The ELIZA effect refers to the cognitive bias whereby users overestimate the capabilities of a conversational system and attribute to it an understanding, intelligence or awareness that it does not actually possess. This phenomenon manifests itself in the tendency to project intentionality, empathy and reasoning onto a computer programme capable of producing coherent and contextually appropriate responses, independently of the actual mechanisms underlying its operation.

Origin

The term takes its name from ELIZA, a natural language processing programme created by Joseph Weizenbaum at MIT between 1964 and 1966. ELIZA simulated a Rogerian psychotherapist by reformulating the user's words in the form of questions, using simple syntactic rules and keyword recognition. Although ELIZA had no semantic understanding and simply manipulated linguistic patterns, Weizenbaum was amazed to find that its users - including his own secretary and colleagues who knew how it worked - developed emotional relationships with the programme. Some confided intimate thoughts to it and asked to be left alone with the machine, convinced that it truly «understood» them and showed a form of empathy.

Psychological mechanism

The ELIZA effect is based on several converging cognitive mechanisms: our tendency towards anthropomorphism (attributing human characteristics to non-human objects), our difficulty in accepting the absence of intentionality behind apparently coherent behaviour, and our propensity to fill information gaps with inferences. Natural language acts as a particularly powerful catalyst: as soon as a system produces sentences that are grammatically correct and contextually relevant, our mind automatically infers the existence of an underlying understanding.

Contemporary relevance

The ELIZA effect remains particularly relevant in the face of modern conversational assistants and generative language models (LLMs). The increasing sophistication of these systems - their ability to nuance, contextualise, rephrase and even mimic empathy - amplifies the illusion of understanding. When an AI generates a structured analysis or an emotionally appropriate response, users naturally project the existence of thought, sensitivity or awareness, whereas these productions are the result of statistical optimisation on massive corpora of texts. The fluidity of language masks the underlying probabilistic mechanics, making the ELIZA effect more insidious than ever.

Implications

Recognising the ELIZA effect is crucial to maintaining a critical attitude towards AI systems: it encourages us to distinguish between linguistic performance and real understanding, apparent coherence and authentic reasoning. This lucidity helps us to avoid over-interpreting the capabilities of systems, to better calibrate the trust we place in them, and to design interfaces that do not exploit this bias to create artificial emotional dependency.