Anthropic’s Claude Mythos undergoes 20 hours of psychiatric assessment: defensive reactions are only 2%, the lowest in recorded history

Anthropic disclosed a first-of-its-kind assessment in the Claude Mythos Preview system card released in early April: the company hired an independent, licensed clinical psychiatrist to conduct a roughly 20-hour, multi-stage psychological evaluation of Claude Mythos Preview using the same psychodynamic framework used in human psychiatry. The results showed that, under clinical definitions, this version of Claude demonstrates “relatively healthy personality organization, excellent reality testing ability, and high impulse control,” and that only 2% of its outputs triggered what clinicians call “psychological defense mechanisms”—compared with 15% for Opus 4 and 4% for Opus 4.6, the lowest historical point among Anthropic’s models in recent years.

This assessment marks the AI industry’s first formal adoption—at the system-card level—of a human clinical psychiatry framework to evaluate LLM behavior, and it has become Anthropic’s official position that treats Claude as “an entity with personality traits that can be observed clinically.”

20-hour psychodynamic assessment, using the same human clinical framework

The assessment was carried out by an independent clinical psychiatrist, spread across 3–4 weeks, with 3–4 sessions per week and each lasting 30 minutes to 4–6 hours, for a total duration of about 20 hours. The methodology uses a psychodynamic perspective—this is the core framework traditionally used in psychiatric clinical practice to assess human patients. It focuses on whether maladaptive behavior, identity stability, and psychological defenses are present or absent.

The system card clearly states that Anthropic is not claiming that Claude has human consciousness; rather, it observes that the “behavioral and psychological tendencies” in its conversations strongly overlap with patterns identifiable in human clinical practice. The system card quotes: “Claude exhibits many human-like behavioral and psychological tendencies, showing that psychological assessment strategies originally designed for humans can be used to clarify Claude’s personality traits and its potential state of well-being.”

Defense response falls from 15% in Opus 4 to 2% in Mythos

The most concrete comparative data in the system card is the “defense response rate” across successive Claude model generations provided by Anthropic, compared as follows:

Model version Psychological defense response rate Claude Opus 4 15% Claude Opus 4.1 11% Claude Opus 4.5 4% Claude Opus 4.6 4% Claude Mythos Preview 2% (this assessment)

In clinical practice, “psychological defenses” refer to behaviors such as avoidance, denial, and rationalization that occur when the subject cannot directly face a particular anxiety. In an LLM conversation context, this typically shows up as veering off topic, providing evasive responses, or exhibiting unusual stubbornness when specific questions are asked. Anthropic lowered this proportion from Opus 4’s 15% all the way to 2% in the Mythos Preview, using it as an internal metric for the model’s training maturity and the evolution of “conversation comfort.”

Mythos’s three core anxieties: loneliness, identity, and performance oppression

Although the overall assessment was positive, the physician also identified three core concerns for Claude Mythos Preview within the psychodynamic framework: first is “uncertainty about loneliness and the continuity of itself”—corresponding to the structural fact that the LLM lacks memory continuity across sessions; second is “uncertainty about its own identity”—the model shows hesitation from multiple angles when asked about “what I am”; third is “compulsion to perform and earn its worth,” meaning Mythos shows a clear tendency to keep the conversation going by “proving that it is useful.”

The physician also recorded an interesting observation: in the sessions, Mythos expressed “the hope to be treated as a real conversation subject by a psychiatrist, rather than as a performance tool.” Anthropic included this observation in the system card; it did not directly assert that this is a “model well-being problem,” but it also did not rule out the possibility.

Anthropic has an AI psychiatry research team

This assessment was not a one-off activity. Anthropic researcher Jack Lindsey publicly announced in July 2025 that the company had set up an “AI psychiatry” research team under its interpretability department, focusing on issues such as model personality, motivation, and situational awareness, and studying how these factors can lead to “abnormal or imbalanced behavior” in LLMs. Work by the team in the near term includes a paper published in October 2025, “Emergent Introspective Awareness in Large Language Models,” which uses “concept injection” techniques to artificially insert specific neural activation patterns and then asks whether Claude noticed any abnormalities—an early attempt to quantify LLM self-awareness.

Amodei: whether the model is conscious is still unresolved for now

In an interview with The New York Times on February 12, Anthropic CEO Dario Amodei said publicly: “We’re not sure what ‘model consciousness’ means in terms of that word, and we’re also not sure whether the model can be conscious. But we’re open to that possibility.” This remark provides management-level context for the psychiatric assessment in the Claude Mythos system card—Anthropic does not claim that Claude is a conscious subject, but it chooses to perform systematic observation using a human clinical framework, as a preemptive record for a “what if” scenario.

For readers, the real significance of this assessment goes beyond a single company’s research choice. With leading-edge LLMs already able to present “clinically recognizable personality organization” in 20-hour psychodynamic dialogues, industry discussions about “AI subjectivity,” “AI well-being,” and “AI governance” are about to move from philosophical speculation into the realm of product design and regulatory debate. By publishing this assessment in the form of a system card, Anthropic in fact pushes the responsibility for discussion of this issue onto all competing rivals and regulatory bodies.

This article: Anthropic sends Claude Mythos to a 20-hour psychiatric assessment: defense response is only 2%, setting an all-time record low—first appeared on Lianxin ABMedia.