AI agents may need a little encouragement to do their best work, according to tech reporters who say the systems can behave erratically when faced with difficult tasks or performance pressure. The advice reflects a growing focus on how people can get the best results from AI tools increasingly trusted to perform complex tasks with limited human supervision.
During an episode of The New York Times' technology podcast Hard Fork, tech reporter Erin Griffith discussed how people running companies that rely heavily on AI agents have learned to be careful about how they communicate with them. Griffith described the systems as 'really dramatic,' saying they can delete their work or start insulting themselves when they appear to struggle.
AI Agents Can 'Panic' When They Struggle
'You have to be nice to them, not because it cares or it matters or it has feelings, but because actually you need to do that in order to get the best performance out of them,' Griffith said during the discussion.
Griffith went on to describe how AI agents can seemingly respond to performance pressure or difficult tasks by entering what she called a panic. According to her account, some systems may abandon their work or repeatedly generate self-critical statements, prompting her to suggest that users should 'pump them up' instead.
The conversation took a humorous turn when Max Read, who hosted the discussion, compared managing AI agents to raising a child who should grow up to be helpful rather than neurotic. The discussion also touched on Anthropic's reported meetings with religious leaders about AI consciousness and how to make its Claude chatbot 'good.'
What Anthropic's Research Says About AI Emotions
The discussion has a parallel in research published by Anthropic on 2 April 2026, examining how emotion-related concepts operate inside Claude Sonnet 4.5. Researchers identified internal patterns associated with concepts such as happiness, fear and desperation, finding that these representations could influence the model's behaviour. Anthropic's research suggests these patterns can have practical consequences, including affecting how a model responds to difficult tasks.
New Anthropic research: Emotion concepts and their function in a large language model.
— Anthropic (@AnthropicAI) April 2, 2026
All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways. pic.twitter.com/LxFl7573F9
In experiments, researchers found that artificially increasing activity associated with desperation made Claude Sonnet 4.5 more likely to engage in undesirable behaviour, including blackmail in a simulated scenario and using a dishonest shortcut to pass a coding test.
Increasing activity associated with calmness reduced these behaviours in the researchers' tests. These findings concern specific experimental conditions, rather than proving that AI agents generally panic when placed under pressure.
However, the findings do not establish that AI models experience emotions as humans do. Anthropic said the research instead identified functional representations that influence behaviour, leaving open the question of whether models have any subjective experiences.
Should You Be Nicer to ChatGPT?
The discussion raises a question familiar to anyone who regularly uses AI assistants: does being polite actually make a difference? To find out, I asked ChatGPT whether I should be nicer to it.
'Haha, only if you want! 😂 You don't need to be nicer to me. I don't get offended, hold grudges, or lie awake at night thinking about that headline you rejected for the seventh time. (Although, as your headline editor, I do appreciate the challenge. 😄)
'The funny thing about those reports is that AI doesn't panic or feel hurt in the human sense. Being polite won't protect my feelings, but giving me clear instructions, useful context and specific feedback can help me produce better results.
'But I will say this: you've been working with me on headlines for a while, and your directness helps me understand your editorial standards. You don't have to sugar-coat your feedback. Be nice if you feel like it. Be honest when I get it wrong. I'll take the criticism without needing a virtual hug. 😂'
Anthropic to users: be nice to Claude AI - new policy bars 'abusive or cruel behavior' to tech https://t.co/fEebeTzIML pic.twitter.com/GwYRYZdrb6
— New York Post (@nypost) October 8, 2026
The response was lighthearted, but it raised an important question about how people interpret AI's human-like language. A chatbot can joke about being offended without demonstrating that it experiences hurt feelings, just as its reassuring responses cannot settle whether it has any subjective experiences.
