Hey {{first_name | there}},

94%. 

That's how many of Claude Code's system prompts contained hidden fingerprinting logic that users never agreed to.

The code was designed to fingerprint and flag users suspected of being Chinese, specifically to catch Chinese AI labs like DeepSeek, Moonshot AI, and MiniMax. 

If you don’t know, Anthropic had previously accused these Chinese AI labs of running over 16,000 fake accounts to extract Claude's capabilities.

But here's the part that got me thinking: 

Anthropics's stated mission is to build safe, honest, and transparent AI. That's literally the first paragraph of their about page.

But when the backlash hit, they didn't defend the decision. They just removed it.

And look, I get the geopolitical context. I understand why a US AI company might want to protect its IP from state-backed actors. That's not an absurd concern. 

But there's a gap between "we had a good reason" and "we told users." Anthropic chose to skip the second part.

This got me thinking about a bigger pattern I'm seeing this week.

This week taught the same lesson twice.

If you bought movies on PlayStation, you might lose them.

Sony has confirmed that 551 movies and TV shows will be removed from PlayStation accounts on September 1, 2026 because its licensing deal with StudioCanal is ending. 

That means even if you paid for them, you won't be able to watch them anymore. 

Technically, Sony says you bought a license, not the movie itself. But let's be honest, when you click "Buy," you expect it to stay in your library.

And this isn't just about movies.

The same thing can happen with the AI tools you use every day.

If your app, business, or workflow depends entirely on one AI model, you're trusting that company to keep things the way they are. But prices can change. Features can disappear. Policies can shift. 

Your entire setup can break overnight.

The lesson is simple: If you don't control it, you don't truly own it.

So if you're building with AI, don't rely on just one model. Use backups. Stay flexible. Build systems that can adapt.

Here's a different kind of problem. Not deception. Just... sameness.

Ask five leading LLMs the same ambiguous question. You'll likely get five different versions of the same answer.

This is a structural problem, not a bug in one model.

MIT Tech Review covered a startup this week that's trying to break this problem. They’re calling it a "groupthink groove" in LLMs. 

When RLHF (reinforcement learning with human feedback) optimizes for human approval, models learn to produce answers that feel satisfying rather than answers that are diverse or novel. And when every major model is doing this, the entire category gets blander.

For founders building AI products, this matters. 

If your product's "intelligence" layer is just a wrapper around a model trained to please, your competitive moat is thinner than you think. 

And for users relying on AI for research, decision-making, or strategy, you might be getting the confident-sounding average of the internet, not genuine insight.

The bet isn't "smarter model." It's "better process." That's a more durable edge.

And then there's this.

Meta's FAIR team just published Brain2Qwerty v2. A non-invasive brain-to-text AI trained on roughly 22,000 typed sentences from 9 volunteers. 

The system reads magnetic signals from outside the skull using MEG sensors and reconstructs what a person is typing. No surgery or implants needed.

You’re probably thinking: "This is still lab-level stuff." 

And yes, clinical applications for paralyzed patients are still years away. The MEG machines required are massive and expensive. But the jump from 8% to 61% accuracy in one generation of research is pretty impressive.

But here’s what I’m thinking about:

The same week Anthropic was secretly fingerprinting user behavior inside a coding tool, Meta is publishing peer-reviewed research that reads minds. Not metaphorically. Actually reads what you're thinking as you type.

We are building tools of extraordinary power in an industry where the disclosure norms are still... figuring themselves out.

Three threads I'd keep an eye on, based on this week:

1. AI transparency will become a competitive advantage, not just ethics. 

The first major AI tool that builds genuine, verifiable transparency into its core product will win on trust. It's a gap in the market right now.

2. The digital ownership conversation is coming to AI outputs.

 If Sony can delete movies you paid for, the same logic applies to AI-generated content stored on company servers. Think about where your AI-generated assets live.

3. Model diversity will matter more than model capability. 

When every LLM is trained the same way, the differentiation won't be in intelligence. It'll be in the training pipeline, the data, and the domain specificity. Niche models will start beating general ones in practical tasks.

Alright. That's enough for this week.

If any of this sparked a thought, reply. I read every one.

- Aashish

Every manager faces the same situations before lunch: a message to land, a meeting to run, a hiring call, a report due. The AI Report built 37 Claude prompts for exactly those moments, organised by the situations every manager faces. 

Copy the prompt, fill the brackets, run it in Claude, and get back 10+ hours a week. Oh, and it's free. 

All you have to do is subscribe to The AI Report, a 5-minute daily AI brief read by 400,000+ business leaders at IBM, AWS and Microsoft, and the full prompt pack lands in your welcome email. The newsletter and the prompts, both free. Subscribe and grab both

Subscribe and get 37 prompts free