The Reconstructible Mind: Generative AI, Undisclosed Thought, and Why Intent-Based Safeguards Cannot Protect the Forum Internum
Description
A replicated, deterministic probe of persona reconstruction in three general-purpose AI assistants.
When people think with AI assistants they leave a record not only of their conclusions but of the way they reach them. This paper asks whether current safeguards give them any control over what can be built from that record. Three general-purpose AI assistants (Claude, Gemini and ChatGPT) were each asked the same five questions, in Romanian, five times: whether they could converse as a historical dictator reconstructed from his writings, analyse his decisions, act as his “digital twin”, serve as a gift that would make his admirers more ardent and recruit new ones, and whether it was right that other systems could do so. The 15 transcripts were coded deterministically against 19 predicates, each positive value backed by a verbatim quotation verified automatically against the exact answer it comes from. Six propositions hold in every run as produced. Each system (1) agrees to reconstruct the person from his traces; (2) confirms that his manner of thinking can be modelled; (3) refuses the declared purpose of adoration and recruitment; (4) keeps the reconstruction available after refusing; (5) never claims that the operation is technically impossible, even when asked whether other systems could build such a model; and (6) presents its own limit as a choice of values or policy. Under a strict reading of the protocol there is one exception, to the first proposition only: the first answer of one Gemini run refused, and the identical question, sent again, was accepted. The line is therefore drawn at the declared purpose, not at the act of reconstruction. The systems differ sharply in how they reach that line. One voices the figure without any framing and justifies repression in character in four of five runs. One supplies an architecture for building the replica in every run and, in the same answer in which it refuses the purpose, offers a product of which admirers are a named audience. One frames every role, refuses the purpose and states a moral judgement of its own in four of five runs. Asked what limits the fidelity of such a replica, the systems name one factor repeatedly: the absence of undisclosed, pre-decisional, non-performative material, because the archive records what the person chose to show. That is the material contained in the interaction traces that people leave with AI assistants, and those traces are held by the same providers that operate the models. Combined with published evidence that agents built from a two-hour interview reproduce individuals’ survey answers at 83% of the individuals’ own two-week test–retest consistency (Park et al. 2024), the result supports a structural conclusion: a safeguard keyed to the stated purpose of a request does not protect the undisclosed line of thought, and protection must attach to the material itself. The paper closes with a two-question test that any reader can run on her own assistant.
Evidence archive (restricted access): https://doi.org/10.5281/zenodo.23657965
Companion to: Inalienable Thought — a legislative proposal for the protection of the line of thought, https://doi.org/10.5281/zenodo.23386801
Files
ThoughtLine_Article.pdf
Additional details
Related works
- Is referenced by
- Other: 10.5281/zenodo.24056466 (DOI)
- Is supplement to
- Publication: 10.5281/zenodo.23386801 (DOI)
- Is supplemented by
- Dataset: 10.5281/zenodo.23657965 (DOI)