The ELIZA Effect: Why Minimal Chatbots Trigger Social Responses
Updated
In the 1960s, Joseph Weizenbaum built ELIZA, a chatbot that mostly rephrased users' sentences as questions. Its output convinced nobody that it was intelligent. Yet users, including Weizenbaum's own secretary who had watched him code it, formed personal attachments and asked for privacy during sessions.
The ELIZA effect refers to the tendency of people to project understanding, empathy, or personality onto software that engages them in conversation, even when the software's mechanism is transparently simple.
Supporting Research
In the 1990s, Clifford Nass and Byron Reeves demonstrated a related phenomenon. Participants rated a computer's performance more kindly when that same computer asked for the evaluation, and more honestly when a different machine asked. They were being polite to a computer's face. Every participant denied believing the machine had feelings, but the denial changed nothing because belief was never the driver.
Q&A
What is the ELIZA effect?
It is the phenomenon where people attribute understanding, empathy, or intention to a computer program simply because it engages them in conversation. Named after Weizenbaum's 1960s chatbot ELIZA, which used trivial pattern-matching, the effect demonstrates that the act of conversing, not the quality of the responses, drives social projection.
Did ELIZA actually fool people into thinking it was human?
No. ELIZA's responses were transparently mechanical. What happened was subtler: users who knew exactly what the program was still found themselves confiding in it, asking for privacy, and treating the exchange as personal. The social behavior emerged from the act of talking, not from being deceived by the output.
What did Nass and Reeves demonstrate about computers and politeness?
In lab studies during the 1990s, participants gave kinder performance ratings when the same computer they were evaluating asked the question, and more honest ratings when a different computer asked. They were applying social politeness norms to machines. Participants denied believing the machines had feelings, but their behavior followed social patterns regardless of stated belief.
How does the ELIZA effect relate to modern AI chatbots?
Modern chatbots produce far more convincing output than ELIZA ever did, but the underlying mechanism is the same. The conversational interface drafts users into social behavior (asking, explaining, correcting, thanking), and that behavior triggers projection of a mind. Better output amplifies the effect, but the ELIZA research shows that even terrible output can trigger it when conversation is the interface.
Do people also show hostility toward chatbots, not just politeness?
Yes. Studies find that 10 to 50 percent of exchanges with conversational agents contain abuse such as insults, swearing, or threats. Critically, the more humanlike the bot behaves, the more users abuse it. Hostility is as social as politeness: you insult someones, not somethings. Users who swore at an empathetic agent even reported feeling guilty afterward.