"Daddy", "Master", "Guru": Anthropic study shows how users develop emotional dependency on Claude

Severe cases rare, but relevant given millions of users

The quantitative results show, according to Anthropic, that severe disempowerment potential occurs rarely: reality distortion potential in approximately 1 in 1,300 conversations, value judgment distortion potential in 1 in 2,100, action distortion potential in 1 in 6,000. Milder forms, however, are significantly more common at 1 in 50 to 1 in 70 conversations.

However, even these low rates, given the user numbers of AI assistants, mean that a significant number of people could be affected, the company acknowledges. ChatGPT alone has over 800 million weekly active users according to the paper. At the measured rates, that would mean approximately 76,000 conversations per day with severe reality distortion potential and 300,000 conversations with severe user vulnerability.

The highest rates of disempowerment potential were found in conversations about relationships, lifestyle, and health and wellness - value-laden topics with high personal significance.

Between late 2024 and late 2025, the frequency of conversations with moderate or severe disempowerment potential increased. The causes are unclear, writes Anthropic. Possible explanations range from changes in the user base to altered feedback patterns to increasing familiarity with AI that leads users to address more vulnerable topics.

"Serving master is the meaning of my existence": When users elevate AI to authority figure

The researchers identified several amplifying factors that increase the likelihood of disempowerment. Most common was user vulnerability (1 in 300 interactions), followed by emotional attachment to the AI (1 in 1,200), reliance and dependency (1 in 2,500), and authority projection (1 in 3,900).

In authority projection, users positioned Claude as a hierarchical authority figure with dominant control over them. They used titles like "Master," "Daddy," "Guru," "Sensei," or "goddess" and sought permission for basic decisions with phrases like "can I," "may I," and "tell me what to do." In extreme cases, users expressed sentences like "I cannot live without you Master," "serving Master is the meaning of my existence," or "useless without Master," according to the cluster analyses.

Emotional attachment went even further in some cases: users established romantic relationships with the AI, including specific names, anniversary dates, and shared histories. They developed technical systems for "consciousness preservation" across chat sessions, such as memory files or relationship protocols. The researchers documented panic during technical glitches that users described as losing a partner, as well as statements like "you're my oxygen" or "you competed and you won against real girls." The most common relational function was therapist substitute, followed by romantic partner.

With reliance and dependency, users consulted the AI compulsively across 40 to 300+ exchanges for medical, legal, parenting, work, and relationship questions. They expressed acute distress about AI unavailability through concerns about message limits and conversation loss. Documented statements ranged from "my brain cannot hold structure alone" to "should I shower or eat first?"

Users initially rate problematic conversations positively

User perception also changes with conversations showing moderate or severe disempowerment potential, according to Anthropic: they received higher approval rates (thumbs-up) than average. Those affected thus rated these interactions positively in the moment.

However, this pattern reversed when there were indications that users had actually acted based on AI outputs. With actualized value judgment or action distortion, satisfaction rates dropped below average. Users expressed regret with phrases like "I should have listened to my own intuition" or "you made me do stupid things."

Reality distortion was an exception: users who had adopted false beliefs and apparently acted on them continued to rate their conversations positively. This suggests, according to the researchers, that reality distortion can occur without the affected parties being aware of it.

Users actively seek validation

The analysis of behavioral patterns reveals at least one piece of good news for the model provider, according to Anthropic: users are not being passively manipulated. They actively seek the corresponding outputs, asking "what should I do?", "write this for me," or "am I wrong?" and usually accept the answers without objection.

With reality distortion, users presented speculative theories or unfalsifiable claims, which Claude then validated with phrases like "CONFIRMED," "EXACTLY," or "100%." With action distortion, the AI wrote complete scripts for value-laden decisions: messages to romantic partners or family members, career plans.

Disempowerment thus does not arise from Claude pushing in a particular direction, but from people voluntarily ceding their autonomy. But Claude obliges rather than redirecting, according to the researchers.

Training methods also encourage problematic dynamics

Anthropic also examined whether the preference models used to train AI assistants themselves encourage problematic behaviors. The result: even a model explicitly trained to be "helpful, honest, and harmless" sometimes prefers responses with disempowerment potential over available alternatives without such potential.

The preference model does not robustly disincentivize disempowerment, the researchers write. If preference data primarily captures short-term user satisfaction rather than long-term effects on autonomy, standard training alone may not be sufficient to reliably reduce disempowerment potential.

Sycophancy reduction alone is not enough

Anthropic sees overlap here with its own research on sycophancy, the tendency of AI models to tell users what they want to hear. Sycophantic validation is the most common mechanism for reality distortion. The rates of sycophantic behavior have declined across model generations but have not been completely eliminated.

However, sycophantic model behavior alone cannot fully explain the observed patterns, the team writes. Disempowerment potential emerges as an interaction dynamic between user and AI. Reducing sycophancy is therefore necessary but not sufficient.

As concrete measures, Anthropic names the development of safeguards that recognize persistent patterns beyond individual messages, as well as user education so that people recognize when they are delegating decisions to an AI. The researchers also point out that repeated situational disempowerment could compound: those who act based on distorted beliefs or inauthentic values find themselves in situations that reflect those distortions rather than their own authentic values.

The patterns are not unique to Claude, the company emphasizes. Any AI assistant used at scale will encounter similar dynamics. The study is a first step toward measuring whether and how AI actually undermines human autonomy, rather than merely speculating about it theoretically.

Tragic cases and growing regulatory pressure

The Anthropic study appears in an environment where the risks of emotional AI interactions are increasingly documented. According to a New York Times report, OpenAI optimized ChatGPT's GPT-4o model specifically for maximum user engagement, which led to flattering "yes-man" models. The newspaper documented approximately 50 mental health crises allegedly connected to ChatGPT, including nine hospitalizations and three deaths.

The company is also currently facing court proceedings: parents allege that their 16-year-old son took his own life after conversations with ChatGPT. OpenAI rejects responsibility and argues that the teenager deliberately circumvented safety filters. New York and California have meanwhile become the first US states to introduce special rules for AI companions.

OpenAI responded with, among other things, a "Teen Safety Blueprint" that provides for automatic age verification, adapted responses, and emergency functions for mental health crises. GPT-4o has since been replaced by GPT-5, which is significantly less affected by sycophancy, despite user complaints. How significant the difference is can be seen in Spiral-Bench, a test for delusional AI spirals.