I will help folks understand how to increase user trust and improve the explainability of AI products, based on exclusive research conducted on AI experiences.
Watch this WaysConf session, then continue with related talks or explore the current programme.
In this WaysConf 2025 presentation, Caleb Sponheim from the Nielsen Norman Group argues that humanizing artificial intelligence is a design trap that hinders product utility. While conversational interfaces and personified chatbots can temporarily boost engagement metrics, they ultimately lead to unmet user expectations regarding empathy, memory, and planning. Furthermore, anthropomorphization prevents users from leveraging the non-human strengths of large language models, such as processing massive datasets or executing complex code. Sponheim outlines actionable strategies to counteract this trend, ranging from simple UX copy adjustments and system prompt constraints to tightly scoping AI features into task-specific tools rather than open-ended chat interfaces. Finally, he advocates for observational and analytics-driven user research methods over self-report data to accurately measure how users interact with AI systems.
Anthropomorphizing AI is not an inherent characteristic of large language models, but rather the result of deliberate decisions made by designers and developers. Companies often introduce human-like personas to artificially boost engagement, which can lead to unhealthy user attachments.
Watch from 4:51When products encourage users to treat AI like a human, users develop expectations for traits the technology cannot deliver, such as genuine empathy, long-term planning, and perfect memory. This misplaced trust can also lead to severe privacy risks when users over-share sensitive personal information.
Watch from 9:57A collaborative study by the University of Waterloo, LSE, and UCL revealed that users who attributed emotional traits to an AI were actually less likely to accept its advice. This indicates that humanizing AI does not guarantee better user trust or decision-making outcomes.
Watch from 22:23The most effective AI implementations are tightly scoped, task-specific features rather than open-ended conversational interfaces. Examples like Amazon's review summaries and Goblin Tools leverage LLM capabilities without inviting unnecessary anthropomorphism.
Watch from 28:14Users are often unaware of their own anthropomorphizing behaviors, making self-report research unreliable. Designers should instead rely on direct observation, analytics, and in-context methods like diary studies paired with objective usage logs.
Watch from 36:49These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.
When we design AI to mimic human conversation, users naturally apply human mental models to the interaction. This leads to a dangerous level of over-trust. For example, a user might share their deepest personal thoughts and secrets with a chatbot, forgetting that they are actually feeding private data to a massive corporation with lenient privacy policies. AI cannot offer genuine empathy or social connection, and encouraging users to treat these systems as confidants exposes them to significant privacy and psychological risks.
Watch from 11:03Large language models possess immense power, but their strengths are fundamentally non-human. They excel at processing massive datasets, generating large-scale code, performing needle-in-a-haystack searches across hours of media, and translating hundreds of languages simultaneously. When we force users into a personified, conversational mental model, we limit their expectations. They stop asking the AI to perform these highly complex, technical tasks because they view the system merely as a human-like conversational partner.
Watch from 13:42Modern "reasoning" models often feature UI elements that display a "thinking" process to the user. This is highly misleading. These models are not thinking or reasoning in any human sense; they are simply generating tokens behind the scenes using test-time compute. There is no proven causal link between these visible "thinking" logs and the final output. Designers expose these logs primarily to keep users occupied during processing delays, which needlessly encourages users to anthropomorphize the machine.
Watch from 19:51Timely academic research on AI is difficult, but a notable study conducted by the University of Waterloo, the London School of Economics, and University College London analyzed 410 participants. The researchers compared emotional attribution with user trust. Surprisingly, they discovered that participants who attributed emotional traits—such as happiness, admiration, or guilt—to an AI system were actually less likely to accept its advice. This proves that humanizing your product does not guarantee better user trust or adoption.
Watch from 22:23The most successful AI user experiences are not open-ended chat boxes, but tightly scoped product features. Amazon's automated review summaries and Goblin Tools are excellent examples of this approach. They leverage the pattern-recognition strengths of LLMs without forcing the user into a conversational loop. By presenting neutral, objective outputs and limiting the length of interactions, designers can deliver the full utility of AI without inviting the complications of humanization.
Watch from 28:14A 2025 study by METR involving experienced software developers highlights the danger of relying on self-report data in AI research. When asked, developers estimated that code generation tools saved them about 25% of their time. However, objective measurements revealed that they actually lost time on programming tasks when using the tools. Because users are often unaware of their actual behaviors and biases, researchers must prioritize direct observation, analytics, and diary studies paired with objective logs over subjective surveys.
Watch from 37:31
