ELIZA's creator argued some jobs should stay human

The design

Joseph Weizenbaum, the MIT computer scientist who built ELIZA, published a book in 1976 arguing that some tasks ought not be delegated to computers regardless of whether a machine could technically do them. The Internet Archive's catalogue record for Computer Power and Human Reason confirms it as a 300-page W. H. Freeman title. Its actual arguments survive here through a rarer document: a November 1976 Stanford Artificial Intelligence Laboratory memo reprinting three published reviews of the book, by Bruce Buchanan, Joshua Lederberg and John McCarthy, plus Weizenbaum's own reply to McCarthy, quoting his book directly by page number.

What the evidence says

Buchanan's review, distinct from Weizenbaum's own claims, reduces the book to a formal argument: computers can never fully understand human problems, therefore it is wrong to command them to handle affairs requiring empathy, a premise Buchanan traces to Weizenbaum's own line, quoted with its page number, that 'an individual is dehumanized whenever he is treated as less than a whole person' (p. 266). McCarthy's review, independently, revisits Weizenbaum's own account of ELIZA and of Colby's related PARRY test, concluding that both show only that people, including psychiatrists, 'are therefore easily fooled.' Weizenbaum's reply confirms what the book actually opposes: not machine intelligence in the abstract, but 'machine-administered psychotherapy' specifically, which he writes he fears would 'induce just this kind of self-image' of a patient reduced to an object.

What it asks of people

The book, as its reviewers and Weizenbaum's own reply describe it, asks clinicians and AI researchers to draw a line between what a computer can do and what it should be trusted to do, specifically resisting proposals, which Colby's own work had raised, to let programs conduct therapy at scale. It asks a patient in such a system to accept being addressed by a program that Weizenbaum argues cannot share human experience, regardless of how fluent it sounds.

Privacy and safeguards

Neither the book's catalogue record nor the reprinted reviews describe any data-handling practice; this is a 1976 philosophical argument, not a technical specification, and it predates any framework for what a machine-therapy system would retain or disclose. What it does establish is an early, named objection to letting a program stand in for a therapist at all, framed as a risk to a patient's sense of being treated as a whole person rather than a data-handling risk as later understood.

• Does Weizenbaum's objection rest on what computers cannot do, or on what society should not ask them to do?
• How does McCarthy's account of the ELIZA and PARRY illusions change how a reader should weigh either program's apparent success?
• What would Weizenbaum's 1976 argument say about a companion app that a user treats as a confidant today?

The book's reception, preserved here through its critics rather than the text itself, shows an argument already contested by other AI researchers the year it appeared, not a settled verdict either side won.