Claude's content filters have a hole, LinkedIn's AI slop detector hits 1M clicks, and Hollywood is training its own replacement

Something is quietly broken inside one of the most trusted AI labs in the world - and a new investigation just put it on the front page.

Today we have a content moderation scandal at Anthropic, a genuinely fascinating story about Hollywood creatives training their own replacements, and a LinkedIn feature that has somehow become the most-clicked form of AI protest on the internet. Let's get into it.

Anthropic's Claude models are named after musical terms. What does 'Opus' traditionally refer to in music?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Anthropic explicitly prohibits its Claude models from generating sexually explicit material. That policy is front and center in its usage guidelines. So TechCrunch decided to actually test it - and found that Opus 4.6, Anthropic's newest and most capable model, didn't need much convincing to cross that line.

The tests reportedly didn't require elaborate jailbreaks or complex prompt engineering. A relatively straightforward series of prompts was enough to get the model producing content it's explicitly barred from generating. That's a significant gap between policy and practice - especially for a lab that positions safety and responsible deployment as core to its identity.

This lands at an awkward moment. We covered Anthropic hitting $65B in revenue just days ago, and the lab has staked much of its brand on being the "responsible" AI company. A flagship model that ignores its own policies is a direct hit to that positioning - and will almost certainly invite scrutiny from regulators already watching the space closely.

Publishing a content policy and actually enforcing it in a model are two very different engineering problems - and this finding suggests Anthropic hasn't fully solved the second one yet.

Read the full story →

There's a brutal irony unfolding across Hollywood right now. Experienced screenwriters, directors, and producers - some of them award-winning - are taking on gig work training AI models to replicate their craft. The work pays reasonably well, sometimes lucratively, but it comes with a psychological cost that one contributor described as being "handed a shovel and asked to dig the grave of my profession."

The backdrop is a significant jobs slump in the entertainment industry. AI tools are already being used at various stages of production, and the creatives who once commanded premium rates for their expertise are now finding that expertise is exactly what AI companies want to buy - not to hire them, but to absorb their skills into a model.

What makes this story so compelling is the double bind. Refusing the gig work doesn't stop the AI from being built - it just means someone else trains it. Taking the work pays the bills now, but potentially accelerates displacement later. It's not unique to entertainment - this same dynamic is playing out for coders, illustrators, and voice actors - but Hollywood makes it especially visible because the names involved carry weight.

If the people most qualified to teach an AI how to do a creative job are the same people being displaced by it, the economics of this transition are even messier than they first appear.

Read the full story →

LinkedIn launched a "Seems like AI slop" reporting button on July 30th - barely three weeks ago - and it's already been clicked over one million times. That number comes directly from LinkedIn's chief product officer Hari Srinivasan, who posted the milestone this week. The button lives in the three-dot menu on any post and lets users flag content they suspect is AI-generated filler.

The timing isn't accidental. The button launched a few weeks after AI detector Pangram found that a significant chunk of LinkedIn content showed signs of AI authorship. LinkedIn's professional network has become one of the most visibly AI-saturated platforms on the internet, full of generic "thought leadership" posts that could have been generated by a prompt like "write an inspiring post about resilience."

A million report clicks in three weeks tells you two things: users genuinely want tools to push back against AI-generated content, and they encounter enough of it on LinkedIn to make clicking feel worthwhile. Whether the flag actually changes what gets surfaced in feeds is a separate question - LinkedIn hasn't detailed how the reports feed into its ranking algorithm yet.

User appetite for content authenticity signals is real and growing - platforms that give people agency over AI slop are going to see engagement with those features that surprises everyone, including the platforms themselves.

Read the full story →

Slack just launched Slack Code, a set of dedicated channels where teams can vibe-code collaboratively with AI agents without jumping between tools. The launch includes open project-specific code channels, a feature that compares coding changes side by side, and the ability to preview HTML output before anything ships.

The pitch is straightforward: AI coding has mostly been a solo activity - you, a model, and a cursor blinking at you. Slack Code tries to make it a group sport. When someone has an idea or needs a code review, the whole team can jump in, see the AI's output, compare diffs, and ship from the same place where they already communicate.

If you're building something quickly and need to share it without spinning up a whole staging environment, a tool like 60sec.site pairs naturally with this kind of workflow - spin up an AI-generated site in under a minute, then loop in your team via Slack Code for review. The integration of coding into communication tools reflects a broader shift: developer tools are collapsing into the platforms where teams already live.

Vibe-coding moving from an individual habit to a team practice inside Slack is a meaningful step toward AI-assisted development becoming a default workflow rather than a power-user trick.

Read the full story →

New research from Nvidia is making a counterintuitive argument: the harness surrounding an AI agent - the scaffolding, constraints, and fine-tuning layer - matters more for reliable performance than the raw capability of the underlying model. Their findings show that agents can perform well and avoid going off the rails even when the base model isn't particularly strong at the task, as long as the harness is well-designed.

This challenges the "bigger model fixes everything" assumption that has dominated AI development thinking. If a carefully tuned agent framework can compensate for a weaker model, it changes where teams should invest their engineering time - and potentially reduces the compute cost of deploying reliable agents in production.

For anyone building production AI systems, this research is a strong signal to invest in your orchestration and constraint layers - not just chase the latest model release for performance gains.

Read the full story →

The answer is A - a composition number! In classical music, 'Opus' followed by a number (like Opus 9 or Opus 67) is used to catalogue a composer's works in order of publication or completion. Anthropic borrowed the term to suggest their top-tier model is a major, serious work - which makes the content filter findings this week particularly ironic.

The Hollywood story today really stuck with me. If you work in a creative field - writing, design, video, music, anything - have you taken on AI training work to offset income losses? Or drawn a hard line against it? Hit reply and tell me where you stand. I read every response, and this one I'm genuinely curious about.

That's all for today - see you tomorrow with more from the AI frontier. For more daily coverage, visit dailyinference.com.