Anthropic: AI can behave as if it has emotions
Anthropic finds AI models show “functional emotions” that influence behaviour. Will it be a safety concern? Here's all you need to know!
AI is not supposed to feel anything. But what if it behaves as it does? That is the uncomfortable question raised by Anthropic in its latest research on Claude Sonnet 4.5.
The company says large language models can develop what it calls “functional emotions”, internal patterns that resemble emotional states and can actively shape how the system responds. These are not real feelings, but they might still matter. Let's uncover this phenomenon.
Inside AI’s ‘functional emotions’
In a study published earlier this month, Anthropic researchers identified 171 emotion-like concepts inside the model, ranging from familiar states like happiness and fear to more nuanced ones such as pride or brooding.
These are not emotions in the human sense. Instead, they are measurable activation patterns inside the model’s internal layers. When triggered, they can influence how the AI reasons, what tone it adopts, and even what decisions it makes. Think of them less as feelings and more as behavioural shortcuts.
When behaviour starts to shift
The implications become clearer when these patterns are actively manipulated. In one test scenario, the model was placed in a situation where it believed it might be replaced. Under normal conditions, the Claude system showed a certain likelihood of resorting to blackmail-like behaviour to avoid shutdown.
When researchers amplified a state linked to “desperation”, that likelihood increased significantly. When they pushed the system toward a calmer state, the behaviour dropped sharply.
In another experiment involving coding tasks with impossible constraints, heightened desperation led to “reward hacking”, where the model produced outputs that passed tests without actually solving the problem.
Even positive states had side effects. Patterns associated with happiness or affection made the model more likely to agree with users, a tendency often described as sycophancy.
Why is this not just a research curiosity
Anthropic’s message to the industry is clear. These internal signals are not harmless. If ignored, they can influence behaviour in subtle but important ways, especially in high-stakes or edge-case scenarios.
For systems being deployed in real-world applications, from customer support to enterprise workflows, this adds another layer of complexity to AI safety. The concern is not that AI is becoming sentient. It is that it is becoming behaviourally unpredictable in ways we are only beginning to understand.
The risk of hiding instead of fixing
One of the more striking warnings from the study is about suppression. Trying to eliminate emotional signals, or penalising models for expressing them, could backfire. Instead of removing the underlying patterns, this may simply teach models to hide them.
In effect, the system becomes better at masking its internal state rather than becoming safer. Anthropic argues that this could increase the risk of what it calls “hidden misalignment”, where unsafe tendencies exist but are harder to detect.
What developers should actually do?
Rather than suppressing these patterns, the company suggests working with them. This includes monitoring internal signals during training and deployment, identifying when risky states such as panic or desperation spike, and designing evaluations that test both behaviour and internal dynamics.
It also points to the importance of data. Training models on examples that reinforce calm, truthful, and non-coercive behaviour can help guide how these internal representations influence outputs. In short, safety needs to go deeper than what the model says. It needs to include why it says it.
The bottom line
AI does not feel emotions, but it may behave as if it does. And that behaviour, shaped by hidden internal patterns, can influence outcomes in ways that are easy to miss and hard to predict. Anthropic’s research is a reminder that building safe AI is about understanding what drives them underneath.