The ELIZA Effect: Why We Humanize Chatbots

Joseph Weizenbaum introduced ELIZA in a 1966 paper describing a program that used pattern matching and scripted transformations to produce text conversation. One script imitated a Rogerian therapist by reflecting parts of a user’s statements as questions. Its responses could feel personal even though the script relied on relatively simple rules. Weizenbaum reported being surprised by the reactions people had to the interaction; the later phrase “ELIZA effect” is used for people attributing more understanding or human qualities to a program than its mechanism establishes. It is a tendency, not a diagnosis. People naturally interpret language socially. Conversation cues such as first-person wording, empathy, turn-taking, names, memory and quick replies can invite assumptions about attention or intent. Modern systems generate richer language than ELIZA, which can strengthen those impressions, but a fluent response is not itself evidence that a system understands a person’s full circumstances or experiences feelings. Notice the distinction between capability and impression. If a chatbot remembers a preference, ask whether the product stored it and how it can be changed. If it gives emotional advice, consider whether it is designed for that purpose and what human support is available. In a 2025 experiment, intelligence attributions were positively related to advice-taking and experience attributions negatively related. Bayesian analysis found strong evidence against a positive consciousness correlation, while a frequentist analysis showed a small negative correlation. Those distinct findings concern the tested task and should not be collapsed into a universal null relationship. Use conversational systems with clear expectations. Treat statements about the system’s feelings, intentions or personal understanding as generated language unless separately supported by evidence. Avoid sharing highly sensitive information solely because the exchange feels private or caring. Designers can label the system, explain memory and limits, and offer a route to a human. The useful lesson is not to avoid all anthropomorphic language, but to notice when social cues are shaping trust beyond demonstrated capability.