Meta's teen defaults go global, under-16 ban begins and why takedowns don't always work

Hello and welcome to Everything in Moderation's Week in Review, your need-to-know news and analysis about platform policy, content moderation and internet regulation. It's written by me, Ben Whitelaw and supported by paid members like you.

At school, if you wanted to get good test scores, you'd get a friend to mark your test for you or, if you were particularly cunning, mark it yourself. While both approaches can deliver results, no-one wins in the long run.

Today's newsletter has a whiff of that. Governments are citing evidence to justify social media bans, platforms validating their own safety features and AI companies setting a course that no-one asked for. Prove your solution works, win the argument, hope everyone moves on.

Talking of scrutiny, welcome to new subscribers joining EiM from Germany, Pakistan, Singapore, Italy, the US, India, the Netherlands, Morocco and elsewhere. EiM is a globally focused newsletter with a global audience. Share your feedback about today's edition or send me stories you think I've missed or that merit inclusion.

Here's your Week in Review — BW

Policies

New and emerging internet policy and online speech regulation

Malaysia became the fourth country globally and the second in southeast Asia (EiM #336) to implement an under-16s social media ban, as measures began rolling out on Monday (June 1). The ban affects platforms with more than 8 million Malaysian users — including Facebook, Instagram, TikTok and YouTube — with users having a month to download their data. CNA has previously reported that a host of other countries in the region, including Singapore, the Philippines and Thailand, are likely to follow suit with a ban.

Money mismatch: Non-compliant platforms could face sanctions of RM10m ($2.5m), far less than the ceiling imposed by EU and UK regulators. For context, YouTube — one of the companies affected by the new law — made an average of $126m A DAY from global ad sales in the last three months of 2026. The fine is basically a rounding error.  

One story that I missed during last week’s UK heatwave: Temu has been hit with a largest fine to date under the Digital Services Act for failing to check its service for illegal products. The €200m fine — which makes the €120m hit that Elon took look modest — is below the 6% annual global revenue, or €2.8bn, that it could have been enforced. Temu has three months to address the issues identified by the European Commission. Politico has the write-up.

Also in this section...

• Europe's tightwalk on tech sovereignty (Financial Times)
• A Parent's Guide to Digital Safety (Stanford)
• EU moderation watchdog says social media giants hate taking down hate speech (The Register)

T&S Insider readers had questions; I answered

Debunking myths, working with AI — and working against it!

Everything in Moderation*Alice Hunsberger

Products

Features, functionality and technology shaping online speech

A new piece in the Financial Times captures a frankly unsettling standoff among AI experts: those who believe in "recursive self-improvement" — AI models capable of building better versions of themselves  — and those who say that such systems could go rogue without human input. The prize at stake? Just the small matter of artificial superintelligence. 

Boom and gloom: Many EiM readers have come across the concept of the AI triad or at least people who subscribe to the idea of “AI accelerationism” (aka AI boomers). But some of the quotes from former DeepMind researchers and VCs veer well beyond the usual tech-optimism. Coming just a week or so after Pope Leo XIV called for more responsible and aligned AI innovation (EiM #340) makes it even more jarring. 

Also in this section...

• Why we're partnering with Cinder to power the world's most trusted AI video platform (Synthesia)
• ROOST and Mila Announce Partnership to Advance Open Source AI Safety, Starting with Youth Safety (Mila)

Platforms

Social networks and the application of content guidelines

Meta has announced that the Teen Default experience that was piloted on Instagram last October (EiM #312) will be rolled out globally as well as on Facebook and Messenger, potentially changing the online experience of millions of 13-18 year-olds. But it’s the justification behind the expansion of the movie-rating-inspired system that I’m really fascinated by.

Halfway down the announcement, Meta explains the “external assessment” that was carried out by Alice (formerly ActiveFence) to validate the impact of last year’s changes. Mike and I talk at length about the methodology in this week’s Ctrl-Alt-Speech (out now for Patreon subscribers, dropping later today in other podcast apps) but essentially researchers created 36 accounts to assess the prevalence of mature content thrown up by Instagram’s Feed, Reels, Explore page and Search and an ‘unnamed competitor’.

The results — 68% less mature content than the competitor for the new default, 96% less mature content than the tighter Limited Experience mode — unsurprisingly paint Meta’s change in a favourable light. There’s lots more to say about this so I’ll be taking over T&S Insider on Monday to share more thoughts.

Florida this week became the first US state to file a lawsuit against OpenAI for what attorney-general James Uthmeieir called “hurting our kids” and aiding mass shooters. It joins Kentucky and Pennsylvania — both suing Character.ai — who have brought action against AI companies. Florida is also pursuing a criminal case against the AI lab for its part in the murder of two people during a campus shooting in April last year.

Also in this section...

• How to Fight AI Brain Rot at School? For One Country, It’s With Free ChatGPT (WSJ) 
• Covalen workers protest at Meta's HQ in Dublin (RTE)

People

Those impacting the future of online safety and moderation

Imagine you’re moderating an online space overrun with conspiracy theories and misinformation. Where do you start? Tackle the most farfetched and implausible contributions first in the hope that they reach fewer people? Or focus on the more plausible but still unevidenced ones?

The challenge of what content to prioritise removing is tackled by Madeline Jalbert (University of Washington) and Morgan Wack (University of Zurich) in an early manuscript due to be published in the Communications Psychology journal. In the study, they found that removing posts containing plausible evidence significantly decreased participants conspiracy beliefs and then asked participants to act as moderators for those posts and to prioritise what to remove.

Their results are fascinating: participants removed posts containing less plausible evidence (I have the Pizzagate conspiracy in mind) at a higher rate than more plausible posts, even though such an approach is not conducive to dampening belief in the conspiracy theory. Not only that but when given the explicit goal of reducing belief in the conspiracy theory, they remove both implausible and plausible at the same rate.

This finding has implications for community-moderated spaces like Reddit, user-led initiatives like X/Twitter’s Community Notes (or at least the bots that write notes nowadays) and how platforms educate users to manage self-governed spaces. It may be that users' intuitions about content removal aren’t necessarily helping.

Posts of note

Handpicked posts that caught my eye this week

• “You'll leave with a tested policy, sharper instincts on precision-recall tradeoffs, and a shared vocabulary for harms the field is still naming.” - Policy folks might be interested in this upcoming workshop with Shubhi Mathur. 
• “I gave that talk because I believed the people closest to the problem needed to be the ones building the solutions. Not consulting on them. Not advising from the outside. Actually building them.” - Alexis Crews on why she’s joining this newly-funded startup.
• “But which risks are most severe? Who is most vulnerable? And who is most responsible for addressing them? To answer these questions, we conducted a three-round expert consultation with 272 AI experts.” - Peter Slattery, from the MIT AI Risk Initiative, with a meaty looking paper.