Testimonial chain

A testimonial chain is a sequence of tellings, A to B to C, along which warrant is preserved, so that the last hearer’s knowledge can trace back to the first source. Chains are what make testimony a social technology rather than a private exchange: through them a fact observed once can be known by people who never met the observer. Burge’s account of content preservation, and the transmission views that Lackey has challenged and Wright has examined at length, are in large part accounts of what a link must do.

On my account a link satisfies the Tracking Minimum derivatively when it preserves, without relevant alteration, a particular assertion whose source satisfies it originally. Quotation, recording and faithful relay carry warrant this way. They transmit what they do not originate, because they preserve a particular assertion together with its provenance.

Language-model links do not. Generation does not preserve particulars. The model’s parameters compress its corpus into distributional constraints, and production selects continuations by likelihood under them, so a claim attested in many unreliable sources is favoured over one attested in a few authoritative ones, and nothing retains the address of a ground. Verbatim reproduction of a memorised string does not help, because the system does not mark reproduction as reproduction, so quotation and synthesis are indistinguishable. Chains of model-generated text reposted by users do not preserve warrant as chains of human testimony do. They preserve text.

The claim is structural, not sociological. Downstream of a model, a person who does the work of a competent instrument reader acquires warrant of her own and can testify on her own account; the chain restarts with her, and her hearers are fine. If she has not done that work, the chain fails at her link, for the ordinary reason that a testifier lacked warrant. What has been shown is what the model link itself transmits, which is nothing. That hearers cannot ordinarily tell such a composite chain from a simple one is true, and is why the structural fact matters in practice, but it is a separate claim.