The Testimony Paradox of AI Collaboration

Paul Ricoeur analyzed testimony as a speech act requiring the witness’s attestation—the staking of personal credibility on a claim. The witness who testifies puts something at risk: a history of reliability that can be damaged by error. Machine output lacks this attestation structurally. When a large language model produces a confident summary, it risks nothing, because it possesses no credibility that can be eroded by being wrong. The testimony paradox follows from this asymmetry: the builder who asks the machine for help in a domain she already understands can evaluate the output critically, catch errors, and benefit from the collaboration. The builder who asks for help in a domain she does not understand—which is frequently the entire point of asking—cannot evaluate the output, cannot catch the errors, and is maximally vulnerable to what [YOU] on AI calls “confident wrongness dressed in good prose.” The paradox is not a defect to be engineered away but a structural feature of any