{"path":"research/what-if-and-lenses/outside-reading-psychology.md","content":"# Outside reading in psychology, deliberation and the philosophy of disagreement\n\nDate: 2026-10-22\n\nPurpose: outside reading for the what-if view and worldview lenses. Read depth is stated per source.\n\nHow to read the \"Read\" labels:\n\n- **long excerpts**: I read several pages of the paper body, not every page.\n- **abstract**: I read the abstract only.\n- **secondary**: I read another source's description of the work.\n- **memory**: from memory, not checked this session.\n\nNo source below was read start to finish in a PDF. Where a number comes from a preprint and the published version may differ, the text says so.\n\nOne caution up front. Almost every study here measures attitude change, animosity or persuasion. The project aims at legibility and at people feeling seen. Almost nothing below measures those. So the findings bear on the project by analogy, not directly.\n\n---\n\n## Summary: what matters most\n\n1. **Understanding another mind needs new information from that mind.** In 25 experiments, being told to imagine another person's view did not make people more accurate about that person. Asking them did. Lenses built from holders' own words fit this. A view built on premises the system guessed does not.\n2. **A what-if view is a \"how would this work\" task.** That kind of task reliably lowers people's confidence that they understand a policy. It does not reliably moderate their views. Close replications failed. A later study found moderation only for issues argued by consequences, not for issues held as protected values. Justify the view by legibility, not by moderation.\n3. **Arguments alone changed no minds in the deep canvassing trials. Exchanging stories did, durably but by little (d about 0.04 to 0.08).** A pure premises-and-consequences view is argument mode. Pair each premise with the holder's own words or experience. Personal experiences of harm earned more respect from opponents than facts did.\n4. **Showing the premises on which people differ (cruxes) is doable, but it did not produce convergence.** In structured AI-risk debates, both sides could restate the other's arguments well and still did not move. The best near-term crux was expected to close about 5 percent of the gap. The two sides' most informative cruxes were mostly different ones. This supports \"legibility, not convergence\". Do not promise agreement.\n5. **Moral reframing is shakier than its reputation.** A 2026 study by two independent teams, five issues, found null results (Bayes factors about 100 to 1 for no effect). The moral foundations questionnaire has long-standing structure problems. Do not use foundations as lens axes without a check. Holders' own words avoid the problem.\n6. **Language models voicing people is a known failure area.** Fidelity is shown at group level, not for a person. Models flatten and misportray groups, and they are unstable across prompts and months. The best result used a two-hour interview with the person, and it still reached only 85 percent of the person's own retest on coarse survey items. Keep the rule \"no model plays a person\".\n7. **Tests for \"did we represent the holder fairly\" exist, but they disagree.** In one study pass rates were 54 to 71 percent. In another, readers could not tell real from imitation. In a third, imitations were less extreme than the real thing. Self-judged accuracy is not reliable. The holder is the only oracle, and accountability to the holder improved effort.\n8. **Deep disagreement is about what counts as relevant or as evidence.** Deliberus already makes relevance claims attackable. A what-if view should be parametrised by which relevance claims a holder accepts, not only by which premises. \"Same conclusion, different reasons\" (overlapping consensus) is a legitimate thing to display.\n9. **Feeling understood and being understood are only modestly linked.** Perspective-taking also raised confidence faster than accuracy. A \"mind-meld\" framing risks false confidence. Show what the holder stated and what the system inferred, and say so.\n10. **Backfire is not a general law, but exposure design matters.** One 52-issue test found no backfire on factual correction. One large field study found a backfire when people were paid to follow accounts of opposing elites. Test the exposure design; do not assume it is safe.\n\n---\n\n## 1. Perspective-taking versus perspective-getting\n\n### 1a. Eyal, Steffel and Epley 2018, \"Perspective mistaking\", JPSP\n\n- **Found.** 25 experiments. People told to take another person's perspective were not more accurate at predicting that person's emotions, lies, spouse's preferences or attitudes. If anything they were less accurate (pooled over Experiments 1 to 15: d about 0.26 lower, 95% CI 0.12 to 0.40). They sometimes grew more confident. The effect was not robust in single experiments (significant in 4 of 17). Perspective-taking did reduce egocentric bias, but what replaced it was not more accurate. The final experiment showed that getting the perspective, by conversation, did raise accuracy. Pretest participants believed perspective-taking would help on every measure.\n- **Read.** Abstract and long excerpts of an author manuscript hosted on a third-party site.\n- **Source.** https://doi.org/10.1037/pspa0000115 (abstract also at https://pubmed.ncbi.nlm.nih.gov/29620401/)\n- **For the project.** This is the strongest outside support for lenses made of holders' own words. The accuracy comes from information that left the other person's head, not from the viewer's simulation. A what-if view must run on premises and weights the holder gave. If the system fills gaps by inference, it is doing perspective-taking on the holder's behalf, which is the move that did not help. The authors say their tests may have missed the conditions where perspective-taking works. All tasks were predictions of individual states, not reasoning about policy.\n\n### 1b. Damen, Pollmann and Grassow 2021, \"The benefits and obstacles to perspective getting\", Frontiers in Communication\n\n- **Found.** Three experiments. People often did not ask, even when asking was possible. Perspective-getting improved prediction in two of three studies. In the attitude study, friends who could ask were more accurate than friends who could not (d about 0.90), and instructed asking beat optional asking (d about 0.68). In the first study it did not help. Confidence was a barrier to asking. Telling people that asking works did not increase asking. It worked over chat and video, not only face to face.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://www.frontiersin.org/journals/communication/articles/10.3389/fcomm.2021.611187/full\n- **For the project.** People will not seek the holder's own account by default, and the more sure they feel, the less they ask. The view should put the holder's own statement in front of the reader without requiring the reader to go and ask. It should not rely on telling readers that the statement is more reliable.\n\n### 1c. Kalla and Broockman 2023, \"Which narrative strategies durably reduce prejudice?\", AJPS\n\n- **Found.** Field studies and a survey experiment that separated three narrative strategies. Conversations using only perspective-getting (hearing the other person's own story) still durably reduced exclusionary attitudes. Prompts to take the other's perspective by analogy, and perspective-giving, were not needed.\n- **Read.** Abstract only.\n- **Source.** https://onlinelibrary.wiley.com/doi/10.1111/ajps.12657 (online 2021; issue 2023)\n- **For the project.** Same direction as 1a, now with durable attitude effects. \"Hear it from them\" is the active ingredient the authors isolate. The effect is on attitudes toward outgroups, not on accuracy about their reasoning.\n\n### 1d. Tuller, Bryan, Heyman and Christenfeld 2015, \"Seeing the other side\", JESP\n\n- **Found.** Writing from the standpoint of a specific opponent moderated attitudes (weight-discrimination issue, d = 0.47 against writing from one's own standpoint), and also on abortion. It worked only when participants had met the opponent and were told that opponent would judge the accuracy of what they wrote. With an abstract \"other\", or with only expert review, there was no extra effect (d = 0.02 in the abstract-other condition).\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1016/j.jesp.2015.02.003\n- **For the project.** Perspective-taking can work when the target can check it. A lens that the holder can read, correct and sign off on is the same structure. A lens of an unnamed \"typical conservative\" is the condition that showed nothing. Samples were small (roughly 85 to 105 per study) and I found no replication.\n\n### 1e. Bruneau and Saxe 2012, \"The power of being heard\", JESP\n\n- **Found.** Two studies, Mexican immigrants and White Americans in Arizona, then Israelis and Palestinians. One person wrote about the difficulties of life in their group (\"perspective-giving\"). The other restated it in their own words (\"perspective-taking\"). Attitudes toward the outgroup improved more after perspective-taking for the dominant groups, and more after perspective-giving for the non-dominant groups. For Palestinians, writing an essay without a partner did nothing, so the effect depended on being heard.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1016/j.jesp.2012.02.017\n- **For the project.** This is direct evidence for the attunement pole: being heard has its own effect, and it is larger for the people who feel unheard. The benefit needs a real listener, not a record. A lens that stores a person's view but is never visibly received may not give this benefit.\n\n### 1f. Schroeder, Kardas and Epley 2017, \"The humanizing voice\", Psychological Science\n\n- **Found.** Four experiments on polarizing issues. When someone disagreed with the reader, hearing them speak made them seem more humanlike than reading the same words (d = 0.46 on one humanlikeness measure and 0.26 on another, in the excerpt I read). With agreement the pattern was inconsistent.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1177/0956797617713798\n- **For the project.** Text hides some of the cues that make an opponent seem like a thinking person. A text-only lens made of holders' words may under-humanize. Short audio of the holder stating a premise is worth testing. This is my inference, not a finding.\n\n### 1g. Reis, Lemay and Finkenauer 2017, \"Toward understanding understanding\", Social and Personality Psychology Compass\n\n- **Found.** Feeling understood and being understood are only modestly related. Many things besides accuracy produce the feeling.\n- **Read.** Abstract and introduction.\n- **Source.** https://doi.org/10.1111/spc3.12308\n- **For the project.** If attunement is measured by \"I felt understood\", it can rise without the lens being accurate, and the reverse. Measure both and keep them apart.\n\n### 1h. Itzchakov and colleagues, listening studies\n\n- **Found.** In five experiments (2018), high-quality listening (attentive, empathic, nonjudgmental) made speakers' own attitudes clearer to them, not more correct, and made them more willing to express them. In four experiments (2017), it raised awareness of inconsistencies in one's own view and lowered extremity.\n- **Read.** Abstracts only.\n- **Source.** https://doi.org/10.1177/0146167217747874 and https://pubmed.ncbi.nlm.nih.gov/27856728/\n- **For the project.** Being listened to changes the speaker. A system that listens well may soften or sharpen the holder's own view as a side effect. That is not convergence, but it is movement, and it should be tracked rather than assumed to be zero.\n\n---\n\n## 2. Moral reframing, moral foundations and their critiques\n\n### 2a. Feinberg and Willer 2015, \"From gulf to bridge\", PSPB\n\n- **Found.** Six studies (same-sex marriage, health care, military spending, English as official language). People write persuasive messages in their own moral language, not the audience's. Messages reframed into the audience's values persuaded more. Mediation: the reframed message made the position seem to fit the audience's values. In a side test, most people could pick the reframed message as the more persuasive one (64 percent of liberals, 85 percent of conservatives) but few wrote one unprompted. A small minority said they would not send it.\n- **Read.** Abstract and long excerpts of the discussion.\n- **Source.** https://journals.sagepub.com/doi/abs/10.1177/0146167215607842 (full text mirror: https://www.scicom-bellagio.com/wp-content/uploads/2017/10/Feinberg-and-Willer-2017_From-Gulf-to-Bridge-When-Do-Moral-Arguments-Facilitate-Political-Influence.pdf)\n- **For the project.** Reframing is persuasion by translation. It aims to move the audience, which is not the project's stance. But the finding that people do not spontaneously write in the other's moral language, and mostly can see why they should, is a use case for a lens: it supplies the other's own phrasing.\n\n### 2b. Feinberg and Willer 2019 review; Voelkel and Feinberg 2018\n\n- **Found.** The 2019 review says moral reframing works across economic inequality, environment, same-sex marriage and candidates, and lists open questions on mediators and moderators. The 2018 paper (two studies, 2016 U.S. election) found conservatives reading a loyalty-based criticism of one candidate supported him less than those reading a fairness-based one, and the reverse for liberals.\n- **Read.** Abstracts only.\n- **Source.** https://compass.onlinelibrary.wiley.com/doi/10.1111/spc3.12501 ; https://pmc.ncbi.nlm.nih.gov/articles/PMC6295651/\n- **For the project.** This is the pro-reframing case. Read it with 2c.\n\n### 2c. Recent null results: Hawkins and colleagues 2026; Voit, Twardawski and Fischer 2026\n\n- **Found.** Two independent teams tested moral reframing on five issues (environment, tax, immigration, English as official language, universal health care). They found null results throughout, including in the sub-groups most likely to respond. Bayes factors put the data about 100 times more likely under no effect. They found no good explanation. Separately, a direct replication in the U.S. (N = 743) and a conceptual replication in Germany (N = 269) of a successful conservative-values climate message did not bridge the liberal-conservative gap.\n- **Read.** Abstracts only. Both are very recent; I did not see commentary on them.\n- **Source.** https://doi.org/10.1080/10584609.2026.2612739 (Political Communication, online 2026-01-07) ; https://www.sciencedirect.com/science/article/pii/S0022103126000016 (JESP, 2026)\n- **For the project.** Do not build a lens feature whose point is \"restate it in their moral language and they will be moved\". The evidence is now mixed enough that this would be an untested bet. If the lens only shows the holder's own words, it does not depend on this effect.\n\n### 2d. Critiques of moral foundations theory\n\n- **Found.** Suhler and Churchland (2011) argue the five foundations look contrived and the innateness and modularity claims do not fit neuroscience. Zakharin and Bates (2021, five studies, total N = 11,496) report that no tested five-factor model of the questionnaire fits adequately. They build a seven-foundation model and a \"moral tilt\" factor. Their introduction collects other critiques: two factors instead of five, poor reliability for some scales, doubtful invariance across countries. Curry (2019, an essay) says the theory lacks a principled basis and omits several well-established forms of cooperation. Gray and colleagues (2022) review evidence against treating purity as one coherent moral domain.\n- **Read.** Abstracts, and the introduction of Zakharin and Bates. Curry is an opinion essay.\n- **Source.** https://doi.org/10.1162/jocn.2011.21637 ; https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0258910 ; https://behavioralscientist.org/whats-wrong-with-moral-foundations-theory-and-how-to-get-moral-psychology-right/ ; https://journals.sagepub.com/doi/10.1177/10888683221124741\n- **For the project.** If a lens dimension were \"which foundations you weight\", it would stand on a measure whose structure is disputed. Using what holders say they find in need of explaining, in their own words, avoids borrowing that measure. I did not research the Bear and Knobe dimension itself; that is a different literature.\n\n---\n\n## 3. Deep canvassing, narratives and the depolarization megastudy\n\n### 3a. Kalla and Broockman 2020, APSR\n\n- **Found.** Three preregistered field experiments, 230 canvassers, 6,869 voters, seven U.S. locations. Face-to-face conversations that used arguments alone had no effect on immigration or prejudice. Otherwise identical conversations that also included a non-judgmental exchange of narratives reduced exclusionary attitudes for at least four months (d = 0.08). The two transphobia experiments gave d = 0.08 and 0.04, including by phone. Canvassers asked people to reach their own conclusions from the stories. Analogic perspective-taking was not necessary. Whether the canvasser belonged to the outgroup did not moderate. The paper describes an earlier 2016 study of the same approach (N = 501) that found a durable reduction in transphobia. A 2014 study of the technique (LaCour and Green) was retracted after data irregularities; I know that from a search summary and from memory, not from this paper.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1017/s0003055419000923\n- **For the project.** Effects are small and need real conversation. They show that a conversation of arguments alone did nothing, and that the active part was non-judgmental listening plus stories. A what-if view is made of arguments. Expect it to be useful for clarity and perhaps respect, but not to move attitudes unless stories ride with it.\n\n### 3b. Kalla and Broockman 2022, canvassing changed the canvassers\n\n- **Found.** In three real campaigns, activists who held in-depth two-way conversations with voters who disagreed became less hostile to out-partisans themselves: about 4.5 points on a feeling thermometer toward Trump voters, and affective polarization down 4 to 6 points. Measured a median of about 15 days after the last conversation (Studies 1 and 2). Liberal activists only. Within-subject design, no control group.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1017/s0003055422000132 (APSR, online 2022)\n- **For the project.** The listener is changed by hearing a real person's story. This is a case for letting the reader of a lens meet a holder's own account, not only a structure derived from it. The effect is not shown to last, and the sample is one-sided.\n\n### 3c. Kubin, Puryear, Schein and Gray 2021, PNAS\n\n- **Found.** 15 studies. Political opponents respected moral beliefs more when they were supported by personal experiences than by facts. The most effective experiences were relevant to the issue and involved harm. The mechanism was seeing the other as more rational. This held in moral disagreements; in agreement and in non-moral disagreement, facts did as well as experiences.\n- **Read.** Abstract and discussion.\n- **Source.** https://www.pnas.org/doi/full/10.1073/pnas.2008389118\n- **For the project.** A what-if view displays \"if you grant these premises, this follows\". That is the facts-and-logic mode. For moral disagreements, the evidence says the same premise attached to a person's lived experience earns more respect. Offer the holder's story as an optional layer under each premise. The result concerns respect, not accuracy.\n\n### 3d. Voelkel and colleagues 2024, the Strengthening Democracy Challenge megastudy, Science\n\n- **Found.** 25 treatments, 32,059 U.S. partisans in an online survey. 23 reduced partisan animosity (by up to 10.5 percentage points). Strongest: showing relatable, sympathetic people with different views, and stressing shared identities. Six reduced support for undemocratic practices (up to 5.8 points), most strongly by correcting misperceptions of the other side's views and by showing the risk of democratic collapse. Five reduced support for partisan violence. The treatments that worked best for animosity differed from those that worked for antidemocratic attitudes. The authors say they did not find large, durable effects from single short online treatments. In the preprint (numbers may differ in the published paper): some treatments backfired on undemocratic attitudes; treatments that gave understandable reasons why out-partisans hold their identity were small (d about -0.09 to -0.10); a \"moral similarities and differences\" treatment was mid-table (d = -0.26); correcting misperceptions worked only for some kinds (correcting beliefs about dehumanization worked, correcting policy positions did not).\n- **Read.** Published abstract and editor's summary, the Science \"discussion\" text as extracted, and long excerpts of the preprint (OSF, 2023).\n- **Source.** https://www.science.org/doi/10.1126/science.adh4764 ; preprint https://doi.org/10.31219/osf.io/y79u5 ; the policy forum by Nyhan and Titiunik (not read) is at https://doi.org/10.1126/science.ads7371\n- **For the project.** Two things bear on lenses. First, \"show what the other side actually believes\" corrected exaggerated beliefs and helped on some outcomes, but only when the correction was about the right thing. Second, explaining why the other side holds its identity was a weak treatment. The results measure survey attitudes right after a short exposure, not understanding.\n\n---\n\n## 4. The illusion of explanatory depth for policies\n\n### 4a. Fernbach, Rogers, Fox and Sloman 2013, Psychological Science\n\n- **Found.** Experiment 1 (N = 198): asking people to explain how a policy works lowered their rating of how well they understood it and made their stated positions more moderate. Experiment 2 (141 recruited, 112 analysed): listing reasons for one's position did not do this; mechanistic explanation did. Experiment 3: mechanistic explanation lowered donations to like-minded advocacy groups.\n- **Read.** Abstract and long excerpts of Experiments 1 and 2.\n- **Source.** https://doi.org/10.1177/0956797612464058\n- **For the project.** This is the paper behind the intuition that a what-if or \"how would it work\" view will humble people. The samples were small, and the replications below did not hold up the moderation part.\n\n### 4b. Crawford and Ruscio 2021, Psychological Science\n\n- **Found.** Two preregistered close replications of Experiment 2 (N = 306 and N = 405) and a close and conceptual replication of Experiment 3 (N = 343). None of the key effects was significant. People again reported understanding the policies less after explaining them. Their positions did not become more moderate, and in-group favouritism did not change. Listing reasons increased extremity, which fits older findings.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1177/0956797620972367\n- **For the project.** The robust part is the drop in felt understanding. That is a legitimate benefit for a legibility tool. The moderation of views is not established. A design that depends on it should be avoided.\n\n### 4c. Sloman and Vives 2022, Cognition\n\n- **Found.** A new experiment (N = 739, Prolific). Understanding and attitude extremity both fell after mechanistic explanation, but only for issues framed around consequences (student-loan cancellation, drug-price limits). They did not fall for issues held as protected values (gun background checks, citizenship rules). Social proof about not understanding made no difference.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1016/j.cognition.2022.105146\n- **For the project.** This is the most useful finding for the lens idea. Whether a what-if view opens anything depends on whether the premise is held as a consequence claim or as a protected value. A lens could mark which premises a holder treats as non-negotiable. Do not expect the what-if view to loosen those. This is the authors' own proposed boundary, tested once.\n\n### 4d. Porter and colleagues 2022 review, Nature Reviews Psychology\n\n- **Found.** The review lists the explanation task among brief ways to raise intellectual humility, and says that a close replication reduced overestimated knowledge but did not change extreme attitudes. Most brief interventions had small samples and no replication. Self-distanced reflection had replicated.\n- **Read.** Long excerpts.\n- **Source.** https://www.nature.com/articles/s44159-022-00081-9 (also https://pmc.ncbi.nlm.nih.gov/articles/PMC9244574/)\n- **For the project.** It confirms the picture in 4b.\n\n---\n\n## 5. Identity-protective cognition, backfire and intellectual humility\n\n### 5a. Kahan, Peters, Dawson and Slovic 2017, \"Motivated numeracy\", Behavioural Public Policy\n\n- **Found.** The identity-protective cognition thesis predicts that people with more numeracy will use it to protect their group's belief. Polarization on a gun-control data problem grew with numeracy rather than shrinking.\n- **Read.** Abstract.\n- **Source.** https://www.cambridge.org/core/journals/behavioural-public-policy/article/abs/motivated-numeracy-and-enlightened-selfgovernment/EC9F2410D5562EF10B7A5E2539063806\n- **For the project.** The standard worry behind \"show people the other side's chain of reasoning and they will use it to defend their own\". Read it with 5b.\n\n### 5b. Persson and colleagues 2021; Tappin, Pennycook and Rand 2021\n\n- **Found.** Persson et al. ran a large preregistered replication of motivated numeracy and found no good evidence for it. Congruent responding was not moderated by numeracy, and where it was moderated it appeared only at the very highest numeracy. Tappin et al. (three studies) found a robust direct effect of political group identity on reasoning, but no evidence that cognitive sophistication magnifies it. Sophistication magnified the effect of prior factual beliefs instead.\n- **Read.** Abstracts, plus long excerpts of Tappin et al.\n- **Source.** https://doi.org/10.1016/j.cognition.2021.104768 ; https://doi.org/10.1037/xge0000974\n- **For the project.** Group identity does bias reasoning, but \"smarter people defend their side harder\" is not well supported. Do not design around it. Do design for the direct identity effect.\n\n### 5c. Kahan and colleagues 2017, \"Science curiosity\", Political Psychology\n\n- **Found.** A new science-curiosity scale predicted willingness to read surprising information even when it ran against one's politics. Among highly curious subjects the polarizing effect of comprehension shrank. The authors flag that this is early work.\n- **Read.** Abstract and conclusion.\n- **Source.** https://doi.org/10.1111/pops.12396\n- **For the project.** Curiosity, not skill, predicted openness. A view that is framed as \"see what you did not expect\" may reach curious users more than a view framed as \"correct yourself\". I did not check whether this has replicated.\n\n### 5d. Bail and colleagues 2018, PNAS; Levy 2021, American Economic Review\n\n- **Found.** Bail et al. surveyed about 1,650 Twitter users, then offered $11 to those randomly assigned to treatment to follow a bot that retweeted opposing political accounts for a month. Republicans moved in a more conservative direction (0.12 on a 7-point scale in the cautious estimate; 0.11 to 0.59 SD depending on estimate). Democrats moved slightly liberal, not significantly. The authors state limits: paid compliance, Twitter only, and no tested mechanism. Levy randomly changed what news outlets people saw on Facebook for two months. The opinion index did not move (a 0.005 SD point estimate). He argues the difference from Bail may be because Bail's design showed opposing opinion leaders rather than news outlets, and forced attention.\n- **Read.** Abstract and long excerpts of Bail et al. A passage of Levy only; I did not read his results on affective polarization.\n- **Source.** https://www.pnas.org/doi/full/10.1073/pnas.1804840115 ; http://www.roeelevy.com/Papers/Social_Media_and_Polarization.pdf\n- **For the project.** Pure exposure to opposing views is not safe by default. What is exposed (opposing elites versus structured reasons) and under what attention conditions matters. The what-if view is a structured, holder-chosen exposure, not a feed. That difference is untested.\n\n### 5e. Wood and Porter 2019, \"The elusive backfire effect\", Political Behavior\n\n- **Found.** Five experiments with more than 10,100 participants and 52 issues. No correction produced a backfire, even on polarized issues. People generally heeded factual information.\n- **Read.** Abstract and press summaries, secondary.\n- **Source.** https://www.doi.org/10.1007/s11109-018-9443-y ; https://papers.ssrn.com/abstract=2819073\n- **For the project.** Factual correction rarely backfires. This does not cover other kinds of exposure (5d), nor changes of attitude as opposed to fact.\n\n### 5f. Intellectual humility\n\n- **Found.** Porter and Schumann (2018): people higher in intellectual humility were more open to the opposing view in imagined disagreements and read more opposing perspectives. A reading that made a growth mindset salient raised intellectual humility and openness (small sample, one study). Bowes et al. (2020): intellectual humility correlates with less affective polarization; correlational. Koetke and Schumann (nine samples, N = 3,248): humility is not linked to belief strength but is linked to a more liberal orientation, a confound for the field. A set of four preregistered German experiments (N = 1,668; authors not captured in my read): a target person's humility, and the participant's own, made participants more willing to approach rather than avoid a person who disagreed. The 2022 review (4d) says short-term boosts are possible but largely unreplicated beyond self-distancing.\n- **Read.** Abstracts, and long excerpts of the review.\n- **Source.** https://doi.org/10.1080/15298868.2017.1361861 ; https://www.sciencedirect.com/science/article/abs/pii/S0092656620300805 ; https://www.tandfonline.com/doi/full/10.1080/17439760.2024.2352739 ; https://www.apa.org/pubs/journals/releases/psp-pspi0000462.pdf\n- **For the project.** The German result is the interesting one: perceived humility in the other person made people willing to engage with them. A lens that shows a holder stating uncertainty in their own words could help engagement. Most of this literature is correlational, so do not claim a causal effect.\n\n---\n\n## 6. Philosophy of bridging\n\n### 6a. Rawls, overlapping consensus (Political Liberalism, 1993)\n\n- **Found.** In an overlapping consensus, citizens with different comprehensive doctrines each endorse a shared political conception for reasons internal to their own doctrine. It is stronger than a modus vivendi, which holds only while power is balanced. Rawls does not claim it exists everywhere or lasts forever.\n- **Read.** Stanford Encyclopedia of Philosophy entry, section 3.5, and a scholarly summary. Primary text not read.\n- **Source.** https://plato.stanford.edu/entries/rawls/\n- **For the project.** The what-if view can show this structure exactly: two holders with different premises, same conclusion. That is a real result about their relation and needs no convergence. It is also a limited one: Rawls is clear that it applies to a political conception among reasonable doctrines, not to all questions.\n\n### 6b. Sunstein 1995, \"Incompletely theorized agreements\", Harvard Law Review\n\n- **Found.** Legal systems often secure agreement on an outcome and on low-level reasons, without agreement on fundamental principle. When people disagree on an abstraction they move to more particulars. This supports stability and mutual respect. Sunstein separates it from Rawls: Rawls seeks agreement on political abstractions among comprehensive views; Sunstein seeks agreement on particular outcomes among people who disagree on principle.\n- **Read.** Abstract, and a long excerpt of the commentary.\n- **Source.** https://doi.org/10.2307/1341816\n- **For the project.** This justifies a view that goes down a level when abstract premises clash: \"you disagree on the principle; do you agree on this case?\" A what-if view over specific policies does that. The cost is that agreement on particulars can hide that the reasons differ. Show both.\n\n### 6c. Deep disagreement: Fogelin 1985 and later work\n\n- **Found.** Fogelin argued that some disagreements cannot be resolved by argument, because the parties differ on framework propositions that argument presupposes. The Stanford Encyclopedia says deep disagreements concern what considerations are relevant and what counts as evidence for what. Opinion splits on whether they are rationally resolvable. The Stanford Encyclopedia lists Fogelin, Ranalli, Pritchard and Lynch among those who hold they are not, and Feldman and Matheson among those who hold they are. Ranalli's paper argues the answer depends on what hinge commitments are: non-epistemic theories support pessimism, an entitlement theory permits optimism. Pritchard's own abstract argues that a hinge epistemology does not lead to relativism and shows how to respond rationally.\n- **Read.** Stanford Encyclopedia section 5.5.4, abstracts of Ranalli and Pritchard, and a secondary quotation of Fogelin. Fogelin's paper itself not read; quotations are from secondary sources.\n- **Source.** https://plato.stanford.edu/entries/disagreement/ ; https://link.springer.com/article/10.1007/s11229-018-01956-2 ; https://doi.org/10.1007/s11245-018-9612-y\n- **For the project.** In Deliberus, every link carries a claim that it is relevant, and that claim can be attacked. That is a built-in handle on exactly what the literature calls deep: disputes about relevance and evidence. A what-if view that lets a holder switch relevance claims on and off would show where the disagreement is deep, not just where premises differ. Where nothing further down exists to appeal to, the view should say so rather than imply resolution.\n\n### 6d. Mercier and Sperber 2011; Mercier and Landemore 2012\n\n- **Found.** Mercier and Sperber argue that reasoning evolved to produce and evaluate arguments for persuasion. Mercier and Landemore apply it to deliberation: reasoning works best when used to produce and evaluate arguments in public deliberation, group reasoning outperforms individual reasoning when diverse opinions are discussed, and confirmation bias drives polarization when people reason alone or with like-minded peers. They add that failures of reasoning are better repaired collectively than individually.\n- **Read.** Mercier and Landemore: abstract and an introduction excerpt. Mercier and Sperber: abstract and a summary only.\n- **Source.** https://doi.org/10.1111/j.1467-9221.2012.00873.x ; https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/why-do-humans-reason-arguments-for-an-argumentative-theory/53E3F3180014E80E8BE9FB7A2DD44049\n- **For the project.** This supports putting different views into one graph rather than keeping them apart. It also warns that a tool for one person alone, reasoning on their own premises, will tend to entrench. The what-if view should be built for two or more holders in one space.\n\n### 6e. Landemore, Democratic Reason (2013), and critics\n\n- **Found.** Landemore argues that diversity of cognitive approaches, not individual ability, gives inclusive deliberation its edge, relying on the Hong and Page \"diversity trumps ability\" theorem. Critics: Ancell argues that Hong and Page's results undercut the conclusion that democracy beats non-democratic alternatives; Benson argues the theorem underestimates the complexity of political problems and offers a new model that keeps the value of diversity but not the strong claim.\n- **Read.** Publisher blurb for Landemore; abstracts and a long excerpt of Benson; abstract of Ancell.\n- **Source.** https://press.princeton.edu/books/paperback/9780691176390/democratic-reason ; https://link.springer.com/article/10.1007/s11229-021-03161-0 ; https://www.cambridge.org/core/journals/episteme/article/abs/democracy-isnt-that-smart-but-we-can-make-it-smarter-on-landemores-democratic-reason/025A01004C498E2CF398B5AD969A2F18\n- **For the project.** The weaker claim, that varied approaches help, is safe. The strong claim is contested. The project does not need it.\n\n### 6f. Fishkin and colleagues 2021, \"America in One Room\", APSR\n\n- **Found.** A deliberative poll with 526 registered voters over a long weekend, and a control group of 844. Briefing materials were vetted by experts on opposing sides. Participants discussed in moderated small groups. The deliberators showed large depolarizing changes in policy attitudes and in affective polarization. Democrats' thermometer ratings of Republicans rose 13 points, Republicans' of Democrats 14, with the control group unchanged. The most extreme partisans gained the most. The design makes no attempt to reach a group consensus and measures individual opinions in confidential questionnaires.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1017/s0003055421000642\n- **For the project.** This is the closest design match: it measures individual change without forcing agreement. The effects are large, but the intervention was a long weekend with balanced briefing, small moderated groups and expert panels, for 526 people. It is not a screen. The authors themselves raise scaling as an open question.\n\n---\n\n## 7. Practices\n\n### 7a. The ideological Turing test\n\n- **Found.** Caplan proposed it in 2011 (known to me from Yeomans's account and from memory; not re-read). Four empirical studies:\n  - Yeomans 2022 (U.S. health-care debate, incentivised): partisans showed a limited ability to simulate opponents. Judges were 55.3 percent accurate against 50 percent chance, with confidence of 77.3 percent. Imitations were less extreme than the real thing. Judges' accuracy tracked effort (word count). Political sophistication improved the representation of one's own side only.\n  - Mastroianni, Dana and Nichols 2024 (preprint; 902 writers, roughly 750 to 800 readers): fake statements (\"I'm a Democrat because...\") were indistinguishable from real ones. Readers performed at chance, and said \"real\" too often.\n  - Brand, Brady and Stafford 2025 (Cognitive Science; 1,200 participants; Brexit, vaccination, veganism, UK): \"pass\" rates of 54 to 71 percent depending on the topic. People who passed rated opponents as less ignorant and less immoral.\n  - The studies differ in task. Short generic \"why I'm X\" statements pass easily; detailed argument content on a specific policy fails more.\n- **Read.** Yeomans: abstract and long excerpts. Mastroianni: abstract and the author's long blog post. Brand: abstract and long excerpts.\n- **Source.** https://doi.org/10.1177/13684302211014582 ; https://osf.io/preprints/psyarxiv/39az8_v1 ; https://www.experimental-history.com/p/ideological-turing-test ; https://doi.org/10.1111/cogs.70126\n- **For the project.** The test supplies a method for judging a lens: does the holder accept it as theirs, and can others tell it from the holder's own words? Do not trust the system's or a reader's sense that the lens is accurate. The test results also show that imitations drift toward less extreme positions, so a system that infers a holder's view may flatten it. The \"pass\" criterion needs care: passing shows ability to mimic, not truth of belief.\n\n### 7b. Rapoport's rules, charitable restatement and receptiveness\n\n- **Found.** Dennett's version of Rapoport's rules: restate the target's position so well that they say \"I wish I'd thought of putting it that way\", list points of agreement, say what you learned, and only then criticise. Rapoport attributes the first rule to Carl Rogers (Dennett's footnote). The one experiment I found on faithful restatement (Younis 2026, three preregistered studies, third-person readers of exchanges about charities) found that misrepresenting the opposition lowered trustworthiness and reasonableness judgments, but did not lower persuasion. Accurate restatement had no advantage over no restatement. The author notes that Dennett's claim concerns the opponent being restated, which his design did not test. Yeomans and colleagues (2020) built an algorithm for \"conversational receptiveness\": acknowledgement, hedging, positive framing. Receptive writers were rated better partners, receptive Wikipedia editors drew fewer personal attacks, and a \"recipe\" intervention made messages more persuasive to opponents (standardised beta 0.24). Writers following the recipe found it harder.\n- **Read.** Dennett excerpts. Younis: abstract and long excerpts. Yeomans et al.: abstract and long excerpts.\n- **Source.** http://tparent.net/Dennett_excerpts.pdf ; https://doi.org/10.1007/s10503-026-09692-5 ; https://doi.org/10.1016/j.obhdp.2020.03.011\n- **For the project.** I found no test of the full four-step sequence. The nearest tests are Tuller (1d) and Younis. The practical point: the person being restated has to be the one who judges it. That is the holder sign-off the project already has in mind.\n\n### 7c. Double crux\n\n- **Found.** A technique from the Center for Applied Rationality. Each party seeks a statement B such that B is a crux for both: if B turned out different, each would change their view on the main claim A. The aim is a shared causal model, not agreement on A. It includes \"operationalise the disagreement\" and \"resonate\" steps.\n- **Read.** Two LessWrong write-ups (Duncan Sabien's 2017 post; Eli Tyre's \"The basic double crux pattern\").\n- **Source.** https://www.lesswrong.com/posts/exa5kmvopeRyfJgCy/double-crux-a-strategy-for-resolving-disagreement\n- **For the project.** I found no controlled test of double crux itself. The nearest measured work is in 9a. The technique treats finding a shared crux as progress even without agreement, which matches the project's stance. Note the requirement that the crux be crux-for-both; the measured work below found many cruxes were crux-for-one.\n\n### 7d. Adversarial collaboration\n\n- **Found.** Mellers, Hertwig and Kahneman (2001) describe a protocol in which disagreeing scholars agree in advance on tests that could change their minds, with an arbiter, and publish together. Clark and Tetlock (2021, chapter 2023) propose it as a science reform, and stress that each side should articulate the other's position so that the other feels fairly characterised, not caricatured.\n- **Read.** Abstract of Mellers et al.; long excerpt of Clark and Tetlock.\n- **Source.** https://journals.sagepub.com/doi/10.1111/1467-9280.00350 ; https://www.researchgate.net/publication/352101647_Adversarial_Collaboration_The_Next_Science_Reform\n- **For the project.** It gives a template for pre-registering what each holder would count as changing their view, which is a crux elicitation. Its measured outcomes in political settings are in 9a.\n\n---\n\n## 8. Simulating people with language models\n\n### 8a. Argyle and colleagues 2023, \"Out of one, many\", Political Analysis\n\n- **Found.** GPT-3 conditioned on real people's sociodemographic backstories produced \"silicon samples\" whose answers matched the human subgroup's response patterns across survey data. The authors call this \"algorithmic fidelity\". They state that it \"does not imply that the model can simulate a specific individual\".\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1017/pan.2023.2\n- **For the project.** The strongest pro-simulation paper says in its own text that it does not license simulating an individual. A lens is about an individual holder.\n\n### 8b. Bisbee and colleagues 2024, Political Analysis\n\n- **Found.** ChatGPT personas answering feeling-thermometer questions matched the averages of the 2016 to 2020 ANES. But the variation was too small, 48 percent of regression coefficients differed significantly from the human data (the sign flipped in 32 percent of those), results shifted with small prompt changes, and the same prompt gave significantly different results three months later.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://doi.org/10.1017/pan.2024.5\n- **For the project.** A simulated holder would give answers that change over time with the model, with no way for the holder to notice. The project's requirement that lenses be reproducible is not met by a model that drifts.\n\n### 8c. Wang, Morgenstern and Dickerson 2025, Nature Machine Intelligence\n\n- **Found.** Analytical argument plus human studies (3,200 participants, 16 identities, four models): the authors argue, and show empirically, that models misportray groups (the paper compares model output with out-group imitations and with in-group self-portrayals) and flatten them (in the supplement, model responses were less diverse than human in-group responses across all question types and groups). Prompts using an identity label can essentialize identities. Inference-time fixes reduce but do not remove the harms. The authors connect this to epistemic injustice about lived experience.\n- **Read.** Abstract and supplementary excerpts.\n- **Source.** https://www.nature.com/articles/s42256-025-00986-z\n- **For the project.** The nearest direct evidence on what goes wrong when a model speaks for a group member. Out-group-style portrayals are what a system would produce when it fills in a holder's weights from a label. It supports the \"holders' own words\" rule and gives a concrete ground for it beyond principle.\n\n### 8d. Santurkar and colleagues 2023; Cheng, Durmus and Jurafsky 2023\n\n- **Found.** Santurkar et al. found substantial misalignment between language-model opinions and those of 60 U.S. demographic groups, about as large as the Democrat-Republican gap on climate change, even after steering the model toward a group. Cheng et al. found that GPT-3.5 and GPT-4 persona portrayals contain more racial stereotypes than human-written ones with the same prompts, with patterns of othering.\n- **Read.** Abstracts only (via aggregator pages).\n- **Source.** https://arxiv.org/abs/2303.17548 ; https://arxiv.org/abs/2305.18189\n- **For the project.** Same direction as 8c from other teams.\n\n### 8e. Park and colleagues 2024, \"Generative agent simulations of 1,000 people\"\n\n- **Found.** Agents built from a two-hour interview transcript with each of 1,052 people (about 6,500 words each) predicted that person's General Social Survey answers at 85 percent of the person's own two-week test-retest accuracy. Agents built from demographics reached 71 percent and from a self-written paragraph 70 percent. Personality scores: 0.80. Economic games: 0.66. Cutting 80 percent of the transcript still gave 0.79. A summary of the interview, which removes the person's wording, gave 0.83. They report lower accuracy bias across racial and ideological groups than with demographic prompts.\n- **Read.** Abstract and long excerpts.\n- **Source.** https://arxiv.org/abs/2411.10109\n- **For the project.** Two readings. For: the method that worked best used the person's own account, not a label; that agrees with a lens made of holders' own words. Against: this still predicts coarse answers to closed survey questions, not reasons or weights; the test is whether the agent gives the same answer as the person, not whether the person would accept the agent's explanation; and the agent cannot be audited from the outside. Notably, a summary of the interview did nearly as well, so the benefit came from content, not voice.\n\n### 8f. Models as mediators and persuaders (not simulating a person)\n\n- **Found.** Tessler and colleagues 2024 (Science; N = 5,734): an AI mediator generated group statements from participants' opinions and critiques. Participants preferred them to statements by human mediators, rated them clearer and more informative, and groups ended less divided; unmediated exchange did not do this. Dissenting views were incorporated. In a published response, the authors note participants did not know the source in the direct comparison; commentators flagged that trust in an AI mediator may change outcomes. Costello, Pennycook and Rand 2024 (Science; N = 2,190): three-round dialogues with GPT-4 Turbo about a person's own conspiracy belief reduced it by about 20 percent, durable at two months; a fact-checker rated 99.2 percent of sampled claims true.\n- **Read.** Abstracts and long excerpts of both.\n- **Source.** https://www.science.org/doi/10.1126/science.adq2852 ; https://www.science.org/doi/10.1126/science.adq1814\n- **For the project.** Models can do useful work in deliberation without playing a holder, but both systems are tuned to produce agreement or belief change. The Habermas Machine optimises group approval. That is the convergence aim the project rejects. If a model is used at all, keep it to tasks like finding candidate premises in a holder's text for the holder to confirm, and not to writing the common ground.\n\n---\n\n## 9. Cruxes, \"what would change your mind\", and minimal differing premises\n\n### 9a. Forecasting Research Institute work on AI-risk disagreement\n\n- **Found.**\n  - Existential Risk Persuasion Tournament (169 forecasters): large disagreement and minimal convergence after months of debate, the largest on AI risk. The authors list as a puzzle why incentivised forecasters did not converge.\n  - Roots of Disagreement (22 participants, \"concerned\" and \"skeptical\"): median risk of AI-caused catastrophe by 2100 went from 25 to 20 percent and from 0.10 to 0.12 percent, so almost no movement. Each side summarised the other's arguments well. Participants engaged 25 to 100 hours. The best crux resolving by 2030 was expected to close about 5 percent of the gap (about 1.2 points of a 22.7-point gap). The two groups' most informative cruxes were mostly different. The skeptics would update on real-world demonstrations rather than theoretical arguments. Agreement was greater over longer horizons (30 and 40 percent for severe harm over a thousand years).\n  - Conditional Trees (21 domain experts and 3 superforecasters, 75 questions): a method for generating cruxes. The average crux question resolving in 2030 was nine times more informative than comparison questions on forecasting platforms (p = .025).\n  - Rosenberg and colleagues 2025 (Risk Analysis): two adversarial collaborations with incentives for respectful perspective-taking. Neither group changed long-term risk estimates, but both preregistered cruxes that would sway them.\n- **Read.** Abstracts and long excerpts of the summary pages; Karger's interview was read in part. The papers themselves not read in full.\n- **Source.** https://forecastingresearch.org/research/existential-risk-persuasion-tournament ; https://forecastingresearch.org/research/roots-of-disagreement-on-ai-risk ; https://forecastingresearch.org/research/ai-conditional-trees ; https://doi.org/10.1111/risa.70023\n- **For the project.** The strongest measured case of \"show people the premises they differ on\". Three results matter. People can be made to understand each other without moving. Cruxes are often crux-for-one, not for both. And listing them may not narrow the gap much. All of this supports displaying disagreement without promising convergence. The sample was expert, small and about one topic. The authors also list a fourth hypothesis, broad worldview differences that go beyond the topic, and say the evidence favours that one and a third (differences about the long run) over the first two.\n\n### 9b. Velez, Liu and Clifford 2025, \"When information affects attitudes\" (preprint)\n\n- **Found.** Using language models, the authors elicited people's attitudes and the \"focal beliefs\" they described as central to them. Participants (N = 2,468 and 1,910) were randomly given a model-written factual counterargument aimed at the focal belief, at an attitude-relevant but unmentioned \"distal\" belief, or a placebo. Both reduced belief strength. Focal-belief counterarguments produced attitude change about twice as large, and more durable (some persistence at about ten days).\n- **Read.** Abstract only. It is not peer reviewed.\n- **Source.** https://doi.org/10.33774/apsa-2025-m6rbw\n- **For the project.** This is the nearest test of \"the belief a view rests on is the one that matters\". It supports letting holders mark which premises their conclusion depends on. Note it measures persuasion with model-written arguments, which the project would not do.\n\n### 9c. Argument maps and biased evaluation\n\n- **Found.** Cullen and Sharma (short report, 2018; pilot on Mechanical Turk; border-wall argument): showing the same argument as a structured map rather than prose removed the link between prior attitude and how intelligent the argument was judged (difference between opponents' ratings in prose and map conditions d = 0.63). A jumbled map that looked similar but had no logical structure did not help. Low-reflection opponents also judged supporters as less morally bad in the map condition. Follow-up studies were preregistered (2021 to 2022); I did not find published results. A small Northwestern pilot (N = 38) of a gamified argument-mapping discussion reported more openness and less antipathy, labelled by the authors as highly exploratory.\n- **Read.** The short report in full, the preregistration abstracts, and a working-paper abstract. None is peer reviewed.\n- **Source.** https://philarchive.org/archive/CULSROv1 ; https://doi.org/10.17605/osf.io/xjtyf ; https://www.ipr.northwestern.edu/our-work/working-papers/2025/wp-25-27.html\n- **For the project.** This is the only evidence I found on a structured claim-graph display as such, and it is weak: a pilot with a control that fits the project's own worry (a map that merely looks smart). It points the right way. Treat it as a reason to run the test, not as a result.\n\n### 9d. \"What would have to be true for you to change your mind?\"\n\n- **Found.** I found no controlled study of the bare question and its later effect on belief. The related work is above: cruxes in 9a (preregistered cruxes were produced but did not change estimates in the short run), accountability in 1d, and Velez et al. in 9b. A small study (N = 155; O'Leary and Fletcher 2024) found that people shown a counterargument updated whether they reflected deeply or shallowly; the advantage of deep reflection was not significant (p = .055).\n- **Read.** Abstract and long excerpts of O'Leary and Fletcher. The rest is by absence of search results.\n- **Source.** https://doi.org/10.3390/jintelligence12050047\n- **For the project.** The claim that this question reliably opens minds rests on practice and opinion, not on a measured effect I could find. If the project asks it, ask it as a record of the holder's stated condition, which has value for legibility whether or not it changes anyone's mind.\n\n### 9e. Showing what the other side believes\n\n- **Found.** In the megastudy (3d), correcting misperceptions of rival partisans' views was the best-performing strategy for support of undemocratic practices and partisan violence, but only when the misperception corrected was about anti-democratic attitudes or dehumanization. Correcting policy-position misperceptions did not help on those outcomes.\n- **Read.** As 3d.\n- **Source.** As 3d.\n- **For the project.** A lens that shows real stated positions has a measured precedent when it corrects exaggerated beliefs about the other side. What it corrects matters more than that it corrects.\n\n---\n\n## What this means for the what-if view and the lenses (my reading, not findings)\n\n- **Source of premises.** Every accuracy result (1a to 1d, 7a, 8c, 8e) points one way: the premises, weights and wording must come from the holder. The what-if view should refuse to run on inferred premises, or show them marked as inferred.\n- **Do not promise humility or moderation.** The robust effect of explaining mechanisms is reduced felt understanding (4b). Moderation is not established, and it fails for protected values (4c). Frame the benefit as clarity.\n- **Pair logic with person.** Arguments alone did nothing in field trials (3a). Experiences of harm earned more respect than facts (3c). Offer the holder's own story as an optional layer under a premise.\n- **Show the structure of the difference honestly.** Overlapping consensus (6a), incompletely theorized agreement (6b) and crux-for-one (9a) are all things the view can display. Deep disagreement (6c) means some differences have nothing underneath. The view should be able to say \"this is where the disagreement bottoms out\".\n- **Make relevance a first-class parameter.** The project's attackable relevance claims fit the literature on deep disagreement better than the premise-level view alone.\n- **Validate with the holder.** Use holder sign-off in the manner of the ideological Turing test and Rapoport's first rule, with accountability (1d). Expect model or reader judgments of accuracy to be unreliable (7a, 8b).\n- **Guard against false confidence.** Perspective-taking raised confidence without accuracy (1a), people overrate their own predictions (1b), and feeling understood is not being understood (1g). A mind-meld metaphor invites this. Label what was said and what was derived.\n- **Keep models out of the holder's voice.** Evidence in 8 supports the rule. If models are used, restrict them to finding candidate premises for the holder to accept or reject. Avoid objective functions that reward agreement (8f).\n- **Measure the right things.** The outside literature measures attitude change, animosity and respect. If the project's aims are legibility and attunement, it needs its own measures: holder-accepted accuracy, felt understanding versus actual (1g), and whether readers can state the other side's position in terms the other side accepts (7a).\n\n---\n\n## What I could not verify\n\n- Fogelin 1985 itself, Mercier and Sperber 2011 in full, Rawls's *Political Liberalism*, and Landemore's book were not read. I used abstracts, a Stanford Encyclopedia entry and secondary descriptions.\n- Megastudy effect sizes in the preprint may differ from the published version.\n- Wood and Porter was read at summary level only.\n- I did not read Levy's results on affective polarization.\n- Two replication theses on the illusion of explanatory depth showed up only in search summaries. I left them out.\n- I found no controlled test of double crux, of steelmanning as a full sequence, or of the bare question \"what would change your mind\".\n- The German intellectual-humility experiments (APA PDF) were read without author names.\n- All studies are Western, mostly U.S. or U.K., and mostly online.\n\n---\n\n## Source list\n\n1. Ancell, A. Review of Landemore, *Democratic Reason*, Episteme. https://www.cambridge.org/core/journals/episteme/article/abs/democracy-isnt-that-smart-but-we-can-make-it-smarter-on-landemores-democratic-reason/025A01004C498E2CF398B5AD969A2F18\n2. Argyle et al. 2023, Political Analysis. https://doi.org/10.1017/pan.2023.2\n3. Bail et al. 2018, PNAS. https://www.pnas.org/doi/full/10.1073/pnas.1804840115\n4. Benson 2021, Synthese. https://link.springer.com/article/10.1007/s11229-021-03161-0\n5. Bisbee et al. 2024, Political Analysis. https://doi.org/10.1017/pan.2024.5\n6. Bowes et al. 2020, J. Research in Personality. https://www.sciencedirect.com/science/article/abs/pii/S0092656620300805\n7. Brand, Brady and Stafford 2025, Cognitive Science. https://doi.org/10.1111/cogs.70126\n8. Bruneau and Saxe 2012, JESP. https://doi.org/10.1016/j.jesp.2012.02.017\n9. Cheng, Durmus and Jurafsky 2023. https://arxiv.org/abs/2305.18189\n10. Clark and Tetlock, Adversarial collaboration. https://www.researchgate.net/publication/352101647_Adversarial_Collaboration_The_Next_Science_Reform\n11. Costello, Pennycook and Rand 2024, Science. https://www.science.org/doi/10.1126/science.adq1814\n12. Crawford and Ruscio 2021, Psychological Science. https://doi.org/10.1177/0956797620972367\n13. Cullen and Sharma 2018 short report. https://philarchive.org/archive/CULSROv1 ; preregistration https://doi.org/10.17605/osf.io/xjtyf\n14. Curry 2019, Behavioral Scientist. https://behavioralscientist.org/whats-wrong-with-moral-foundations-theory-and-how-to-get-moral-psychology-right/\n15. Damen, Pollmann and Grassow 2021, Frontiers in Communication. https://www.frontiersin.org/journals/communication/articles/10.3389/fcomm.2021.611187/full\n16. Dennett, *Intuition Pumps* excerpt (Rapoport's rules). http://tparent.net/Dennett_excerpts.pdf\n17. Eyal, Steffel and Epley 2018, JPSP. https://doi.org/10.1037/pspa0000115\n18. Feinberg and Willer 2015, PSPB. https://journals.sagepub.com/doi/abs/10.1177/0146167215607842\n19. Feinberg and Willer 2019, Compass. https://compass.onlinelibrary.wiley.com/doi/10.1111/spc3.12501\n20. Fernbach et al. 2013, Psychological Science. https://doi.org/10.1177/0956797612464058\n21. Fishkin et al. 2021, APSR. https://doi.org/10.1017/s0003055421000642\n22. Forecasting Research Institute: XPT https://forecastingresearch.org/research/existential-risk-persuasion-tournament ; Roots of Disagreement https://forecastingresearch.org/research/roots-of-disagreement-on-ai-risk ; Conditional Trees https://forecastingresearch.org/research/ai-conditional-trees\n23. Gray et al. 2022, Personality and Social Psychology Review. https://journals.sagepub.com/doi/10.1177/10888683221124741\n24. Hawkins et al. 2026, Political Communication. https://doi.org/10.1080/10584609.2026.2612739\n25. Itzchakov et al. 2018, PSPB. https://doi.org/10.1177/0146167217747874 ; 2017 https://pubmed.ncbi.nlm.nih.gov/27856728/\n26. Kahan et al. 2017, Behavioural Public Policy. https://www.cambridge.org/core/journals/behavioural-public-policy/article/abs/motivated-numeracy-and-enlightened-selfgovernment/EC9F2410D5562EF10B7A5E2539063806\n27. Kahan et al. 2017, Political Psychology. https://doi.org/10.1111/pops.12396\n28. Kalla and Broockman 2020, APSR. https://doi.org/10.1017/s0003055419000923\n29. Kalla and Broockman 2022, APSR. https://doi.org/10.1017/s0003055422000132\n30. Kalla and Broockman 2023, AJPS. https://onlinelibrary.wiley.com/doi/10.1111/ajps.12657\n31. Koetke and Schumann, J. Positive Psychology. https://www.tandfonline.com/doi/full/10.1080/17439760.2024.2352739\n32. Kubin et al. 2021, PNAS. https://www.pnas.org/doi/full/10.1073/pnas.2008389118\n33. Landemore, *Democratic Reason*. https://press.princeton.edu/books/paperback/9780691176390/democratic-reason\n34. Levy 2021, American Economic Review. http://www.roeelevy.com/Papers/Social_Media_and_Polarization.pdf\n35. Lieberman et al. 2023, Open-mindedness interventions review (preprint). https://doi.org/10.31234/osf.io/rg46s\n36. Mastroianni, Dana and Nichols 2024 (preprint). https://osf.io/preprints/psyarxiv/39az8_v1 ; blog https://www.experimental-history.com/p/ideological-turing-test\n37. Mellers, Hertwig and Kahneman 2001, Psychological Science. https://journals.sagepub.com/doi/10.1111/1467-9280.00350\n38. Mercier and Landemore 2012, Political Psychology. https://doi.org/10.1111/j.1467-9221.2012.00873.x\n39. Mercier and Sperber 2011, BBS. https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/why-do-humans-reason-arguments-for-an-argumentative-theory/53E3F3180014E80E8BE9FB7A2DD44049\n40. Nyhan and Titiunik 2024, Science policy forum (not read). https://doi.org/10.1126/science.ads7371\n41. O'Leary and Fletcher 2024, J. Intelligence. https://doi.org/10.3390/jintelligence12050047\n42. Park et al. 2024. https://arxiv.org/abs/2411.10109\n43. Persson et al. 2021, Cognition. https://doi.org/10.1016/j.cognition.2021.104768\n44. Porter and Schumann 2018, Self and Identity. https://doi.org/10.1080/15298868.2017.1361861\n45. Porter et al. 2022, Nature Reviews Psychology. https://www.nature.com/articles/s44159-022-00081-9\n46. Pritchard 2018, Topoi. https://doi.org/10.1007/s11245-018-9612-y\n47. Ranalli, Synthese (online 2018; issue 2020). https://link.springer.com/article/10.1007/s11229-018-01956-2\n48. Rawls entry, Stanford Encyclopedia. https://plato.stanford.edu/entries/rawls/\n49. Reis, Lemay and Finkenauer 2017, Compass. https://doi.org/10.1111/spc3.12308\n50. Rosenberg et al. 2025, Risk Analysis. https://doi.org/10.1111/risa.70023\n51. Sabien 2017, \"Double Crux\". https://www.lesswrong.com/posts/exa5kmvopeRyfJgCy/double-crux-a-strategy-for-resolving-disagreement\n52. Santurkar et al. 2023. https://arxiv.org/abs/2303.17548\n53. Schroeder, Kardas and Epley 2017, Psychological Science. https://doi.org/10.1177/0956797617713798\n54. Sloman and Vives 2022, Cognition. https://doi.org/10.1016/j.cognition.2022.105146\n55. Stanford Encyclopedia, Disagreement. https://plato.stanford.edu/entries/disagreement/\n56. Suhler and Churchland 2011, J. Cognitive Neuroscience. https://doi.org/10.1162/jocn.2011.21637\n57. Sunstein 1995, Harvard Law Review. https://doi.org/10.2307/1341816\n58. Tappin, Pennycook and Rand 2021, JEP: General. https://doi.org/10.1037/xge0000974\n59. Tessler et al. 2024, Science. https://www.science.org/doi/10.1126/science.adq2852\n60. Tuller et al. 2015, JESP. https://doi.org/10.1016/j.jesp.2015.02.003\n61. Velez, Liu and Clifford 2025 (preprint). https://doi.org/10.33774/apsa-2025-m6rbw\n62. Voelkel et al. 2024, Science. https://www.science.org/doi/10.1126/science.adh4764 ; preprint https://doi.org/10.31219/osf.io/y79u5\n63. Voelkel and Feinberg 2018. https://pmc.ncbi.nlm.nih.gov/articles/PMC6295651/\n64. Voit, Twardawski and Fischer 2026, JESP. https://www.sciencedirect.com/science/article/pii/S0022103126000016\n65. Wang, Morgenstern and Dickerson 2025, Nature Machine Intelligence. https://www.nature.com/articles/s42256-025-00986-z\n66. Wood and Porter 2019, Political Behavior. https://www.doi.org/10.1007/s11109-018-9443-y\n67. Yeomans 2022, Group Processes and Intergroup Relations. https://doi.org/10.1177/13684302211014582\n68. Yeomans et al. 2020, OBHDP. https://doi.org/10.1016/j.obhdp.2020.03.011\n69. Younis 2026. https://doi.org/10.1007/s10503-026-09692-5\n70. Zakharin and Bates 2021, PLOS ONE. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0258910\n71. Northwestern IPR working paper, \"Point Taken\" pilot (WP-25-27). https://www.ipr.northwestern.edu/our-work/working-papers/2025/wp-25-27.html\n72. Sgambati and Ayduk 2023 (intellectual humility and polarization over time). https://doi.org/10.1111/pops.12890\n73. Effects of intellectual humility in the context of affective polarization (APA, authors not captured). https://www.apa.org/pubs/journals/releases/psp-pspi0000462.pdf\n"}