountability and human oversight are receiving greater attention. At the AI for Good Global Summit 2026, experts examined what years of global content moderation could teach policymakers, companies and civil society about governing AI. During the discussion, speakers described how AI can generate images, videos and text at scale, making it more difficult for users to determine where content came from or whether it is authentic. They also examined how AI could help platforms detect harmful material and moderate content across different languages. This combination of risks and opportunities formed the basis of a discussion at the AI for Good Global Summit 2026 on whether lessons from global content moderation could inform the oversight of AI.

The panel brought together Paolo Carozza, a member and co-chair of the Oversight Board; Elonnai Hickok, Managing Director of the Global Network Initiative; Tomas Lamanauskas, Deputy Secretary-General of the International Telecommunication Union (ITU); and Alex Walden, Global Head of Human Rights at Google.

Carozza explained that Meta created the Oversight Board to provide independent oversight of content moderation decisions on its platforms. Its work focuses on accountability and the application of international human rights standards, particularly where freedom of expression intersects with other rights. As AI expands into more areas of human activity, he said the challenge is to determine how institutions can place human rights at the centre of their decisions and translate those principles into effective oversight.

Governance starts inside companies

Walden argued that companies should begin by making a formal commitment to the United Nations Guiding Principles on Business and Human Rights.

Such a commitment, she said, should extend beyond a general statement of support. Senior executives need to understand the company’s human rights responsibilities, while corporate policies should explain how those responsibilities will be addressed. Companies also need teams or individuals responsible for incorporating human rights considerations across their operations.

“That sort of governance is the first step into having accountability inside the company,” Walden said.

Hickok expanded the discussion beyond companies’ internal structures. She argued that stakeholder engagement is central to implementing human rights principles.

Companies need to consult the people and communities affected by their services and policies, particularly in the markets where those services are deployed, she said. These conversations can help companies identify risks and assess whether policies intended to protect freedom of expression or prevent harm are working as expected.

According to Hickok, engagement with civil society, academics and the technical community can also provide evidence that complements companies’ internal research, creating a feedback loop between corporate policies and their effects in practice.

“Companies need to be going and talking to the stakeholders that are going to be impacted by their policies, by their services,” Hickok said.

The limits of global alignment

Lamanauskas placed the discussion within the wider development of technology governance. He said the relationship between technology and human rights remains relatively new, noting that technical discussions rarely addressed human rights directly two decades ago.

Although companies and international institutions are paying greater attention to these questions, he argued that establishing a binding global framework was, under current political conditions, effectively impossible. In practice, national regulations, regional rules and company policies may be applied across borders through global technology platforms.

“Having those actually binding global frameworks is basically impossible,” Lamanauskas said.

This raises a further question: whether those policies reflect universally recognised human rights or the values of the countries and regions in which technology companies are based or regulated.

He pointed to the UN Inter-Agency Working Group on Artificial Intelligence, which brings together around 60 UN agencies, as one forum for coordinating work across the UN system. He also described technical and policy standards as one way to establish common approaches where binding global rules may not be possible.

Carozza said the Oversight Board had deliberately based its work on international human rights standards rather than general ethical principles. He argued that these standards provide a more consistent foundation because they are recognised through international law and treaties across different countries.

AI as a risk and a moderation tool

The discussion then turned to the role of AI in content moderation.

Carozza said AI-generated content was creating new human rights risks on social media. He pointed to synthetic intimate imagery and the manipulation of information during elections and crises as examples. At the same time, he said AI could help platforms moderate content in languages that have previously received limited coverage and incorporate local context into content decisions. Carozza presented AI as both a potential moderation tool and a source of new forms of harmful content.

Walden described the growing volume of synthetic media as a challenge for people trying to determine whether an image or video is authentic. She said there was no single solution, but companies were developing tools that could provide users with more information about the origins of digital content. As one example, Walden described SynthID as an imperceptible marker embedded in content generated by Google products. According to her, the marker can help identify whether an image was created using a Google product.

She also referred to Google’s participation in a cross-industry initiative focused on content provenance and technical standards. Such approaches could provide information about who created a piece of content and how it was produced, she said.

“Again, there are many ways in which technically we should be looking to address this so that users can have more ways of identifying the provenance and context for any content that they’re looking at,” Walden said.

Lamanauskas identified a similar tension in efforts to address child sexual abuse material. AI can contribute to the creation and distribution of such material, he said, but it can also be used to detect it. He described this as a continuing contest between those using the technology to cause harm and those developing systems to prevent it.

Lamanauskas also pointed to ITU’s work with companies and standards bodies on watermarking and deepfake detection. Because the technology continues to change, he said these systems must respond to a moving target. He argued that AI education should go beyond teaching people how to operate particular tools. It should also help them understand how to live safely in an environment increasingly shaped by AI, including systems designed to agree with users and reinforce their existing views.

The limits of automated moderation

Hickok cautioned that using AI for content moderation introduces its own risks. Automated systems must respond to continuously changing forms of harmful content, she said. Hate speech, for example, may be communicated through emojis or language whose meaning depends heavily on context. AI systems may also perform unevenly across different languages.

According to her, companies need consistent policies explaining where and how AI is used in moderation. They must also consider whether an automated system should only flag content for human review or be allowed to remove it directly. She argued that strong appeals procedures are particularly important when automated systems make moderation decisions. Without effective remedies, legitimate content could be removed with no meaningful way for users to challenge the decision.

“I think that points to the importance of having really strong appeal mechanisms when we’re using AI in content moderation and an effective remedy,” Hickok said.

Other questions concern whether automated moderation should be voluntary or required by regulation, which tools should be used and how those tools should be tested. Hickok also raised the need for public transparency and independent oversight of these systems.

Carozza said these unresolved questions were among the reasons the Oversight Board remained committed to involving people in the oversight process. The board functions in part as an appeals body, with users bringing cases for review. He believed every decision made by the board had confirmed the continuing need for human evaluation, even where automated tools could accomplish a great deal.

Carozza then turned to the emerging oversight questions surrounding large language models and chatbots as they become more prominent in social interactions and workplaces. The Oversight Board is therefore considering whether its experience with independent content moderation oversight could be adapted to interactions involving these systems.

Making embedded values visible

Lamanauskas argued that the responses produced by different AI models may reflect embedded political, philosophical and cultural values. Because societies do not necessarily share a single value system, he questioned whether those values could be harmonised globally. Transparency, however, may offer more scope for alignment.

“You probably can only harmonize transparency,” Lamanauskas said.

He suggested that standards could help users understand the values embedded in different models, even if those standards could not determine which values the models should adopt. Greater transparency could then allow people and institutions to make more informed choices about the systems they use.

Walden agreed that standardisation could be useful in areas such as transparency. Some common practices could be developed voluntarily and more quickly than through regulation, she said.

Other areas may require regulation to establish consistent expectations across the industry. Walden pointed to non-consensual intimate imagery and other forms of harmful content as examples where voluntary measures may not be sufficient. She also cautioned against regulations that require companies to use a particular technology. As AI develops, she argued, companies need enough flexibility to revise their methods and respond to new forms of harm.

Assessing risks as technology changes

Hickok identified human rights risk assessments and impact assessments as tools that could help companies, governments and other stakeholders respond flexibly to emerging technologies and understand the challenges associated with their implementation. She pointed to mandatory risk assessments under the European Union’s Digital Services Act as one model. She said further work was needed to determine how mandatory assessments would operate in practice and how they could be made meaningful.

She argued that assessments should form part of a wider system of accountability. Civil society organisations need access to relevant reports and opportunities to provide evidence, while companies should include stakeholder engagement throughout the assessment process.

The discussion did not identify a single institution or regulatory model that could resolve every question raised by AI. Instead, speakers described oversight as a combination of internal company governance, international human rights standards, technical standards, stakeholder engagement, regulation, transparency and access to human review.

In closing, Carozza called for continued dialogue across industry, civil society, oversight institutions and public bodies. He argued that cooperation across these groups would be necessary to account for different value systems and the people affected by AI technologies.

GENOCIDE WATCH