Regulators finally act, AI models shift and platforms scramble

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 members like you.

What a frantic 24 hours. This time yesterday, none of us had any inkling about the H1-B visa bombshell for T&S workers (see Posts of the week). Then, just before I hit schedule on today's newsletters, the European Commission announced a chunky fine for X/Twitter for a breach of the Digital Services Act. Phew.

EiM tries to make sense of all this in real time, bringing you the latest developments, the broader context and the implications for the people actually doing this work. If you find that valuable, consider becoming a member. Every contribution keeps EiM independent and ensures Week in Review or T&S Insider doesn’t disappear behind a full paywall.

Welcome to new subscribers from Telus Digital, eSafety Commision, Valiant TMS, HWCPA, Preply, Product Madness, Deloitte and elsewhere. This is your Week in Review — BW

Policies

New and emerging internet policy and online speech regulation

Ofcom this week issued a £1 million fine to a pornography company for failing to implement adequate age checks, in what will be seen as a win of the Online Safety Act’s enforcement muscle. AVS Group Ltd, which runs 18 adult sites, will face a further daily £1000 penalty unless it fails to implement what the regulator calls “highly effective age assurance” by Monday.

The plot thickens: However, a BBC report suggests AVS Group Ltd may be part of a complex web of companies, some of which appeared in the 2016 global investigation known as the Panama Papers. The concern, according to the BBC’s technology editor, is AVS may not even know it has been fined.

Australia’s online safety regulators has said it will be “watching the migratory patterns” of teen users to smaller platforms ahead of next Wednesday’s under-16 social media ban deadline (EiM #273). According to Crikey, some of the 10 big platforms aren't happy that fast-growing apps like RedNote, Lemon8, Yubo and Bluesky aren't covered by the ban. But eSafety commissioner Julie Inman Grant said it had been in touch with them to make them aware of their responsibilities.

Better the devil you know? As I highlighted in this week’s Ctrl-Alt-Speech, some of these alternative social platforms have less than basic T&S provisions in their terms and conditions. In one case, users are asked to email a generic email address to report impersonation. Can’t see that going well.

Also in this section...

• TikTok and LinkedIn face illegal content reporting probe in Ireland (Euractiv) 
• Considering Trust and Safety's Past, Present, and Future (Tech Policy Press)
• The Global Retreat from Content Moderation Is Endangering Free Expression: Kenya Shows Why (Just Security) 

What if the entry-level T&S apocalypse hasn’t happened? (yet)

The narrative says that junior T&S roles have disappeared. In reality, some still exist — but they’ve concentrated in companies far from the US platforms that everyone watches.

Everything in Moderation*Alice Hunsberger

Products

Features, functionality and technology shaping online speech

If you’ve been the foundational model ‘space race’, you’ll have seen how Google’s Gemini 3 Pro has put the cat among the pigeons. This week, new data saw it it also rank highly among users when measured for trust, ethics and safety, according to VentureBeat. The testing — conducted using a methodology built by University of Oxford researchers called HUMAINE — found the model's “breadth of knowledge and the flexibility of the model across a range of different use cases” made it popular among users.

Model behaviour: Accurate models are one thing but that doesn’t always translate into user trust. Research shows users often rely on clarity, consistency and perceived intent as much as technical performance. So trust — in all of its messy, context-dependent forms — could become an important metric for AI models.

Also in this section...

• The VPN panic is only getting started (The Verge) 

Platforms

Social networks and the application of content guidelines

AI-generated anti-immigrant narratives on TikTok have amassed 4.5 views in a month , according to The Guardian, in what is the latest in a string of generative AI platform incidents (EiM #307). The research, conducted by a Paris-based non-profit, also found that less than 2% of posts carried an AI label, despite the platform joining two industry-wide content authenticity initiatives back in May 2024. Looks like that hasn’t yielded a lot.

A BBC analysis of one weekend of football fixtures found thousands of racist slurs, rape threats and death threats targeting players across Instagram, Facebook and X/Twitter. More than 2,000 posts analysed by data science company Signify were believed to break platform policy but left up by platforms. 39 posts were so serious that they warranted the involvement of football clubs or law enforcement.

Champions of abuse: Abusive posts directed at players is on the rise, according to research, and has led to sports teams increasingly enlisting T&S vendors and social media monitoring tools to filter and hide abusive content on platforms. Most fans won’t know — and I wonder what they’d think about it if they did.

Also in this section...

• ChatGPT-5 offers dangerous advice to mentally ill people, psychologists warn (The Guardian)  
• Valve accused of pulling LGBTQ+ game from Steam at request of Russian  government (Games Industry)  

Stuck in the Middleware with Youth - Ctrl-Alt-Speech

In this week’s roundup of the latest news in online speech, content moderation and internet regulation, Ben is joined by Vaishnavi J, former head of youth policy at Meta and founder and principal of Vyanams Strategies, a product advisory firm that…

Buzzsprout

People

Those impacting the future of online safety and moderation

For a long-time, I’ve wondered about the crossover between journalism — which I’ve worked in for 15 years — and online safety. What is it about the latter that has drawn me in, from the first time I moderated comments on a new websites back in the early 2010s? Well, Deep Ganguli might have just put words to it. 

Ganguli was the first member of Anthropic’s nine-person societal impacts team that was founded before Claude — the company’s widely-used foundational model — was launched in 2023. In this excellent Verge profile, he explains how the team work across the 2,000-strong company to identify emerging risks, scrutinise model behaviour and surface issues long before they become public problems.

Ganguli describes the role as telling “inconvenient truths” about AI — not just to Anthropic’s leadership, but to the wider world. In other words, surfacing uncomfortable facts, challenging assumptions and making hidden dynamics legible. A bit like — you guessed it — a journalist.

Posts of note (H1-B visa reaction)

Handpicked posts that caught my eye this week

Reactions to yesterday's story that the US administration will seek to deny H-1B visas to social media fact-checkers and content moderators.

• "Holy shit" - Kate Klonick doesn't milk her reaction.
• "This is what government censorship actually looks like: defunding entire fields, deporting researchers, and quite effectively chilling speech that helps people understand how our information systems are being exploited and calls out powerful people and platforms for their bullshit." - Kate Starbird calls it out for what it is.
• "If you were a tech company employee who moderated content overseas — say, trying to prevent a repeat of the Myanmar genocide! or maybe policing CSAM! Or hate speech illegal in your country! — you are a “censor” under the definitions of the Republicans in power." - Renee DiResta baulks at the illogical nature of it all.