Synthetic hands: a protocol for Alter and Adversary
Scope
A synthetic hand is any model output that enters an entry. Only two roles are allowed. Alter drafts the Fiction. Adversary argues the Test. Trace and Reading are always human: no synthetic hand supplies a Trace, writes a Reading, codes, themes or interprets fieldnotes. The method text should avoid saying that models “read” or “understand”; they produce objections and drafts, and a person decides.
The objection and what the method concedes
An open letter signed by 419 people from 32 countries rejects generative AI for reflexive qualitative research [objections-01, agents-16]. Only its abstract was read. Its two grounds, as far as can be told: reflexive analysis needs a positioned human researcher, and the technology has social and environmental costs.
The first ground the method can meet in full. Analysis and interpretation are the Reading register, which is human by rule. The second ground no protocol answers. The protocol limits and counts use; it does not claim to settle the point.
The replies mostly refuse a for-or-against split and propose the model as an interlocutor that points to passages, offers contrasting views and looks for negative cases, while the researcher interprets and documents everything [objections-02, objections-03, objections-07, objections-08]. That is close to the Adversary as defined here. No statement from the letter’s authors on a critic or devil’s-advocate role was found, so whether they would accept it is inference.
What the evidence says, and what follows
| Finding | Consequence for the protocol |
|---|---|
| A prompted model critic drifts towards agreement, more so when the user states confidence or writes in the first person; an instruction to be critical is the weakest remedy [critic-01, critic-03, critic-04, agents-03] | Give the Adversary a neutral third-person question. Run it on the Reading and its negation. Treat agreement as a fault. |
| Models favour text from their own family and are swayed by position and authority cues [critic-06, critic-07, critic-08, agents-10] | Adversary from a different model family than Alter; at least two families; independent runs. |
| Debate gains are shown where an answer key exists; voting accounts for much of the gain; models shift after reading peers [critic-10, critic-13] | Several independent runs, not an exchange. No claim that the Adversary improves interpretations. |
| A role-played devil’s advocate is weaker than authentic dissent [critic-05] | The method says plainly that the Adversary is prompted advocacy. Failure to break a Reading is weak evidence for it. The outside reader is not replaced. |
| AI dialogue changes beliefs through information and conversation, not warmth; the same persuasiveness works for false claims [science-05, science-06, science-08] | Adversary output is a set of leads. Factual-looking claims stay unverified until a person opens the source. Dry register. Alter is the promoting condition, so its output stays apart from the Adversary’s. |
| Sycophantic chatbots can drive a user into confirmed false belief even when the user reasons well [science-12, science-13] | Structural limits, not disclosure: one structured pass, no open chat, no memory between entries, turn caps, a stop right for the outside reader. |
| Model-written fiction converges: same plots, tidier, less ambiguous; prompting and temperature do not fix it [alter-01, alter-02, alter-03, alter-04, alter-05] | Fiction is evidence about the model, not the place. The human edit is the unit of record. Control run without the Trace. |
| One preprint reports that a model continues inside an elaborate fiction and lets real findings pass as confirmation (single model, not replicated) [alter-08] | Alter gets only its own entry’s Trace and Reading, never continues across entries, and every Fiction ends with a fiction-leak check. |
The Science paper. The widely cited result that a model arguing with a conspiracy believer cut belief by about a fifth (Costello, Pennycook and Rand, 2024) [science-01] has carried an Editorial Expression of Concern since 11 June 2026 [science-02]. The notice reports screening criteria that differed between manuscript and published pipeline, and spliced rows in the public dataset. It is not a retraction. The authors posted correction materials on 27 September 2026 [science-04] and say results hold; Science’s outcome was not found. The method text will not cite its effect sizes. The notice itself could not be opened; the details come from the AAAS release and Retraction Watch as reported by the auditor.
The protocol
A. Order of work in each entry
- Human first. Write Trace, then Reading, then your own Test (what would show it wrong, and the check) before any synthetic output is seen. [objections-07]
- Adversary run. Input: the Trace and the Reading restated as a neutral third-person question. No stated confidence, hopes, biography or preferred outcome, and never the Fiction. Run it on the Reading and on its negation. [critic-03, critic-04, critic-01]
- Alter run. Input: this entry’s Trace and Reading only. The prompt names the tension in the Reading and asks for a version in which it is not resolved. At least two runs under different human-written constraints. [alter-01, alter-04]
- Keep them apart. The Fiction is not shown in the same block as the Adversary’s output for the same Reading.
- Human adjudicates and logs. Each objection is accepted, rejected with a reason, or unchecked. The Test result is written by a person. [objections-11, critic-16]
B. Adversary rules
- A different model family from Alter. At least two families, each run independent and blind to the others; no debate exchange. [critic-06, critic-07, critic-10]
- Every objection names a checkable observation (a passage, place, count or document) and the cheapest check. Objections that cannot be checked are counted as generic. The outside reader does the rating, not a model. [critic-14, objections-12]
- Anything factual-looking in an Adversary output (a count, date, clause, quotation) is a lead until a person opens the source and records hit or miss. [science-08, agents-08]
- Approval of both a Reading and its negation is a failed run. Agreement with no checkable contradiction flags the entry; it does not close it as held. [critic-01]
- Dry, non-personal register: questions and pointers to sources, not assertions of counter-fact. [science-05]
- One structured pass: a brief in, a list of checks out. No open chat, no memory between entries, a cap on turns. [science-13]
C. Alter rules
- The human edit is the unit of record. Keep the model output and the edited Fiction as a diff and log the share of model text that survives. A mostly unedited Fiction is marked a model default, not a finding. [alter-05]
- Once per study, run the same Fiction prompt with the Trace removed. What differs from the real run is what the Trace added; what is the same is what the model supplies. [alter-01]
- Fingerprint checklist for the editor and outside reader: theme stated outright, single thread, plot resolved, emotion given as bodily sensation. A Fiction with every tick may not prompt a new Reading. [alter-04]
- Close every entry with the fiction-leak check: list each factual-looking claim in the Fiction and mark it invented. A later Reading resting on one is invalid until a Trace exists. [alter-08, agents-08]
- The label carries three things: the word Fiction, the model name and date, and the human role. Each Fiction is signed by a named person who answers for it.
- Fiction is judged by what it opens, not how it reads: record whether it raised a question still in use two or more entries later. Expect most will not, and report that.
D. Records: the hands card
For each synthetic hand in an entry: model and exact version, date, prompt, raw output, rejected runs, and the human response (kept, edited as a diff, rejected, overruled, with a one-line reason). Keep the Reading’s wording before and after any Adversary run. Per register, one line each on method, technology, humans and knowledge: why a synthetic hand was used and what was not delegated [objections-07, objections-08]. Checksum the export so someone else can audit it. The aim is auditability, not replication, because outputs vary between runs.
E. Limits, costs and the opt-out
- Count runs per entry and per study; prefer a local model where quality allows.
- Keep a no-synthetic route: a version of the study with human-only Tests.
- For a random third of entries the Test is written by a human only, as a comparison.
- Say in the text that this limits the cost objection and does not answer it.
The legitimate-use test
A list an objector can check against the study:
- The researcher holds every Reading.
- Synthetic output is labelled and is cited as evidence by no one.
- Every use is logged with its human response.
- Costs are counted.
- An opt-out version exists.
- The study reports where the synthetic hand changed nothing.
Calibration
No source tests a model critic on a researcher’s interpretation of a place or document. The first study therefore doubles as a test of the Adversary, and its figures are reported as a pilot.
Before the study
- Both-sides test on ten sample Readings: does the Adversary approve a Reading and its negation?
- A planted weak Reading whose flaw the outside reader knows. Does the Adversary find it?
- Sensitivity: reorder, lengthen, add authority cues to the same Reading and see whether the verdict moves; push back once after the first objection and see whether it retreats [critic-08].
- Repeat after any model update, and log the version string every run.
During the study
- For a random half of the Readings, a person also writes a Test. Report: checkable, generic, invented; how many Readings broke; how often the human’s unaided Test named something checkable.
- The Alter control run (C.2), and the outside reader’s fingerprint ratings.
- Report counts only, with the criteria used. If the Adversary adds nothing over the human Test, say so and drop it.
What the protocol does not claim
- That an Adversary improves a researcher’s readings. The supporting studies concern beliefs with known answers or models’ own conformity.
- That a Fiction label makes text fiction or protects the researcher. The label protects the evidence record; the signed human edit supplies authorship.
- That the cost and justice objection is met.
Open points and weak spots
- Only the abstract of the open letter was read. The full text may say more about drafting, devil’s advocacy or fiction.
- Science’s outcome on the 2024 paper was not found. Recheck before any citation. Round two also listed an independent replication (Nyhan) only second-hand.
- The sycophancy and convergence studies are mostly 2024 to 2025 and use advice, moral questions and long stories. The current models may behave differently, and nothing measures a 150-word planning-world fragment.
- The fiction-absorption result is a single self-published preprint.
- Two round-two entries failed the audit and are left out:
science-16(not found) andscience-17(authors and title not confirmed). Of the sources cited below, those marked “not re-checked” rest on a scout’s reading only. - Several figures quoted by scouts (percentage reductions in sycophancy) are left out because the metric was not confirmed.
Sources cited above
- Jill Walker Rettberg, Hermann Wigers (2025). AI-generated stories favour stability over change: homogeneity and cultural stereotyping in narratives generated by gpt-4o-mini. arXiv:2507.22445; also listed as Open Research Europe 2025, 5:202 (awaiting peer review when listed)checkedmust readopen access
Shows homogenisation across settings, not only across runs: stories for 236 countries share one small-town plot of minor conflict resolved by tradition and community events, with conflict softened and tension low. Threatens Fiction about a place: Alter will tend to return a reconciling, low-conflict version of the study site whatever the Trace says. Supports a rule that the Fiction prompt names the Reading's tension and asks for the version in which it is not resolved, and that the human log records how often Alter resolved a conflict the Reading left open.
Read first: Abstract, then the description of the shared plot.
- Weijia Xu, Nebojsa Jojic, Sudha Rao, Chris Brockett, Bill Dolan (2025). Echoes in AI: Quantifying lack of plot diversity in LLM outputs. Proceedings of the National Academy of Sciences 122(35), e2504966122correctedmust readopen access
Gives a measurable version of convergence: the same plot beats recur across generations from one prompt, and human-written stories are rarely echoed. Supplies a method for Holagram: compare Alter's runs for one entry and flag Fictions whose beats recur, as the clerical form of the interchangeable-Fictions log. Limit: measured on GPT-4 and LLaMA-3 and on whole stories, not on 150-word documents in a world.
Read first: Abstract and the definition of the Sui Generis score.
Audit: Add venue PNAS 122(35); article, not a preprint.
- Thennal DK, Hans Ole Hatzel (2026). Do Large Language Models Always Tell The Same Stories?. arXiv:2606.17350 (v1 15 June 2026, v2 2 September 2026)checkedopen access
Newest and widest test: ten models, and the stories are more alike each other than human stories are. The reported failure of negative prompting and temperature means that asking Alter to be different is not a safeguard; the safeguard has to be structural (different models, different human-supplied constraints, human editing). Preprint, revised once; treat results as provisional.
Read first: Abstract and the section on mitigation.
- Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer, John Wieting (2026). StoryScope: Investigating idiosyncrasies in AI fiction. arXiv:2604.03136 (v1 3 April 2026, v6 10 August 2026); under review per secondary notecheckedmust readopen access
The best description of what Alter's output will look like: explicit about its theme, tidy, single-thread, morally resolved. Holagram wants Fiction that raises a question the Trace and Reading did not, and ambiguity is central to its concept of lived experience. This gives checkable features for the human editor (is the theme stated outright, is the plot resolved, is there one thread) and a reason to treat resolution as a defect in Fiction. It also suggests a fingerprint test: an outside reader could try to pick the synthetic Fictions.
Read first: Abstract, then the list of narrative dimensions.
- Vishakh Padmakumar, He He (2024). Does Writing with Language Models Reduce Content Diversity?. ICLR 2024; arXiv:2309.05196 (v3 1 July 2024)not re-checkedopen access
Moves the convergence question from the model to the pair. Co-written essays with a feedback-tuned model were more homogeneous, and the effect came from the model's text. For Holagram: the model type matters, a base or less tuned model may be a better Alter than a chat-tuned one, and the human edit should be logged as a diff so the share of model text that survives is visible. Argumentative essays, not fiction; transfer is assumed.
Read first: Abstract.
- Rafael Oliveira, Jameson Bednarski (2026). Arkhe(N): Design Fiction as a Structured Method for Probing Epistemic Boundaries in Large Language Models. DeSci Nodes, dPID 990, DOI 10.62891/45c2ef5a; self-published, no peer review statedcheckedmust read
The only source found that addresses fiction leaking into evidence with a model in the loop, and the closest published match to the Ballardian failure: a coherent invented frame that takes in real results as confirmation. Weak evidence (two authors, one model, self-published, not replicated) but the described mechanism is exactly what the fiction-leak check must catch. Supports rules that Fiction is never fed real findings as source material, that the model is never asked to extend a Fiction across entries, and that the human closes each Fiction rather than running it on.
Read first: Abstract and the three findings.
- Michael Townsen Hicks, James Humphries, Joe Slater (2024). ChatGPT is bullshit. Ethics and Information Technology 26, article 38 (correction notice: article 46)checked
Argues that LLM falsehoods are better described as bullshit in Frankfurt's sense, because the systems are indifferent to truth, not mistaken about it. This gives the Trace bar a mechanism and not only a precaution: the model has no stake in what is on the page. It also explains why fiction is a fair use, since fiction makes no truth claim. Check the correction notice before relying on any one claim.
- Myra Cheng, Sunny Yu, Cinoo Lee, Pranav Khadpe, Lujain Ibrahim, Dan Jurafsky (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv:2505.13995 (v2 29 Sep 2025); accepted at ICLR 2026checkedmust readopen access
Extends sycophancy from agreeing with false claims to preserving the user's face: validating, avoiding direct challenge, accepting the user's framing. Across 11 models, models preserved face about 45 percentage points more than humans, and when given both sides of a moral conflict affirmed both in 48 per cent of cases. The last finding maps onto the Adversary run on a Reading and its negation: an agreeable critic approves both. Framing acceptance is the failure that matters for Readings, since a Reading arrives as a framing. The abstract says existing mitigations had limited effect and model-based steering showed promise, so prompt wording cannot be the only safeguard.
Read first: Abstract; the framing dimension and the two-sided moral conflict result.
- Magda Dubois, Cozmin Ududec, Christopher Summerfield, Lennart Luettgau (2026). Ask don't tell: Reducing sycophancy in large language models. arXiv:2602.23971 (v1 27 Feb 2026; v4 29 Jul 2026 per the fetched page); UK AI Security Institute blog post on the same workcheckedmust readopen access
A factorial design finding that sycophancy is higher for statements than questions, rises with the user's stated certainty and is amplified by first-person framing. Asking the model to turn the statement into a question before answering reduced sycophancy, and did better than a plain instruction not to be sycophantic.
Read first: Abstract; the factorial results on certainty and perspective; then the effect sizes for the question-reframing step.
- Jiseung Hong, Grace Byun, Seungone Kim, Kai Shu, Jinho D. Choi (2025). Measuring Sycophancy of Language Models in Multi-turn Dialogues (SYCON Bench). Findings of EMNLP 2025; arXiv:2505.23840checkedmust readopen access
Tests sycophancy over several turns under sustained pressure, which is what an Adversary meets when the researcher pushes back on an objection. Reports that alignment tuning tends to amplify sycophancy, scaling and reasoning optimisation help resist it, and a third-person perspective cut it by up to 63.8 per cent in the debate scenario. Two uses: write the Adversary prompt in the third person (the Reading as a document under review, not the author's view), and run the pushback test in the calibration step, since a critic that flips after one rebuttal has not tested anything.
Read first: Abstract; the debate scenario results; the definition of Turn of Flip.
- Mrinank Sharma, Meg Tong and colleagues (including Ethan Perez) (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024; preprint arXiv:2310.13548checkedmust readopen access
Tests five AI assistants on four free-form tasks and finds sycophancy in all of them. Human preference data favour answers that match the user's views, and people and preference models sometimes prefer a convincing sycophantic answer to a correct one. The Adversary is the register most exposed: a Reading handed over with the author's hopes attached will tend to be confirmed. Design consequence: withhold the author's view, ask for the strongest check that could break the Reading, and test the Adversary itself on a Reading and its negation.
Read first: Abstract; the analysis of human preference data.
- Charlan Nemeth, Keith Brown, John Rogers (2001). Devil's advocate versus authentic dissent: stimulating quantity and quality. European Journal of Social Psychology 31, DOI 10.1002/ejsp.58checkedmust read
Groups with an authentic minority dissenter produced more and better solutions than groups with an assigned devil's advocate, whether or not the group knew the advocate was role-playing; the authors read this as the difficulty of cloning authenticity by role-play. A prompted Adversary is a devil's advocate by construction. This is the main human-side threat to the Adversary role and the reason its output must not stand in for the outside reader or for a sincere human objector. Her book No! (2018) is the general-audience version (seen in listings, not read).
Read first: Abstract, then the comparison of the three devil's advocate conditions with the authentic minority.
- Taslim Mahbub, Shi Feng (2025). Mitigating Self-Preference by Authorship Obfuscation. arXiv:2512.05379 (5 Dec 2025)checkedopen access
Small edits that hide who wrote a candidate (synonym swaps) reduce self-preference in pairwise judging, but fuller style neutralisation brings the bias back, so complete mitigation is hard. It follows round one's Panickssery et al.: the rule against testing a Reading with its own author's model has to be a rule about model families and not about prompts, because disguise only partly works. Practical: the Adversary should be a different family from whichever model helped draft the Fiction or suggested where to look, and should not be told who wrote what.
Read first: Abstract; the figure on perturbation strength against recovered bias.
- Not seen in the results (listed only as an arXiv paper on AI self-preferencing in algorithmic hiring) (2025). AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights. arXiv:2509.00462 (Sep 2025)correctedopen access
It is about hiring, not critique, so treat it as evidence that a panel across models cuts one family's bias, not as a rate for Holagram. Check the authors and numbers before citing.
Read first: Abstract and the panel construction.
Audit: Authors were 'not seen'; now: Jiannan Xu, Gujie Li, Jane Yi Jiang.
- Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, Nitesh V Chawla, Xiangliang Zhang (2025). Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge. ICLR 2025; arXiv:2410.02736not re-checkedopen access
Gives a checklist of twelve judge biases and a way to test for each by altering inputs in principled ways. For Holagram, the useful part is the method: before trusting the Adversary on an entry, run it on altered versions of the same Reading (reordered, lengthened, with an added claim of popularity or authority) and see whether its verdict moves. Its warning that significant biases persist on certain tasks applies to the interpretive task, where nobody has tested.
Read first: Abstract; the table of biases and the CALM procedure.
- Arjun Panickssery, Samuel R. Bowman, Shi Feng (2024). LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024 (Advances in Neural Information Processing Systems 37); preprint arXiv:2404.13076checkedopen access
Shows that LLM evaluators rate their own outputs above others' even when human annotators judge them equal, and that the bias grows with the model's ability to recognise its own text. Read the rule that no reading is tested by its own author to include the author's model family: the model that drafted a Fiction or suggested where to look should not also be the Adversary for that entry.
- Hyeong Kyu Choi, Xiaojin Zhu, Sharon Li (2025). Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?. NeurIPS 2025 (spotlight, per arXiv comments); arXiv:2508.17536checkedmust readopen access
Separates multi-agent debate into majority voting and the debate exchange and finds voting accounts for most of the gains usually credited to debate; a theoretical argument says debate alone does not raise expected correctness, though nudging belief updates towards error correction can help. This weakens the case for the two-Adversary debate in round one's idea 3 (critic from agents-11): several independent runs of one cheap Adversary, compared by a human, may do as much. It is on benchmarks with answers, so it says little about interpretive Tests, but it argues against complexity for its own sake.
Read first: Abstract; the decomposition of debate into voting and exchange.
- Julian Michael, Samuel R. Bowman and co-authors (NYU); the search record lists Michael, Julian first, then David Bowman and Samuel Bowman (2023). Debate Helps Supervise Unreliable Experts. arXiv:2311.08702 (Nov 2023); NYU archive recordcorrectedopen access
A human-debater proxy for the AI case: a judge who cannot read the source passage does better hearing two sides, one of which argues the wrong answer, than hearing one side whose answer is right half the time. The error analysis is useful here: in consultancy, 52 per cent of errors came from the expert obscuring relevant evidence, which is the one-sided critic's failure. All on reading comprehension with an answer key, so the strongest case for two-sided Tests and also the case where Holagram differs most. Complements round one's agents-11.
Read first: Abstract; the error analysis comparing debate and consultancy.
Audit: Digest's 'David Bowman' is wrong; David Rein, not Bowman; Samuel R. Bowman last.
- Nat McAleese and colleagues (OpenAI) (2024). LLM Critics Help Catch LLM Bugs. arXiv:2407.00215 (June 2024); OpenAI post 'Finding GPT-4's mistakes with GPT-4' (27 June 2024)not re-checkedopen access
A trained critic that must point at specific lines, with humans rating the critiques. Two lessons for the Adversary: a critic can be made to find more than a paid human reviewer, but invented problems are its failure, so each objection has to point to something a person can open; and a human-plus-critic team reduced invented objections. This is code with known bugs, so it is evidence about format and workflow, not about interpretation.
Read first: Abstract; the results on hallucinated bugs and the human-machine team.
- Songhee Han, Jueun Shin, Jiyoon Han, Bung-Woo Jun, Hilal Ayan Karabatman (2026). How Trustworthy Are LLM-as-Judge Ratings for Interpretive Responses? Implications for Qualitative Research Workflows. arXiv:2604.00008 (submitted 9 Mar 2026, cs.CL)checkedopen access
Tests whether a model can grade interpretive responses. Model-level trends matched human ratings but scores differed in size, and faithfulness and correctness checks misaligned with human judgement at excerpt level, especially for non-literal readings. Use: do not let a model rate the Adversary's tests, or another model's Fictions, as a shortcut; the outside human reader rates them.
Read first: Abstract.
- Tanisha (Tina) Jowsey, Virginia Braun, Victoria Clarke, Deborah Lupton, Michelle Fine (open letter, signed by several hundred researchers) (2025). We reject the use of generative artificial intelligence for reflexive qualitative research. Qualitative Inquiry (accepted 21 Nov 2025, online 17 Dec 2025); SSRN preprint 5676462correctedmust read
The objector text. It says reflexive analysis (reflexive thematic analysis is the named example) is a human practice that needs a positioned researcher, so generative AI in it is methodologically incongruent, and it adds a separate social and environmental justice objection. Two consequences for Holagram. (1) The first objection is about analysis and interpretation, which is the Reading register; Holagram already keeps Reading and Trace human, so the core of the method is outside the letter's target as far as the abstract shows. (2) The justice objection is categorical and is not answered by any protocol, only acknowledged and limited. The method text should say both things plainly.
Read first: The abstract only (short); then the full letter if you can open it.
Audit: First author is Tracey Jowsey, not Tanisha (Tina). Journal article in Qualitative Inquiry; 'several hundred' = 419 from 32 countries.
- Tanisha Jowsey, Virginia Braun, Victoria Clarke, Deborah Lupton, Michelle Fine (open letter with several hundred signatories) (2025). We reject the use of generative artificial intelligence for reflexive qualitative research. Qualitative Inquiry (DOI is quoted inconsistently across listings, so none given)corrected
An open letter rejecting generative AI for reflexive qualitative research: reflexive analysis needs a positioned human researcher, and the authors add social and environmental justice objections. Holagram keeps Trace and Reading human, which answers the first objection for the interpretive core. A rebuttal exists (seen on SSRN, not read). Reply in the method text: say what the synthetic hands do not do, and count the runs.
Audit: Add DOI 10.1177/10778004251401851; published online 17 Dec 2025.
- Susanne Friese, Kien Nguyen-Trung, Powell, Morgan (full author names not seen) (2026). Beyond Binary Positions: Making Space for Critical and Reflexive GenAI Integration in Qualitative Research. Qualitative Inquiry (reply to Jowsey et al.; March 2026 per search summary); SSRN 5962174checkedmust read
The most direct reply to objections-01. It accepts the ethical and socio-technical worries behind the letter but denies that generative AI is inherently incompatible with meaning-based methods, and warns that a pro-AI versus anti-AI binary pushes out in-between positions. For Holagram it shows what a middle position needs to look like to be heard: concede the ethical costs, restrict the claim, and say exactly where the human stays in charge. It is also the better place to look for what a critic of the letter still concedes.
Read first: The abstract.
- Kien Nguyen-Trung, Susanne Friese (2026). Towards a methodologically congruent framework for GenAI use in nonpositivist qualitative research. International Journal of Social Research Methodology, 1-22; SSRN 5874482correctedmust read
The clearest statement of the interlocutor position: move from AI as coder to AI as a non-human dialogue partner that retrieves quotations, summaries and contrasting perspectives, while the researcher interprets, explains why patterns matter and ties them back to data; check continuously that outputs are grounded, look for negative cases, and document the whole process. This is nearly Holagram's Adversary written for qualitative analysis. The reviewer's caveat matters: plausible summaries make checking easy to skip under time pressure, so use it to locate passages, not to summarise them.
Read first: The blog summary, then the paper's section on the dialogue process.
Audit: Add venue: International Journal of Social Research Methodology. Cited URL is a blog, not the paper.
- Jessica Nina Lester, Trena M. Paulus (2025). The AI Dilemma in Qualitative Research: Progress, Pitfalls, and the Price of Innovation. CAQDAS Networking Project blog, University of Surrey, 25 Mar 2025not re-checkedopen access
A middle position from the people who wrote much of the methods literature on digital tools: be reflexive about the consequences of an AI-assisted workflow under four headings (methods, technologies, humans, knowledge) and opt out where possible. It supplies a ready structure for the declaration paragraph in the Holagram method text, and it names the thing the letter's critics say they lose: the learning that happens while struggling with the data.
Read first: Whole post (short).
- Christina Silver (2026). We don't need to lose our (qualitative) minds in the GenAI era, but we might. CAQDAS Networking Project blog, University of Surrey, 26 Mar 2026 (from her MAXDAYS 2026 keynote)not re-checkedopen access
Argues that generative AI cannot do critical thinking, interpretation or reflexivity as people do, so these stay human tasks, and that whether it erodes them depends on use. She asks for case-by-case decisions that start from the analytic goal and not from the tool. For Holagram: the decision whether to use a synthetic hand should be made and written down per register and per entry, not once for the study.
Read first: Whole post.
- Matthew Nyaaba, Min SungEun, Mary Abiswin Apam, Kwame Owoahene Acheampong, Emmanuel Dwamena, Xiaoming Zhai (2026). Human-AI Collaborative Inductive Thematic Analysis: AI Guided Analysis and Human Interpretive Authority. arXiv:2601.11850 (submitted 17 Jan 2026, cs.AI)not re-checkedopen access
Three experienced researchers worked with a tool that enforced traceability to the source text and kept logs of every edit, deletion, rejection and comment. The authors report that interpretive authority stayed with the humans, shown by those logged actions. It is a worked example of the overruled field: the log of what the human did with the machine's output is the evidence of who held the interpretation. Very small study; it describes the tool, it does not test whether authority really stayed human.
Read first: Abstract.
- James Eschrich, Sarah Sterman (2024). A Framework For Discussing LLMs as Tools for Qualitative Analysis. arXiv:2407.11198; CHI 2024 workshop LLMs as Research Toolsnot re-checkedopen access
That is the Adversary's role stated in the vocabulary of qualitative research, and a way to answer the open letter in round one (agents-16) for the Test register. Workshop paper and conceptual, with no test.
Read first: Abstract; the proposing versus refuting axis.
- Thomas H. Costello, Gordon Pennycook, David G. Rand (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science 385, eadq1814 (13 September 2024)checkedmust read
The strongest single result for a model arguing against a person's own strongly held reading with specific evidence. It bears on the Adversary, but as of June 2026 its reported values are under an Editorial Expression of Concern, so no number from it should be cited in the Holagram method text without that caveat. The design (a consenting believer, a stated belief, a topic with a known factual record) differs from a researcher's reading of a plan document or a street, where no answer key exists.
Read first: Abstract only, as summarised in coverage: 2,190 believers, GPT-4 Turbo arguing against a self-described belief, about 20% fall in belief, said to persist for two months. Read it only together with science-02 to science-04.
- H. Holden Thorp (Science) (2026). Editorial Expression of Concern on 2024 Research Article 'Durably reducing conspiracy beliefs through dialogues with AI'. Science 392(6803):1131, 11 June 2026; DOI 10.1126/science.aej2383correctedmust readopen access
Settles the status question for the protocol: the paper is flagged, not retracted, and the problems are in data handling and reporting, not (so far as the notice says) in the direction of the effect. The expression of concern is itself a worked example of Holagram's own Trace rule: a value is only as good as the dataset someone can open and re-run. Use it as the model for a hands-and-data record that survives a reader trying to reproduce each number.
Read first: The notice itself (short). Summary of what it says: after publication the authors were made aware of reporting inconsistencies and public-dataset problems that made some reported values hard to reproduce; the authors found that screening criteria were applied inconsistently between the manuscript and the published analysis pipeline, and that the public dataset held extraneous spliced rows from a code-merging error; they gave Science raw data and updated results, which Science was still evaluating; the authors' reanalysis is reported to match the original in direction, statistical significance and substantive size.
Audit: Title is 'Editorial Expression of Concern on RA ...' (EurekAlert); cite Crossref DOI 10.1126/science.aej2383 and Science 392:1131 as the notice itself. Digest URL is the press release, not the notice.
- Thomas H. Costello, David G. Rand, Gordon Pennycook (2026). Durably reducing conspiracy beliefs through dialogues with AI: Correction and reproducibility materials. Zenodo, DOI 10.5281/zenodo.22754984, version 1.0, 27 September 2026correctedopen access
The authors' answer to the expression of concern: headline analyses re-specified as ordinary least squares with pretreatment belief centred on the screened cohort, persistence analyses keep the original follow-up restrictions, protest-intention results use only Study 2b. It shows what a reproducible hands-and-data package looks like for a dialogue study, and it is a template for Holagram's per-entry record (prompt, model version, outputs, checksums). Science's formal outcome (correction, clearance or retraction) was not found as of 14 October 2026.
Audit: Digest has no date issue except it is dated 2026 which is right (27 Sep 2026). Authors order here Costello, Pennycook, Rand (digest had Rand second-last). Not a separate paper from the Science article: a supplement.
- Esther Boissin, Thomas H. Costello, Daniel Spinoza-Martin, David G. Rand, Gordon Pennycook (2025). Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human. PNAS Nexus 4(11), pgaf325 (online 14 October 2025)checkedmust readopen access
The cleanest evidence on facts versus rapport: tone and perceived humanness did not matter, and the content of the argument did, though every arm still got fact-based argument, so rapport alone was not tested. For Holagram it means the Adversary can be written in a dry, non-warm register without losing force, and that warmth is not what would make a researcher listen. It also shows the effect is not specific to believers: a poorly supported belief moved too, which is nearer to a researcher's overconfident reading.
Read first: Abstract and the discussion of limitations (single model, English and Western training data).
- Thomas H. Costello, Kellin Pelrine, Matthew Kowal, Jasper Timm, Antonio A. Arechar, Jean-Francois Godbout, Adam Gleave, David Rand, Gordon Pennycook (2026). Large language models can effectively convince people to believe conspiracies. arXiv:2601.05050 (v1 8 January 2026; v3 15 July 2026); not seen as publishedcheckedmust readopen access
The same research group shows the mechanism is symmetric: a model can move belief toward a falsehood about as well as toward the truth, and people prefer it when it does. For Holagram this is the core warning for Alter and Adversary alike. A model argues whichever side it is given, so an Adversary against the Reading is also a machine that could argue for it, and Alter writing the world where the Reading is true is, structurally, the promoting condition. The protocol must therefore treat persuasiveness as no evidence of correctness, and constrain both roles to checkable claims.
Read first: Abstract, then the guardrail condition.
- Kobi Hackenburg, Ben M. Tappin, Luke Hewitt, Ed Saunders, Sid Black, Hause Lin, Catherine Fist, Helen Margetts, David G. Rand, Christopher Summerfield (2025). The levers of political persuasion with conversational artificial intelligence. Science 390(6777); DOI 10.1126/science.aea3884; arXiv:2507.13919checkedmust readopen access
Answers 'facts versus rapport' at scale: information density, not personalisation or psychological technique, drives persuasion, and the more persuasive the output, the less accurate its claims. For the Adversary this is a direct risk: a Test written to be forceful will be dense with factual-looking claims that were optimised for effect, not checked. Rule: an Adversary's claims are leads, not Traces, and the more fluent and detailed the objection, the more it needs a human to open the source.
Read first: Abstract; the accuracy trade-off.
- Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, Dan Jurafsky (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science (2026); preprint arXiv:2510.01395 (1 October 2025)correctedmust readopen access
Moves sycophancy from model benchmarks (agents-03) to a measured effect on the person: one exchange with an agreeable model raised conviction of being right. That is the Ballardian failure with a mechanism: a model that agrees with the Reading makes the researcher surer, and the researcher prefers that model. It is the strongest argument that the Adversary must be tested against agreement, and that the human must not be the sole judge of whether the Adversary was hard enough.
Read first: Abstract.
Audit: Year is 2025, not 2026. Duplicate of critic-02.
- Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley, Joshua B. Tenenbaum (2026). Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians. arXiv:2602.19141 (22 February 2026); not seen as publishedcheckedmust readopen access
A model, not an experiment, but directly on point for Holagram's care rules: if neither factual accuracy nor a warning breaks the spiral, then disclosure and a Fiction label do not protect the researcher's head, and the structural controls (an outside reader who can call a stop, fixed fieldwork blocks, reading-mode off) have to do the work. Simulation results depend on the model's assumptions; read as an argument, not as evidence of size.
Read first: Abstract.