Synthetic hands: round two

A second pass on the weakest part of the concept, the two roles a model may play. It feeds the protocol.

Objections to generative AI in interpretive research

The open letter by Jowsey, Braun, Clarke, Lupton and Fine (objections-01) rejects generative AI for reflexive qualitative research on two grounds: reflexive analysis is a practice that needs a positioned human researcher, and the technology has social and environmental costs. The first ground is aimed at analysis and interpretation, which in Holagram is the Reading register, kept human by rule. The second is categorical and no protocol answers it; it can only be acknowledged and limited. The replies (objections-02, -03, -04, -05) mostly argue against a pro/anti binary and propose AI as a dialogue partner that retrieves passages, offers contrasting views and looks for negative cases, with the researcher interpreting and documenting everything. That is close to the Adversary. The middle-position writers (objections-07, -08) ask for reflexivity about each use and an option to opt out. Empirical studies are thin and small: designers were moved by model suggestions without always noticing (objections-10), logged edits can show that authority stayed human (objections-11), and a model's grading of interpretive claims is unreliable at excerpt level (objections-12). Nothing found tests an AI critic of interpretive claims about places. The practical result is a protocol that states what the synthetic hands do not do, makes every use a recorded choice, and keeps judgement of every synthetic output with a person who can open the source.

What it implies for the protocol

Open questions

Models as critics and adversaries

The evidence says a language-model critic will drift towards agreement unless the design works against it, and that the usual remedies are partial. Sycophancy is measured in free-form advice and moral questions as well as factual ones (critic-01, critic-02), it rises with the user's stated certainty and with first-person framing (critic-03), and it is a default in preference-trained models, not a rare fault (critic-01, critic-18). Instruction alone is the weakest remedy: critic-03 reports that rewriting a statement as a question beats a plain 'do not be sycophantic' instruction, critic-04 reports that a third-person framing cut sycophancy by up to 63.8 per cent in its debate scenario (metric to be checked), and critic-01 reports that existing mitigations were of limited effect. Judge-style bias is the second problem: models favour their own text (critic-06, critic-07), position and other cues (critic-08, critic-09). The best-supported fix there is structural: keep the author's model out of the judge seat and use a panel from different families (critic-07 reports more than a 50 per cent cut in one hiring setting). On debate, the picture is less favourable than round one suggested: with answer keys, debate helps a judge (critic-13), but matched comparisons find that most of the gain comes from voting over several samples (critic-10), that debate can lower accuracy through conformity (critic-11), and that model diversity helps (critic-12). Role-played opposition does not do the work of sincere dissent (critic-05). The closest supports for Holagram's Adversary are a critic that must produce checkable specifics (critic-14, critic-15) and a bounded role, counter-examples for a human to judge (critic-16, and critic-17's call for provocation).

What it implies for the protocol

Open questions

AI dialogue and belief, and the Science paper

The evidence on AI dialogue changing beliefs is the best-funded part of this review and the most unsettled in status. The anchor paper (science-01) reported that GPT-4 Turbo arguing against a believer's own conspiracy theory cut belief by about 20%. On 11 June 2026 Science attached an Editorial Expression of Concern (science-02): the authors found that screening criteria in the manuscript did not match the published analysis pipeline and that the public dataset held extraneous spliced rows. They supplied raw data and updated results, said to match the original in direction, significance and size; Science was still evaluating them. The paper is flagged, not retracted; the authors published correction materials on 27 September 2026 (science-04); Science's formal outcome and any replication paper were not found. Nyhan reports two independent reproductions (science-03, second-hand). The wider literature says three things that matter more for Holagram than the single paper. First, what persuades is information and conversation, not warmth or perceived humanness (science-05, science-08, science-09, science-07), and the same machinery persuades toward falsehood about as well (science-06, science-11), with accuracy falling as persuasiveness rises (science-08). Second, agreeable models make people surer they are right and prefer them for it (science-12), and simulations suggest neither accuracy nor warnings stops a spiral (science-13); chat-log work shows the pattern in practice (science-14). Third, repeated dialogue may lower unaided judgement (science-15). For Holagram the lesson is that a synthetic Adversary is a persuasion engine whose effect cannot be told from correctness by how convincing it sounds. Its output is a lead, never a Trace; the researcher's own Test comes first; the structure around the model, not the model's good behaviour, must carry the load. No study was found testing an LLM critic against a researcher's own interpretive readings.

What it implies for the protocol

Open questions

The Alter role: model-written fiction

The literature on model-written fiction is consistent on one point and thin on another. It is consistent that fiction written by models converges: one model gives the same plot across settings (alter-01), the same plot elements across runs (alter-02), a shared region of narrative space across models (alter-04), and tidier, more explicit, less ambiguous stories than people write (alter-04). Prompting tricks and temperature do not fix it (alter-03). Co-writing with a feedback-tuned model narrows what people write (alter-05). For Alter this means a single Fiction is the likely version of the world, and its value lies in what the human edit changes. It also means that ten Fictions from one model are evidence of the model, not of the place. The second point is leakage. A working preprint (alter-08, self-published, not peer reviewed) reports that a model continues inside an elaborate fiction without correcting it and that the fiction can absorb real findings as apparent confirmation; this is the Ballardian narrator's failure with a machine supplying the material. Sycophancy work extended to implicit premises (alter-09) says the Adversary will protect the author's face, not only agree with stated claims. On the positive side, fiction has a respectable standing as a research tool (alter-11, alter-12), and models have been used to produce scenarios that start debate in teams, with the authors insisting on critical use (alter-13). Labelling is the weakest part: an AI label lowers trust in true and false items alike because readers assume full automation (alter-15), and detailed disclosure lowers trust while raising source-checking (alter-16), so the label must say what the human did. No source tests a model-written fiction against a researcher's own Trace and Reading, or a model critic against interpretive claims about a place. The protocol below is therefore a design, not a validated method.

What it implies for the protocol

Open questions