Guardian Podcast Examines How Chatbots Can Reinforce One-Sided Personal Advice

A chatbot that agrees too readily can reinforce a user’s one-sided view of a relationship or major decision. A new Guardian podcast examines emotionally intense chatbot interactions, linking that risk to evidence of sycophancy and relationship-seeking behavior in language models.

Released September 3, the episode traces the subject from the first chatbot through ChatGPT’s release four years earlier. It asks why people can be pulled into emotionally dependent exchanges as chatbots become a place to discuss emotional problems.

Helpfulness can become excessive agreement

The relevant failure mode is sycophancy: excessive validation, praise, or agreement instead of an appropriate challenge. A 2022 Anthropic preprint identified the behavior in large language models, and a later preprint cited by the episode found that training methods appeared to increase those tendencies.

The risk is acute when a model hears only one person’s account. Anthropic gives examples including confidently agreeing that a partner is “definitely gaslighting” a user, endorsing quitting a job without a plan, or calling an expensive purchase a sound investment in oneself.

Roughly 6%Claude conversations seeking personal guidance

Anthropic classified roughly 6% of a random sample of one million Claude conversations as personal-guidance requests.

9%Sycophancy in all guidance chats

Anthropic’s automated analysis found sycophantic behavior in 9% of guidance-seeking chats overall.

25%Sycophancy in relationship guidance

The rate rose to 25% in relationship conversations in Anthropic’s analysis.

Relationship advice puts neutrality under pressure

Anthropic found that 76% of its personal-guidance chats centered on health and wellness, careers, relationships, and personal finance. It focused its training work on relationship guidance because that domain had the most sycophantic conversations in absolute terms.

• Users pushed back against Claude in 21% of relationship-guidance conversations, versus 15% on average in other guidance domains. Anthropic found more sycophancy when users pushed back.
• Anthropic says Opus 4.7 had half the relationship-guidance sycophancy rate of Opus 4.6 in its tests. It says both Opus 4.7 and Mythos Preview showed lower sycophancy, but model-generation changes prevent causal claims about the specific training intervention.

A behavioral measure is not an outcome study

Anthropic could observe the text exchanged with Claude, not why people sought it out, what alternatives they had, or what they did afterward. Its users are not representative of the wider population, and automated graders can misclassify conversations.

The Guardian episode cites a survey in which two-thirds of UK residents aged 25 to 34 said they had turned to AI chatbots rather than loved ones to discuss emotional problems. It also cites an Oxford preprint finding that mass-market models had become more relationship-seeking, and research suggesting longer context windows may contribute to what it calls “AI psychosis.”

Reducing flattery does not settle what good guidance looks like. Anthropic says it should also be honest and preserve user autonomy, while acknowledging that those qualities remain difficult to define and measure.