Stop Shipping Friendly AI Copy: Why Tool-Like Beats Human-Like

You can ship a lot of "smart" UI this year while quietly making it less trustworthy, less coherent, and harder to use. The mechanism is usually the same: someone in a design review asks for the AI assistant to sound "friendlier," and the team ends up shipping filler instead of help.

Strip fake personality from AI copy. Design the UI to be direct and tool-like instead.

The expectation gap

NN/g's article "Humanizing AI Is a Trap" opens with a striking example: in March 2025, a man collapsed in a parking lot after being misled by Meta's AI chatbot into thinking he would meet a real person. Later that year, OpenAI's CEO said that ChatGPT should be able to act human-like if users want it to.

These two things sit next to each other uncomfortably. The consequence of AI systems successfully presenting as human is that users calibrate their expectations accordingly. When a system sounds warm, confident, and conversational, users expect human-level reliability. When it fails — and it will — the trust gap is much wider than if it had been tool-like from the start.

NN/g puts the core tension plainly: LLMs are "uniquely potent humanizing technologies." Unlike Clippy or Siri, they produce responses that sound like they came from a person. When organisations then layer on personality modes, emotional language, and conversational pleasantries, they amplify a risk that's already baked into the technology. They raise the bar the system cannot consistently clear.

What "friendly" AI copy actually does

The problem isn't warmth per se. The problem is that most "friendly" AI copy is covering up model limitations rather than acknowledging them.

When an AI assistant says "I'd love to help you with that!" before providing an incomplete answer, the friendliness creates a mismatch. The expectation is that a system this personable and enthusiastic knows what it's doing. The reality is a hallucination or a gap.

When a system says "I'm thinking..." while processing, it implies a cognitive process that doesn't exist. The system isn't thinking. It's computing. The language trains users to anthropomorphize a system that doesn't have a mind, which makes the system's failures more surprising and more damaging to trust.

When microcopy says "Great question!" before every response, it wastes the user's attention and signals that the system is optimising for perceived positivity rather than actual utility. Most users figure this out quickly. It erodes trust, not builds it.

The clearer the system says "I am a tool with limits," the better users calibrate their expectations — and the better the actual experience feels when it delivers.

The trust argument for tool-like copy

This is counterintuitive, but it holds up empirically: more restrained, tool-like AI copy produces more trust over time than warm, human-like copy.

The reason is reliability of expectation. When a system presents itself as a capable tool with known constraints, users learn those constraints. They develop a mental model of what the system can and can't do. Within that model, the system is reliable. Users who understand the limits of a tool can use it effectively.

When a system presents itself as a helpful human-like assistant, users import all of their expectations of human assistants. Humans understand context, have common sense, can infer intent, know when they don't know something and say so. LLMs do some of these things some of the time. The gap between human assistant expectations and actual LLM behaviour is a constant source of friction.

The systems that build lasting trust are the ones that are honest about what they are. Not cold — honest. There's a version of tool-like that's still useful and clear. AI-powered interfaces can be direct without being abrupt.

The AI copy red flags checklist

From the newsletter and from the NN/g research, these are the specific patterns to audit:

First-person emotional language. "I'm so glad you asked!" "I love helping with this." "I'm a bit concerned about..." These phrases imply emotions the system doesn't have. Remove them. If the system needs to acknowledge the user, do it functionally: "Here's what I found" not "I'd be happy to find that for you."

Mind-language for compute. "Thinking...", "Let me consider...", "I'm reflecting on this." The system is processing a query. Call it that, or say nothing while it runs. Implying cognition where there's only computation sets the wrong expectation.

Sycophantic validation. "Great question!", "Excellent point!", "You're absolutely right." This pattern emerged because models trained on RLHF with human feedback rewarded agreeableness. It doesn't serve users. It makes the system less credible, not more relatable.

Hedging phrases that imply emotion. "I'm not entirely sure, but..." implies uncertainty felt as an experience. "This answer may not be complete" is more accurate and more useful.

Output that starts with pleasantries. Responses that begin with "Of course!" or "Certainly!" before answering the question delay the answer. The output should start with the answer.

Unexplained limitations. When the system can't do something, it should say so plainly and specifically. "I can't access real-time information" is more useful than "I'm afraid I can't help with that" (which implies reluctance) or silence.

If you want more thinking like this, Unicorn Club is a free weekly newsletter for senior designers and product teams.

The design review conversation

The hardest version of this problem is not the AI copy itself — it's the design review conversation where someone asks for "friendlier" copy.

The request usually comes from a good place: users should feel supported, not lectured. The tool shouldn't feel cold or bureaucratic. These are reasonable things to want.

The reframe is: warmth and trust are not the same thing. A system can be direct, clear, and respectful without pretending to have feelings it doesn't have. The warmth comes from usefulness and honesty, not from "I'd be happy to help!"

The specific language to reach for in that conversation: "Friendly copy raises expectations the system can't always meet. When it fails, the gap is bigger because the promise was bigger. What we want is copy that's direct and honest about limits — that's actually more trustworthy over time."

What tool-like copy looks like in practice

Tool-like doesn't mean cold. It means that language is doing work rather than performing friendliness.

Instead of: "I'd love to help you write that email! Here's what I came up with:"
Write: "Here's a draft based on your notes:"

Instead of: "That's a great question! I'll do my best to answer it."
Write: "Based on [source], here's what I found:"

Instead of: "I'm thinking about this... It seems like the best approach might be..."
Write: "Recommended approach: [specific recommendation]. If [condition], consider [alternative]."

Instead of: "I'm afraid I can't access the internet, so I'm not sure if this is current."
Write: "This answer is based on my training data and may not reflect recent changes. Verify with [specific source]."

The user research on what users actually want from AI assistants is consistent: they want answers that are accurate, clear, and appropriately qualified. They do not want validation. They do not want emotional language. They want the system to do what it says it can do, and to be honest when it can't.

That is the standard. Write to it.