Uncanny Valley

Technology is designed to become increasingly human-like. The assumption among designers is that greater realism and human resemblance should lead to greater acceptance. However, users report feelings of unease with near-realistic interfaces, rejection of highly developed chatbots, and avoidance of avatars that appear too human. The question is: When does increasing realism tip into rejection? What factors determine this critical threshold—and what evidence exists about this phenomenon?

Studies

Mori's original hypothesis

Masahiro Mori formulated the Uncanny Valley hypothesis in 1970 as a theoretical concept rather than an empirical study. Working at the Tokyo Institute of Technology on robotics, he observed that while industrial robots were perceived neutrally, nearly human prosthetics triggered discomfort. His core thesis: affinity initially increases with growing human likeness but then drops dramatically into a "valley" at approximately 95% realism before rising again at 100%. Mori illustrated this with examples including prosthetics, corpses, and zombies. Remarkably, this theory remained largely untested for over 30 years yet became the dominant design principle in robotics and animation.

Empirical Validation with Morphed Faces

Karl MacDorman conducted the first systematic experiment on the Uncanny Valley at Indiana University in 2009. A total of 365 participants rated 80 computer-generated faces on scales ranging from "very artificial" to "very human" and from "very eerie" to "very reassuring." The faces represented morphed transitions between obviously synthetic and photographically realistic faces. The results confirmed Mori's curve: at 70–85% realism, likability dropped dramatically by an average of 2.3 points on a 7-point scale. Simultaneously, ratings for "eerie" increased by 1.8 points. The valley was real and measurable, with statistical significance of p < 0.001.

Uncanny Valley in Chatbots and Voice Assistants

In 2016, Maya Mathur and David Reichling at Stanford University investigated whether the effect also occurs with non-visual interfaces. They recruited 1,267 participants to interact online with chatbots of varying sophistication: from obviously scripted bots and GPT-based systems to concealed human operators. After each five-minute conversation, participants rated their trust, discomfort, and intention to use the system. The surprising result: the uncanny valley effect occurred even without a visual component. Bots that seemed "almost human" but occasionally made mistakes were rated 34% worse than those clearly recognizable as bots. The critical zone occurred at 65-80% perceived human-likeness.

Principle

Which principle for Customer Experience Design can be derived from this? The core principle is: deliberately choose stylized over near-realistic and avoid the critical zone before perfection. In customer experience, this means avatars, chatbots, or virtual assistants should either be clearly recognizable as artificial or—if technically feasible—appear absolutely, perfectly human. The Uncanny Valley Effect is particularly relevant at digital touchpoints where customers interact with virtual representatives, as well as in product design and advertising featuring CGI elements. This principle works best when companies clearly understand their target audience's expectations and make a conscious decision to rely on either appealing artificiality or high-fidelity realism. The following guidelines demonstrate how to implement this principle in practice.

Guidelines

Deliberately stylize avatars and bots

Design chatbot avatars, voice assistants, and digital assistants to be deliberately non-realistic. Use cartoon aesthetics, simplified forms, or abstract visualizations. Clearly signal through visual language "I am a bot" rather than attempting to appear human. This approach avoids the critical 70-85% realism zone where user discomfort peaks.

Clearly communicate capabilities

Clearly communicate the bot's capabilities at the beginning of each interaction by explicitly stating what it can and cannot do. For example: "I am a bot and can help with X. For complex questions, I will connect you with a human." This expectation calibration prevents disappointment when the bot doesn't respond in a perfectly human-like manner. Users who know they are interacting with a bot are more tolerant of its limitations.

Enable frictionless handoff to human agents

Provide a one-click option to escalate to a human agent at any time. "Speak to a human" should be prominently accessible without requiring users to navigate through multiple steps. When the bot reaches its limits, proactively offer: "This exceeds my capabilities. Would you like me to connect you with a colleague?" This prevents frustration at the critical moment when users realize the bot cannot fully address their needs.

Disclose AI use, don't conceal it

Label AI-generated content, automated responses, and bot interactions transparently. A simple badge such as "Automatically generated" or "Bot response" provides clarity. Never attempt to pass off AI as human—discovery creates a breach of trust and discomfort. Transparency builds trust and sets realistic expectations.