Beyond Sounding Human: Four Better Goals for AI Chat Design
For designers shaping an AI chat interaction, “make it feel real” is too vague to guide useful decisions. A better task is to help someone get a clear answer, explore an idea, understand what the system has done, and choose what happens next. Research on anthropomorphic cues suggests that human-like signals can change how accurate people believe an AI’s information to be. That makes clarity, creative play, reliable task state, and user control more useful design targets than imitation of a human identity.
Why should AI chat aim beyond human likeness?
A conversation can sound fluent without giving the user a reliable picture of the system. In an experiment with 2,165 participants, researchers varied whether a pseudo-language model used text alone or speech with text, and whether it referred to itself as “I” or “the system.” Speech plus text increased participants’ ratings of anthropomorphism and information accuracy; first-person wording raised accuracy ratings and lowered perceived risk in one context. These findings are specific to the study setup, but they show that a cue that makes a system seem more human can also affect judgments of its information. Google Research, “Believing Anthropomorphism: Examining the Role of Anthropomorphic Cues on User Trust in Large Language Models”
That is why anthropomorphic cues and interaction quality deserve separate audits. An anthropomorphic-cue audit asks what the interface implies about the AI’s identity or human qualities. A judgment-risk audit asks whether a cue might lead users to overestimate the accuracy or dependability of its output. The design goals below address a different question: what should a person be able to understand, do, and decide during the exchange? A warm tone can coexist with these goals, provided it does not obscure the system’s capabilities or task state.
Microsoft’s Human-AI eXperience Toolkit organizes guidance around how an AI system behaves at first use, during interaction, when it is wrong, and over time. That structure is useful for chat design because it moves attention from a character’s personality to the changing needs of the person using the system. Microsoft, “Guidelines for Human-AI Interaction”
Make clarity a visible feature of the conversation
Clarity begins with the user’s task, not with a perfectly conversational voice. A person asking for three weekend activity ideas needs options they can compare; someone asking for a rewrite needs a draft they can revise. The interface should make the requested action and the returned result easy to distinguish. If the AI is suggesting, summarizing, or transforming material, label that contribution plainly and preserve enough context for the user to judge it.
A useful interaction can also show what the AI can and cannot do at the point where that information matters. For example, if a chat assistant can suggest a plan but cannot make a booking, the response should not leave the action ambiguous. State what is ready, what still requires the user, and which details remain uncertain. This is a design recommendation inferred from HAX’s emphasis on appropriate behavior throughout an interaction, rather than a claim that one wording pattern works in every product.
A simple clarity check is to ask someone after a short exchange: What did you ask the system to do? What did it actually do? What would you need to check or decide before using the result? If answers diverge from the interface’s intended meaning, revise the wording, structure, or next-step cues before adding more personality.
Use play to support exploration
Playfulness gives people a way to explore a tool and its possibilities. In a study of 372 user-generated posts about ChatGPT, researchers identified playful exchanges that included joking, challenging the system, and testing its boundaries. The authors describe these interactions as a way people explore AI’s capabilities and agency, while noting that the sample reflects posts from an early period of ChatGPT use. Nikghalb and Cheng, “Interrogating AI: Characterizing Emergent Playful Interactions with ChatGPT”
For designers, the practical implication is not to turn every assistant into a comedian. It is to make low-stakes experimentation easy: offer ways to try a different angle, compare alternatives, or revise a draft without losing the earlier version. A short creative prompt such as “Give me three unexpected themes” can invite exploration while keeping the user’s goal in view. Play should be an available mode, not a performance the user must endure.
Research on playful human-AI authoring tools distinguishes playfulness from creativity while describing play as a possible facilitator of creative responses. It discusses interface, AI behavior, and dialogue as design opportunities, while treating playfulness as an experience to support rather than an outcome that can be guaranteed. Liapis et al., “Designing for Playfulness in Human-AI Authoring Tools”
Keep task state legible across turns
A chat interface can feel natural and still lose track of the work. Task state includes the user’s current goal, decisions already made, options still open, and the next step. When the exchange becomes multi-step, a compact recap can show the system’s current understanding: “You have two options left; you prefer a short walk; I haven’t chosen a location yet.” That gives the user a chance to correct the record before the conversation continues.
The interaction pattern should fit the task’s complexity. Ding and Chan’s paper on human-AI text co-creation distinguishes fixed-scope curation, independent creative tasks, and complex, interdependent creative tasks. It maps these to different interaction patterns, from limited intervention to iterative collaboration. The useful design inference is that a one-turn answer may suit a bounded summary, while a developing creative task benefits from turns that expose choices and let the person shape the result. Ding and Chan, “Mapping the Design Space of Interactions in Human-AI Text Co-creation Tasks”
For a practical state check, trace a representative task through the conversation. At each turn, ask whether the person can tell what has been completed, what remains unresolved, and how to correct a mistaken assumption. If the answer relies on remembering several earlier turns, add a concise recap or a visible list of decisions. If the system cannot preserve a detail reliably, make that limitation apparent rather than implying continuity it does not have.
Give users meaningful control
Control means giving the person a usable way to steer or revise the interaction. In chat, this can be as straightforward as letting someone choose between options, edit a generated draft, change a constraint, or ask the system to stop and summarize. These controls are most valuable when they affect a decision the user cares about; a menu of adjustments that changes nothing meaningful adds complexity without agency.
An experiment involving a conversational agent in group discussions compared ways to set expectations about its ability to identify participants who contributed little. It tested information about accuracy, an explanation of the detection approach and data, and an adjustment control. In that particular study, the explanation helped users understand the algorithm and was most effective overall; simply giving an adjustment control led to a more negative perceived discussion experience. The result cautions against treating a control as useful just because it exists: explain what it changes and make the consequence understandable. Do et al., “Inform, Explain, or Control: Techniques to Adjust End-User Performance Expectations for a Conversational Agent Facilitating Group Chat Discussions”
A practical design sequence is: show the current interpretation, offer a small number of meaningful next actions, and make correction easy. For example, after generating an outing plan, the system might ask whether to shorten the route, swap an activity, or keep the draft. Each option names a concrete change. The example is illustrative; it describes a general interaction pattern, not a feature of any particular product.
Turn the goals into a review checklist
When evaluating a chat design, review one ordinary task from beginning to end. Check whether the first exchange makes the system’s role clear; whether a user can distinguish a suggestion from a completed action; whether an exploratory turn invites experimentation without forcing a playful persona; and whether the conversation shows decisions and unresolved steps accurately. Then check whether the person can correct the system and understand the effect of each available control.
Finally, review the language and presentation for cues that may imply human identity or special dependability. Consider whether first-person claims, voice, emotional language, or a human-like avatar could affect a user’s judgment of an answer. That review complements, rather than replaces, testing whether the interaction is understandable and useful. The stronger measure of a chat design is whether a person can pursue their own task with a clear view of the AI’s contribution and the next choice—not whether the exchange passes as a conversation with a person.