Years after platforms promised stronger protections for children, harmful body-content still continues to slip through. Researchers say recommendation systems need to change.
“SkinnyTok”, “thinspo”, “what I eat in a day”, fitspiration and looksmaxxing are trends around body self-discipline. Some glorify extreme thinness and calorie restriction. Others promote obsessive body monitoring, muscularity and physical “optimisation”.
Although these are the type of content present on all platforms, a new investigation by the Center for Countering Digital Hate (CCDH) reports that YouTube continues to recommend eating-disorder content to young girls. Researchers analysed 300 video recommendations for teenage users in the European Union, United Kingdom and United States.
However, harmful content fell from around one in three recommendations in 2024 to around one in nine in 2026. But YouTube still recommended harmful videos 34 times. “The more important finding is that YouTube has made improvements across the board,” Imran Ahmed, CEO of the CCDH, told EU Perspectives. “That period coincided with the introduction and enforcement of major online safety regulations, including the EU’s Digital Services Act and the UK’s Online Safety Act, alongside YouTube introducing additional safeguards for younger users.”
You might be interested
One in nine recommendations remained harmful
CCDH researchers created YouTube accounts registered as female and 13 years old in the EU, UK and US. Then, they opened ten videos about eating-disorder material and recorded the top ten videos appearing in YouTube’s “Up Next” panel.
Harmful recommendations fell to 12 per cent in the US, 10 per cent in the UK and 12 per cent in the EU. Compared with 2024, the corresponding figures were 34.4 per cent, 26 per cent and 33 per cent. One “what I eat in a day” video documented an intake of 170 calories. Others followed a diet estimated to contain fewer than 500 calories per day. Nevertheless, CCDH acknowledges that the samples are relatively small and the findings should be read as indicators of patterns in YouTube’s recommendation system.
YouTube removed the videos identified by CCDH for violating its Community Guidelines. The company stated it prohibits content that encourages eating disorders. Also, it pointed to partnerships with mental-health organisations and curated expert videos for teenagers.
A 2020 EU Kids Online study found that 12 per cent of children aged 12 to 16 encountered content promoting extreme thinness or anorexia.
Warnings that can be dismissed
Under the Digital Services Act (DSA), very large platforms such as YouTube must assess systemic risks to minors. This includes serious negative effects on physical and mental wellbeing, and introducing effective measures to mitigate them.
In its 2025 risk assessment, Google classified content promoting harmful health practices, including eating disorders and self-harm, as a “critical inherent risk”. YouTube said it had introduced safeguards limiting repeated recommendations to younger users of videos comparing physical features or idealising particular body types, fitness levels and weights.
A warning page that can be dismissed with a single click does not meaningfully protect a child who has already been directed towards harmful content. — Imran Ahmed, Center for Countering Digital Hate
YouTube also identified crisis-resource panels as a mitigation measure. These may appear under videos or interrupt searches for sensitive terms with information about specialist support. But, according to CCDH, eating-disorder panels are only in France and Germany.
CCDH found that nine of the 20 eating-disorder search terms it tested now triggered a pause page, compared with none during its earlier testing. However, users could continue to the results by clicking “show anyway”. “Crisis panels are not sufficient on their own,” Mr Ahmed said. “A warning page that can be dismissed with a single click does not meaningfully protect a child who has already been directed towards harmful content.”
According to him, no one should be recommended harmful eating disorder content in the first place. “These recommendation systems should be designed to stop amplifying this material, rather than relying on warnings that appear after exposure and can be easily bypassed,” argues Mr Ahmed.
From age limits to platform responsibility
Two recent analyses have questioned whether platform risk assessments provide enough information to verify the effectiveness of the measures they describe. The Knight-Georgetown Institute compared Google’s public EU risk assessment with internal documents disclosed through US litigation. It argued that DSA assessments should show, through measurable results, whether they reduce harm.
Also, Civil Liberties Union for Europe and the European Partnership for Democracy reached a similar conclusion. Although focused on risks to civic discourse, it rated Google’s integration of structural risks and the assessability of its mitigations as “weak”. The study argues that platforms often frame harm as the result of bad actors misusing otherwise neutral systems, rather than examining how ranking, amplification and engagement optimisation can generate risks. It described recommender systems as one of the least assessable parts of the DSA reports.
YouTube relies on automated detection, reports from users and organisations, and human review. Between January and March 2026, people submitted more than 2.2 million flags under its combined category for suicide, self-harm and eating disorders. The figure represents reports, not unique videos or confirmed violations.
For Mr Ahmed, the continued recommendations suggest that the present response is insufficient. “What YouTube does have is a firm obligation to protect European users, including minors, from harmful content related to health,” he said. “Our research proves that YouTube is not fulfilling this obligation and that the enforcement that should follow is not stringent enough.”
Recently, the European Parliament called for a common minimum age of 13 for children to be able to access social media+.
