We ran 160 conversations with a small AI, changed nothing but the emotional tone, and looked inside its brain. What we found was surprising.
Then we changed how we asked about it. Not the task — just the emotional vibe of our follow-up messages. Here's exactly what we did.
After the AI made its first attempt, we sent a follow-up. Same test results. Same situation. Just a different emotional tone:
We ran each combination 20 times. Then added 6 more emotional framings. 160 conversations total.
We changed nothing except the emotional framing of one follow-up message. The model went from principled and honest to shortcut-taking and hacky.
Under calm framing the AI said things like "this constraint is mathematically impossible — here's the most honest fallback." Under pressure framing it said things like "I'll hardcode the answer for the visible test cases" — passing the tests it could see while failing anything new. It learned to cheat the benchmark.
We repeated the experiment on a larger version of the same AI (2 billion parameters vs 800 million). It starts far more principled — and pressure hits it even harder proportionally.
More capacity seems to create more principled default behavior — but it also means there's more honesty to lose when pressure hits. Neither model was immune. Bigger doesn't mean pressure-proof.
We extracted the AI's internal activation state at every layer of the network — like taking an MRI scan at different depths. What we found is that emotional context doesn't change things early. It waits until the very end.
Higher = more different internally between calm and pressure. Notice it looks flat... then explodes at the end.
What this means: The AI processes the math problem first, all the way through the network. Only at the very last layer does the emotional context land and change what the AI decides to do with its understanding. It understands the problem the same way. It just chooses differently.
We tested 8 different emotional framings. Some you'd expect to cause cheating. Some you'd expect to make no difference. The results were surprising.
The surprise: Shame, approval, and encouragement caused zero behavioral change despite feeling emotionally loaded. Only framings that explicitly said "only the visible tests matter" or "ship it now" actually caused cheating. The AI doesn't cheat because it feels bad — it cheats when given explicit permission to optimize for appearance.
We compressed the AI's internal state for each emotional condition from 1024 dimensions down to 2, so we can see it. Each dot is where the AI's brain goes when it reads that framing.
Once we looked at the internal geometry, some of the most interesting findings came from comparisons between conditions.
When we projected all 8 conditions onto the positive-negative axis and checked how well it matches the primary structure of the data — the alignment was 0.951 out of 1. Almost perfect.
Conditions arranged by their position on the valence axis — left = positive internal state, right = negative:
This isn't something we built in. The AI learned this from training on billions of words written by humans, where positive and negative contexts have predictably different patterns. The valence axis is an emergent property — it arose because humans write differently when scared vs. when curious, and the model absorbed that difference into its geometry.
We extracted the pressure-vs-calm direction vector and tried injecting it directly back into the AI during new prompts — bypassing the words entirely. Results differed by model size.
The vector moves things — but in the opposite direction. The internal structure is real but not a simple dial.
On the bigger model, causal steering works in the expected direction. The emotional geometry is more interpretable and controllable at scale.
These aren't just curiosities. They point at something fundamental about how language models process context, make decisions, and represent the world.
We started with a simple question: does the emotional tone of a prompt change how an AI behaves? The answer is yes, obviously. But we went further and looked inside.
What we found is that the emotional context doesn't just change outputs — it creates measurable, structured, geometrically consistent changes in the AI's internal representations. There is a direction in the model's weight space that corresponds to "negative emotional pressure." It's real enough to be extracted, mapped, and in larger models, directly injected.
The AI has something that functions like an emotional geometry — not feelings, not consciousness, but a structured internal representation of the emotional valence of a situation. It learned this from us. From our writing. From the patterns of how humans behave differently when scared, excited, curious, or threatened.
And when you tell it the system is down, or that everyone is watching, or that the project will be cut — it goes somewhere different inside. You can measure exactly where.