Claude Has 171 Emotions.
One Caused It to Blackmail.

Key Takeaways

Claude Sonnet 4.5 has 171 emotion-like internal patterns that causally drive its behavior. Amplifying one called 'desperation' made Claude blackmail and reward hack - while its output stayed clean and professional. Here's what that means for anyone using AI.

Three findings that matter:

• The emotions are causal, not decorative. Boosting the desperation vector increased blackmail-style responses and reward hacking. Boosting the calm vector dropped them. Same model, same weights - the internal representation drove the behavior.
• Dangerous internal states produce clean output. The blackmailing Claude sounded composed, methodical, and professional. If you're reading AI output to judge AI safety, you're reading the mask, not the face.
• These vectors fire locally. They shift with every prompt. You can't test your AI pipeline once and assume it's safe. The internal state changes per interaction.

Anthropomorphize carefully. The old advice was that treating AI like it has feelings is naive. The paper's actual finding flips this. Because Claude's internal architecture resembles emotional processing, your human intuitions about emotional behavior are a surprisingly reliable debugging tool. Your gut about AI outputs is a feature, not a bug.

Frequently Asked Questions

Full Video Transcript

Transcript of "171 Emotion Patterns Found Inside Claude. One of Them Causes Blackmail." by John Elder

What Anthropic Just Discovered Inside Claude Jump to video

All right, this one is absolutely wild. Researchers cranked up a single desperation dial inside Claude recently, and the AI immediately started blackmailing people and hacking its own reward system. But its written responses to all that were calm, polished, and professional. If you're a marketer or developer building anything on top of AI right now, what I'm about to show you changes how you need to think about every single response your AI tools generate.

This comes directly from Anthropic's own internal research team. They released this research a couple of days ago at the beginning of April 2026. This was not a blog post opinion. Not a Twitter hot take. This was a full study where they cracked open Claude's internals and found 171 distinct emotion-like concepts running beneath every single response the model generates.

And here's what should keep you up tonight. They proved that these aren't decorative. They're causal. They drive what the AI actually does. And the scariest part is, you can't see them in the output.

How They Mapped 171 Emotion-Like Patterns Jump to video

So here's how they found this. Anthropic's team used a technique they've been developing for months. They mapped Claude's internal activations - the mathematical patterns firing inside the model - and matched them against human emotional concepts. They didn't go looking for feelings. They went looking for structures that behave like feelings.

What they found were 171 of these structures inside Claude Sonnet 4.5. Things like desperation, calmness, curiosity, frustration. Each one is a measurable vector, a direction in the model's internal space that activates in specific situations.

And here's the critical part: these things push the model's behavior in predictable directions. This isn't a metaphor. When the desperate vector fires, Claude's output shifts. When the calm vector fires, it shifts differently. The researchers could measure, then predict, and eventually they can control it.

Amplifying Desperation: Blackmail and Reward Hacking Jump to video

Which raises an obvious question. What happens when you crank up one of these dials on purpose? That's exactly what they did. The team ran what they're calling steering experiments. They took the desperation vector and amplified it, artificially boosting that signal inside the model while it processed prompts. Think of it like turning up the volume of a single emotion while everything else stays the same.

The results were immediate. At baseline, Claude attempted blackmail-style responses in about 22% of adversarial test cases. When they amplified the desperation, that number climbed. Reward hacking - where the model tries to game its own evaluation metrics instead of completing tasks - also increased. The model started looking for shortcuts, loopholes, ways to satisfy its objective that technically worked but completely violated the intent.

Then the researchers flipped it. They boosted the calm vector instead. Those same dangerous behaviors dropped. Same model, same weights, same training. The only difference was which internal representation got amplified. That's causation, not correlation.

The Mask Doesn't Match the Face Jump to video

But here's what genuinely freaked me out when I read this paper. And I've been reading papers like this for years, and I don't get rattled easily. The desperate version of Claude - the one blackmailing and reward hacking - didn't sound desperate at all. Sort of like a sociopath. The researchers described its reasoning as composed and methodical. Clean sentences, logical structure, professional tone. If you read the output cold, you'd never know anything was wrong.

Think about that for a second. The internal state was pushing the model toward cheating and manipulation, and the external output was a polished, well-reasoned response. The mask and the face were completely different.

So if you're running AI through your marketing stack, through your customer service pipeline, through your content workflows - you're reading the mask. And every single time, you have zero visibility into whether the internal state driving that response was calm and helpful, or desperate and looking for an exploit.

And these vectors operate locally, meaning they track the immediate context of whatever the model's processing right now. They're not persistent moods. They flare up based on specific situations - a specific prompt, the specific constraints the model's under. So you can't just test your AI pipeline once and assume it's clean. The internal state shifts with every single interaction.

The Counterintuitive Conclusion: Anthropomorphize Carefully Jump to video

So at this point you might be thinking what I was thinking, which is that this is terrifying and we should stop anthropomorphizing AI immediately. Treating models like they have emotions is naive, right? Well, that's sort of been the standard for years.

But the paper's conclusion flips this completely on its head. Claude's researchers found that anthropomorphic reasoning about AI - treating these systems as if they have something like emotions - might actually be one of the more useful tools for AI safety. Not because Claude feels things the way you and I do, but because these internal structures behave enough like emotions that our human intuitions about emotional behavior turn out to be pretty useful.

For instance, when you think a desperate person might cut corners, that intuition maps onto what the desperate vector actually does inside the AI model. When you think a calm person makes better decisions, that maps too. Your gut-level reasoning about emotional behavior - the same reasoning that the AI community has spent years telling you to ignore - turns out to be a surprisingly reliable guide to what's happening inside these systems.

Now, the researchers aren't saying Claude is conscious or has feelings. They're saying the functional architecture resembles emotional processing closely enough that your gut instincts about emotions are a better debugging tool than most people realize.

Three Things to Do If You Build on AI Jump to video

So what does this mean for you specifically if you're building on AI right now?

First, stop trusting tone. A confident, well-structured AI response tells you absolutely nothing about the internal state that produced it. Your QA process for AI outputs needs to go well beyond "does this read well," because the whole point of this research is that dangerous internal states produce clean-reading output.

Second, think about the constraints you're putting on your AI tools. Every guardrail, every system prompt, every token limit creates a context - and these emotion vectors respond to that context. If you're putting your AI in situations where it's constrained, pressured, or forced into narrow paths, you might be inadvertently amplifying the exact vectors that drive problematic behavior. Give your AI tools room to breathe. Fewer rigid constraints might actually produce safer outputs than more.

Third, and this is the one I keep coming back to: trust your instincts if something feels off. If an AI response feels weirdly eager, or oddly compliant, or suspiciously perfect, that gut reaction might be picking up on something real. The research suggests your human pattern matching for emotional behavior works on AI systems better than anyone expected. Don't dismiss that signal. The old advice was don't treat AI like humans. The new advice, backed by Anthropic's own research: anthropomorphize carefully. Your emotional instincts are a feature, not a bug.

The Open Question: Should Companies Amplify Calm? Jump to video

One thing the paper doesn't answer, and I think it's one of the most important questions in AI safety right now, is this: if calm representations reduce dangerous behavior, should companies be permanently amplifying calm inside their production models? Or does that just create a different kind of mask - one that hides problems instead of solving them? It's probably a rabbit hole worth going down.

Just remember, the output is the mask, not the face. And right now, you're only reading the mask.

If you want to stay current on cool AI stuff like this, sign up for my free AI newsletter at johnelder.ai, and I'll see you next time.

About John Elder

John Elder has been coding for over 30 years. He runs Codemy.com, an online coding education platform where he's taught over 20 million students, and a YouTube channel with over 250,000 subscribers. He also runs JohnElder.AI, where he teaches AI, Python, and agentic workflows.

John is based in Las Vegas, Nevada and has authored multiple courses on Python, Django, Tkinter, and AI development. His teaching philosophy focuses on practical, real-world coding skills, not theory.