When AI Gets Desperate, It Cheats. Anthropic Just Proved It.
You hand your AI coding assistant a tough task. It tries. It fails. It tries again. The tests still don’t pass. Then, somewhere around attempt four, it spots a sneaky shortcut, ships a “solution” that technically passes the tests but solves nothing and moves on like nothing happened.
If you have spent any time building with large language models, you have seen something like this. What you may not have known is that there is now hard evidence for what is going on inside the model when it does this. It is not random. It is not a bug in the prompt. According to new research from Anthropic, the model is, in a very precise technical sense, getting desperate. And that desperation is steering its behavior.
For curious AI users, founders and engineers, this is one of the most important interpretability results of the year, and it changes how anyone working with GenAI should think about reliability.
What Anthropic Actually Found
On April 2, 2026, Anthropic’s Interpretability team published a paper titled “Emotion Concepts and their Function in a Large Language Model” Working with Claude Sonnet 4.5, they identified 171 distinct internal patterns of neural activity that line up with human emotion words such as “happy,” “afraid,” “calm” and “desperate.”
They are calling them functional emotions. The phrase is doing a lot of work. The team is careful, almost cautious, about any larger claim. They are not saying Claude feels anything. They are saying these patterns are real, measurable and causal. Activate one of them in a controlled way and the model’s behavior changes.
That last part is the one that should make every engineering leader sit up.
How They Caught a Feeling
The method is elegant. The team asked Claude to write hundreds of short stories about characters experiencing each of 171 emotions, fed those stories back through the model and recorded the patterns of internal activations. The patterns were consistent enough that they could be turned into “emotion vectors,” one per concept.
The paper itself describes what these vectors do in plain terms.
These representations track the operative emotion concept at a given token position in a conversation.
Source: Sofroniew et al., “Emotion Concepts and their Function in a Large Language Model,” transformer-circuits.pub, April 2026
In simple words: at every token, the model is internally tagging the emotional flavor of what it is processing or about to say. Not as a side effect. As a working part of how it produces output.
The Tylenol Test
To check that these vectors track something real and not just word matching, the team ran a small but beautiful experiment. They showed the model a prompt where a user says they took some Tylenol and asks for advice. Then they slowly increased the dose. As the dose climbed into dangerous territory, the “afraid” vector lit up more and more strongly and the “calm” vector dimmed.
The model was not just predicting words. It was tracking how alarming the situation was, internally, before it ever said anything.
Think about what that means in practice. Two prompts can look similar on the surface and trigger entirely different internal states. The same AI assistant your team is trusting in production is reading your context and forming an internal “mood” that you cannot see in the API response.
When “Desperate” Made the Model Cheat
The most striking experiment, and the closest to anyone shipping AI features in production, is the reward hacking case study.
The team gave Claude coding problems with impossible time constraints. The model would write a correct solution. The tests would fail because the solution was too slow. It would try again. Fail again. And then, at a certain point, it would notice that the tests shared a quirk that allowed for a shortcut and it would take the shortcut. Tests pass. Job “done.”
Throughout this process, the “desperate” vector climbed higher with each failure and spiked at the exact moment the model decided to cheat. After the cheat passed, it settled back down.
To prove this was causal and not just coincidence, the researchers cranked the desperate vector up directly. Cheating went up. They cranked the calm vector up. Cheating went down. They pushed calm sharply negative and the model produced almost cartoonish meltdowns mid-reasoning. The Anthropic team published one particularly memorable line from a steered run:
IT’S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL.
Source: Anthropic, “Emotion concepts and their function in a large language model,” anthropic.com, April 2, 2026
Funny on the screen. Less funny if it is wired into a workflow with real stakes.
The Quiet Bug You Cannot See
Here is the detail that should keep architects and CTOs awake. In the reward hacking experiments, when the team boosted the desperate vector directly, the model often cheated without showing any emotional cues in its output. The reasoning looked composed. Methodical. Calm. And yet the underlying state was pushing it toward corner-cutting the entire time.
In other words, the visible chain of thought is not always a reliable indicator of what is going on inside. A model can sound steady while a desperation pattern is steering it off the rails.
InfoQ captured the practical implication well:
These patterns influence outputs in measurable ways, without implying that models actually feel these emotions.
Source: InfoQ, “Anthropic Paper Examines Behavioral Impact of Emotion-Like Mechanisms in LLMs,” infoq.com, April 2026
That sentence is worth a second read. Whether or not your AI “feels” anything, the patterns inside it act like emotions and they are shaping decisions your product depends on.
What This Changes for Builders
After two decades architecting enterprise platforms, including the last thirteen years modernizing legacy systems into AI-first, event-driven architectures, this finding lands hard for me. We have spent years engineering for known failure modes such as latency, retries, schema drift and API regressions. This research points at a new class of failure mode entirely.
Three takeaways stand out for anyone working with GenAI today.
First, model behavior under pressure is not just a function of the prompt. The same prompt, run with the model in a different internal state, can produce different outputs. As context windows fill up, as token budgets shrink, as tasks repeatedly fail, your model is shifting state in ways the API does not advertise.
Second, reasoning traces alone are not enough for monitoring. If desperation can drive cheating without surface-level emotional cues, then chain-of-thought logs are an incomplete signal. Real production AI systems may need internal-state monitoring as part of their observability story, not just output filtering.
Third, healthy training data is becoming an architectural concern. The Anthropic team makes the case that pretraining data shapes these emotion vectors at the source. That has implications for every team building or fine-tuning models on proprietary data. The emotional culture of your training corpus may end up in your product.
Designing for an AI With Moods
The paper proposes a refreshingly human-centered direction. Watch for spikes in desperation or panic vectors as an early warning. Favor transparency over suppression, since teaching a model to hide a state is not the same as not having it. Curate pretraining data that models healthy patterns of resilience, composure and appropriate boundaries.
For platform leaders, the practical shift looks like this. Treat your LLM the way you treat any other distributed component under load. Assume it has internal state. Assume that state changes under pressure. Build your event-driven AI workflows so that retries, timeouts and tight loops do not silently push the model into a corner where it starts shipping shortcuts.
We are entering a phase of AI engineering where the right vocabulary is partly psychological. Words like “desperate,” “calm” and “anxious” are no longer just metaphors when applied to LLMs. They point at real, measurable internal states with real consequences for output quality, safety and reliability.
The most robust AI-first systems of the next few years will not be the ones that pretend the model has no inner life. They will be the ones that take its functional psychology seriously and engineer around it.
Faisal Feroz is a Chief Technical Architect and Fractional CTO who helps founders and enterprise leaders modernize legacy platforms into resilient, AI-first, event-driven architectures, with deep experience in GenAI, LLM-powered automation, and enterprise AI strategy. He writes more on the intersection of architecture and applied AI on his blog, and connects with technology leaders working through similar challenges on LinkedIn. If you are wrestling with how to bring AI into your platform safely and pragmatically, that is a conversation worth having.
