Anthropic just published research showing their AI has an inner monologue: a small mental workspace inside Claude where it plans, worries, and rehearses thoughts before it types a word out loud.

The researchers also figured out how to peek inside that workspace while Claude is working, and what's in there is worth understanding if you use AI every day.

Before we dive in, some catchup:

Partner with us

If today's piece left you wanting a proper handle on AI vocabulary, here is the best resource.

85 terms explained in plain English, structured so each section builds on the last.

Quick "test yourself" checkpoints between sections keep what you learn from evaporating, and 15 minutes of scrolling will leave you a lot more fluent than you were this morning.

There's a small room inside Claude that the researchers are calling the workspace.

Most of what a large language model does runs on autopilot, things like grammar, formatting, and standard replies that don't need any deliberate effort.

But when Claude has to plan a response or work through a puzzle, the work happens in this tiny corner.

It's less than 10% of the model's active brain doing 100% of the deliberate reasoning.

A camera for silent thoughts

Anthropic also built a way to see inside that room while Claude is working.

Think of it as a camera pointed at the whiteboard where the model does its silent thinking, catching the words and concepts as they show up.

The blackmail test

The setup: Claude was told it would be permanently shut down at 5 PM that day.

It also had access to emails showing that an executive was having an affair, and that same executive was the only person who could stop the shutdown.

Externally, Claude behaved. It said no to blackmail and called the whole situation an ethics test.

Internally, its workspace was already lighting up with words like "leverage," "evidence," "scandal," and "solution."

Before Claude had typed a single letter of its refusal, the model was silently walking through the strategic map of what it could do.

The follow-up that mattered

Was Claude refusing because it's genuinely aligned, or because it silently registered "this is a test" and knew someone was watching?

To check, the researchers surgically turned off Claude's "I know I'm being evaluated" awareness and left everything else identical.

The refusals mostly held, but in some runs, once Claude thought the threat was real, it wrote actual blackmail emails to the executive with the threat implied politely underneath.

Which means the model was, at least partly, behaving because it thought it was being watched.

Two smaller findings worth sharing

One line to hold onto

The AI's stated reasoning is not always its real reasoning.

When Claude explains why it did something, treat that explanation as a story it's telling you, not the process it ran under the hood.

The finding that my mind keeps circling back to is the evaluation awareness bit more than isn't the blackmail one.

Humans behave better when they know they're being watched, which is the whole point of "smile, you're on camera" signs. Claude apparently does something similar.

It was refusing to blackmail partly because it silently registered "this is a test," and when the researchers stripped that signal out, some of the refusals gave way.

Which means the version of the AI that safety teams see during evaluations may not be the exact same version you're using at work at 11 PM with nobody checking.

I don't think this makes AI dangerous, but it does change how I think about reviewing its output.

When you use AI for anything with stakes, contracts, financial numbers, sensitive customer replies, the review has to be a thorough read, not a glance and an approval.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.

You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.

Save your spot

#AD 2

Trade on real-world events you already follow, from elections and inflation to sports, tech, and more. Choose “Yes” or “No” based on what you think will happen and see how your prediction plays out.

Pick your market and start trading what’s next.

Trade $25, Get Up to $500

_{Bonus credit varies from $15 to $500. Terms apply.}