Trust & Safety Forum: the three fundamentals for building a safer digital world
The fifth edition of the Trust & Safety Forum is by far the richest to date. No fewer than 700 participants, 57 speakers, 15 sessions and 3 workshops, 2 days of discussions including 1 closed-door day devoted to young people’s voices in the regulation of their digital uses. Doing justice to the richness of the debates in 1,200 characters is a dilemma: should one produce an AI-style summary, to convey the content in the most condensed and exhaustive way possible? It is not certain that this summarising function, in which artificial intelligence excels (provided that a fine critical analysis of the result produced is carried out), is the most relevant here.
Let us opt instead for an intuitive approach, more interesting because the vehicle has not yet taken hold of it. That is the whole strength of intuition: intangible, indefinable, it is the product of objective data and subjective perceptions, it is the fruit of experience, but also of the individual from the earliest age, it is within us, and no one (nor anything) has access to its process of elaboration.
Our intuition, fed by the discussions held during and after the event, as well as by the figures from Stanford University’s AI Index 2026, is as follows: the collective sens of vertigo in the face of the power of AI — both a tool of disruption and a lever for protecting our digital lives — seems to have reached its high point.
Listening to the speakers at the Trust & Safety Forum, the fundamental elements that will enable our societies to absorb the shock of this AI revolution have already been identified. These fundamentals are not clearly expressed as such; they may even seem out of reach, yet they are already there in the background. Whether we are actors from civil society, platforms, regulators, providers of technological solutions, each of us is dazzled by AI as one can be by the powerful headlights of a mysterious vehicle in the night heading straight towards us.
The dazzling AI hides the human at work
We are dazzled by the perceived power of the vehicle, which seems to act in place of the human. Because the vehicle does not show us the role of the human, just as headlights in the night do not show us their driver. Akash Pugalia, Digital Chief Officer at TP (formerly TelePerformance), a global Trust & Safety player, insists on this point: “What is often presented as ‘automated’ moderation is, in practice, a multi-layered system combining AI models, policies and human expertise, which work together at scale.” This complexity is a source of hope, insofar as one can progressively understand the different components of this complexity, one by one. There is not a monolithic block imposing itself on us; there is an ecosystem of elements that can gradually be isolated, identified and dealt with. The matter is not simple, and promises do not come without disappointment. Josephine Ballon, co-founder of the trusted flagger HateAid in Germany, one of the largest in Europe against discrimination and online sexual violence, regrets that the DSA, which was designed to bring transparency to moderation decisions by platforms, has not yet borne fruit, due to the lack of transparency on the algorithms themselves, and also due to the unintuitive reporting procedures put in place by platforms. Flore Bouhey Dwan, Director of Platform Supervision at Coimisiún na Meán, the Irish regulator, acknowledges that the system in place is not perfect, but invokes an element that may appear dilatory, but is on the contrary a fundamental element: time.
Time is needed
Time is needed. Time that is lacking in the face of a dazzling and appetising technological proposition, which two experts from the workshop on youth participation, moderated by researcher Ioanna Noula, compared (without having consulted each other) to the powerful fast-food industry: affordable, easy to use, creating a false sense of satisfaction, and resulting in collective (inf)obesity, affecting all ages.
The regulator faced with such a massive phenomenon needs time, as we have seen.
Trusted flaggers also need time: Robbert Hoving, president of the global INHOPE network, warns about the announced withdrawal of European funding granted for 25 years to many trusted flaggers across the EU, seriously destabilising 14 of the structures working to remove child sexual abuse content. How can they be asked to build independence from the major platforms while their public funding is withdrawn?
Research needs time. It is almost a cry of alarm that Andras Molnar, Director of Online Safety at the TUM Think Tank, from the University of Munich, is issuing. Under the pressure of events, can we still take the time to research and to go into nuance? Yet, for Molnar, no actor can fight online harms alone: cross-sectoral, cross-border and interdisciplinary collaboration is a structural necessity, not a mere option. Except that innovation focuses on rapid deployment to the market, while regulation, for its part, rests on the quality of factual information and institutional sustainability. The same observation is made by Julien Bellaiche, research director at the Global Network on Extremism and Technology (GNET): research is nuance and time.
Companies need time. To make things no simpler, technology is itself a source of compartmentalisation. As Tariq Krim, head of the Cybernetica think tank, notes, the technological solutions used in companies have contributed to the rigidification of silos between teams, each using a solution adapted to its needs… and isolating them in their own professional universe. Faced with this ossification of organisations, he calls for the creation of a Chief Resiliency Officer, who “would work transversally to adapt strategy to a world where geopolitics, cyber, politics or sometimes its absence require decisions to be made outside the usual scope”. To the Chief Risk Officer, formal audits, risk registers by unit, the vertical approach; to the Chief Resilience Officer, informal check-ins, the identification of underlying trends, the transversal approach. In other words, organisations, which are based on a pyramidal and vertical model, will need, in order to test and sustain such a horizontal role, … time.
Quality data is needed
Under the pressure of events, a world traditionally as closed as that of judicial authorities and police services has opened up in recent years. Régis Villette, head of the CNAIP (National Centre for the Analysis of Child Sexual Abuse Images) at the Ministry of the Interior, says so spontaneously, and private experts on extreme violent content such as Bjørn Ihler, founder of Revontulet and actively working with Europol and regulators in New Zealand or Canada, confirm it. The explanation is to be found in the need for reliable data. Whether for investigations, OSINT, AI, quality data is the backbone of everything. Akash Pugalia and Tariq Krim, during their conversation on AI, converge on this observation. For Pugalia, “data is the oil of AI”. For Krim, “if you have incredible datasets, you will thrive with AI”. The latter adds, speaking about agentic AI, that one does not need to be an engineer to master a tool like Claude Code, but it is necessary to “know the context and know how to describe what one wants to do”. In this sentence appears the third fundamental element: context.
Context is needed
Context can be defined from factual data, and AI can absolutely restore part of this context. In the field of image-based sexual violence, such as sextortion and deepfakes, much information is publicly available, making it possible to know whether these forms of violence are taken into account in a given country (public discussions, law, legal action, statistics, etc.). For many professionals engaged against these forms of violence, such as Josephine Ballon or Laura-Blu Mauss from Respect Zone, it is incomprehensible that nudification applications are available on the largest app stores. If the author of a nudification may find it amusing to nudify a woman, that is to gravely deny the sense of the people represented in the image. Depending on the context, the same image may be harmless or cause serious harm. Arda Gerkens, the president of the Dutch regulator, acknowledges that it is difficult to understand the availability of these applications, but that the current regulatory framework makes it difficult to block them in the absence of a law or an administrative ban… which is coming soon in Europe, as early as this year. Gabi Walshaw, VP Product at Qoria, sees this difficulty between data and context in the field of child protection within schools: she constantly observes a gap in understanding between what Qoria measures and what professionals on the ground feel. Let us give the floor to Andras Molnar, who summarised the challenge and the opportunity before us: “Trust & Safety is a systemic challenge: it is no longer about isolated tools or policies, but about how AI, humans and governance frameworks work together coherently.” With time, quality data and contextual information, there is no reason to doubt that we will get there. In any case, that is our intuition.