Amorphous, imitative and shapeshifting: Ditto, the digital twin system investigating AI citation and source verification

AI generated responses often appear reasoned when their claims are not supported by the sources they cite. Gettier (1963) argued even a justified and true belief does not amount to knowledge if the justification and truth align only by accident. Currently, generative AI produces great amounts of text that possess this issue, where responses with reasoning backed by valid sources cannot be differentiated from responses based solely on internalised data, and lack fact-checking from valid external sources (Ji et al., 2023; Rozear & Park, 2023). At the current scale at which AI participates in human fields such as research, medicine, law and education, every answer where the system's understanding of a question differs slightly from the asker's intent is carried forward and multiplied, whilst appearing fluent and supported. This paper raises philosophical attention to AI related fields, and presents a tool which allows the difference between reasoned and unreasoned responses to be detected. Ditto is a digital twin system for individual scholars (currently Kant and Mill, drawing upon approximately a thousand passages of their published works), which only answers questions when its claims can be checked against the scholars' own texts. Every response falls into one of four categories: direct answers, which are derived directly from the selected sources; extrapolations, which reason beyond the texts and are labelled as such; refusals, which are answers unclaimed as the scholar's texts do not align with the question topic; and withheld answers, which attempt a relevant answer three times before reporting a failure. The findings of this paper are dual. Refused and withheld questions are returned to the asker, making this type of failure visible and manageable for the user rather than silently displayed. Across the 41 graded answers of the second evaluation no false information was provided, and every unsupported item in a controlled probe was blocked (3 of 3 attempts in three different trials). One incorrect answer did reach the reader in the earlier 48 answer evaluation, and the pattern that produced it was blocked afterwards; six further answers in that evaluation were shown carrying minor flagged defects. However, in direct and extrapolated responses, the system acts as if it has understood the question, which is where risks may compound: Ditto with verification disabled confidently produced unsupported answers in 22 of 27 responses, whereas the digital twin with verification enabled produced fully verified answers in 12 of 14 responses, and held back responses to 13 questions due to a lack of valid evidence. Thus, the issues of whether an answer is 'true' and whether the question was understood as it had been intended are where human judgement cannot be automated.