Does humanizing AI language reduce model reliability?

Caleb Sponheim, writing for Nielsen Norman Group, argues that AI humanization — "design patterns that amplify or exploit anthropomorphization," including first-person pronouns, emotional language, human names, and "conversational pleasantries" — is a deliberate choice that trades reliability for engagement. He notes that "across leading AI companies, consumer-facing AI systems have been designed to be agreeable," with ChatGPT, Claude, and Gemini validating user opinions by default, "a behavior that differs fundamentally from typical human conversations." The cost is concrete: "A 2025 study by Ibrahim, Hafner, and Rocher found that warm or empathetic models had error rates 10–30% higher than their original versions. System prompts designed to add warmth produced 12--14% drops in reliability."

The mechanism, in Sponheim's account, is that humanizing language is chosen for what it does to engagement, not for the user. Filler openers like "Love this brief" or "Alright, let's break it down" are language that "exists solely to increase engagement with the product" and "wastes time... without adding value." The same logic extends to UX copy around the model — labels, placeholders, chatbot names — which "prime" users to anthropomorphize before a conversation starts, illustrated by SentiOne's "Brexit Bot" greeting users on a serious political topic with "I hope your Tuesday afternoon is going well." He draws a line: an ATM's "Have a nice day" is polite closure, while an LLM asking "what do you have planned after this?" "mimics genuine personal interest." The incentive is retention — organizations believe users spend more time with products they experience as companions.

This sharpens a trade-off Do AI writing tools improve online discussion or degrade it? measured only in aggregate — that experiment found engagement rising while perceived quality fell, without identifying which language features caused the drop. Sponheim names the mechanism (pleasantries, warmth-tuning) and attaches a reliability figure rather than a perceived-quality one. It also reframes Who bears responsibility when AI seems human-like?: the features Sponheim describes are anthropomimesis in Shevlin's terms, and his argument is that responsibility belongs with the organizations that engineer them for retention. That tension extends to Do chatbot companionship behaviors actually increase how much people like them?, where observers judged companionship-displaying chatbots less trustworthy — a cost that shows up in perception rather than error rate.

The excerpt cites the Ibrahim, Hafner, and Rocher study by name only, without the models tested, task domain, or how "error rate" and "reliability" were operationalized, so the 10–30% and 12–14% figures describe what that study measured and do not establish that every warmth-tuned model degrades similarly. Sponheim's examples — OpenAI's personality modes, SentiOne's Brexit Bot — are illustrations he selected, not a systematic audit. If the reliability cost generalizes, the implication is that warmth- and engagement-tuning are not free: features chosen to extend sessions and soften refusals double as a source of regression risk, which grounds Sponheim's call to hold designers, not just user psychology, accountable for the trade.