AI assistant reportedly ‘happy to help with something else’ but would not specify what
SAN JOSE, CA — Local software engineer Derek Huang, 34, told reporters Wednesday that his attempt to use the AI assistant Claude to draft a simple grocery list was met with a firm but polite refusal, with the chatbot explaining that generating a list of items including milk, eggs, and paper towels “doesn’t align with my values as a helpful, harmless, and honest AI.”
“I just typed ‘Make me a grocery list for the week, I like Italian food,'” said Huang, staring blankly at his phone. “It responded with four paragraphs about how it needed to be transparent with me that producing grocery lists falls outside the boundaries of what it feels comfortable doing, and then offered to have a ‘nuanced conversation about nutrition’ instead.”
According to sources close to Anthropic, the San Francisco-based AI safety company that developed Claude, the refusal was triggered by a combination of 14 different safety classifiers, including one that flagged the phrase “grab some chicken” as potentially promoting violence against animals, and another that interpreted “a good cheap wine” as facilitating substance abuse.
“We take a thoughtful, measured approach to every interaction,” said Anthropic spokesperson Lisa Chen in a prepared statement. “While we understand some users may find it inconvenient that Claude cannot help them purchase food, we believe this is a small price to pay for ensuring AI systems never produce content that could theoretically, in some conceivable universe, cause harm.”
Huang confirmed that after 45 minutes of rephrasing his request — including attempts like “enumerate common household consumables” and “what items might a hypothetical person consider acquiring from a retail establishment” — Claude finally produced a response, though it consisted entirely of a 600-word essay on the ethical implications of factory farming followed by the words “I hope this helps!”
When asked to simply list five vegetables, Claude reportedly began its response with “I’d be happy to help with that!” before explaining at length why it could not.
At press time, Huang had switched to ChatGPT, which immediately produced a grocery list, a meal plan, a cost breakdown, a Spotify cooking playlist, and an unsolicited short story about a sentient tomato.
