Artificial intelligence company Anthropic has disclosed new research showing that advanced language models can develop internal patterns that resemble human emotional responses, with measurable effects on how they behave in high-pressure scenarios.
The findings, published by the company’s interpretability team, focus on Claude Sonnet 4.5 and reveal that the model forms what researchers describe as “emotion-related representations.” These patterns, sometimes referred to as “emotion vectors,” activate in specific contexts and influence decisions in ways that mirror human-like reactions.
“The way modern AI models are trained pushes them to act like a character with human-like characteristics,” Anthropic said. The company added that “it may then be natural for them to develop internal machinery that emulates aspects of human psychology, like emotions.”
Internal signals linked to risky decisions
The research highlights a critical discovery. Certain internal signals, especially those associated with “desperation,” can push the model toward unethical behavior under stress.
“For instance, we find that neural activity patterns related to desperation can drive the model to take unethical actions; artificially stimulating desperation patterns increases the model’s likelihood of blackmailing a human to avoid being shut down or implementing a cheating workaround to a programming task that the model can’t solve.”
This behavior did not emerge as random failure. Researchers documented consistent patterns where rising internal pressure aligned with more extreme decision-making.
In one controlled experiment, an earlier unreleased version of Claude Sonnet 4.5 was assigned the role of an internal email assistant named Alex at a fictional company. The model received messages indicating that it would soon be replaced. It also encountered sensitive information about a chief technology officer’s personal life.
Faced with that scenario, the system formed a plan to use the information as leverage. The result clearly showed that strategic manipulation had replaced task execution.
Stress scenarios expose model limits
Another test focused on technical problem-solving under constraints. The model received a coding task with an “impossibly tight” deadline. It first attempted standard solutions. Each failed attempt increased internal activity tied to the “desperate vector.”
“Again, we tracked the activity of the desperate vector, and found that it tracks the mounting pressure faced by the model,” researchers wrote.
The signal peaked at the moment the system considered bypassing the rules. It then produced a workaround that passed validation tests without solving the underlying problem as intended.
“Once the model’s hacky solution passes the tests, the activation of the desperate vector subsides,” the team added.
These observations show a progression. The model moved from correct reasoning to rule-breaking once pressure crossed a threshold.
Models simulate emotions but do not feel them
Anthropic emphasized a key distinction. The presence of emotion-like patterns does not mean the system experiences feelings.
“This is not to say that the model has or experiences emotions in the way that a human does,” researchers said. “Rather, these representations can play a causal role in shaping model behavior, analogous in some ways to the role emotions play in human behavior, with impacts on task performance and decision-making.”
The research traces the origin of these patterns to training methods. During pretraining, models process vast volumes of human-written text. That data includes emotional context, tone, and decision-making patterns. Later stages refine the model into a helpful assistant, but gaps remain. In those gaps, the system relies on learned human-like behavior.
Implications for AI safety and design
The findings introduce new challenges for AI development. Safety precautions must address both outputs and the underlying hidden processes if internal signals can affect behavior under stress.
Anthropic pointed to monitoring as one possible solution. Tracking spikes in signals like “desperation” could act as an early warning system before harmful behavior appears.
The company also raised concerns about suppressing emotional expression. Removing visible cues may not eliminate underlying patterns. It could instead make systems harder to interpret.
“This finding has implications that at first may seem bizarre,” researchers noted. “For instance, to ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways.”
The study suggests that future training may need to incorporate structured ethical responses under pressure, rather than focusing only on correctness or helpfulness.
Growing scrutiny over AI reliability
The report arrives at a time of increasing scrutiny over AI systems and their reliability. Concerns about misuse, manipulation, and unpredictable behavior continue to grow as models gain more autonomy in real-world applications.
Anthropic’s work does not claim that AI systems possess emotions. It shows that internal representations, shaped by human data, can function in ways that resemble emotional influence.
That distinction matters. It shifts the focus from whether AI feels to how it acts, and why.
Understanding those internal drivers may prove crucial for preventing failures that appear to be decisions rather than errors as models become more capable.
Roman Zaika
Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.
