Notes
Research, experiments, and reflections on AI and design, published along with the process of making them.
AI and society (13)
What is happening to people and society as AI spreads, across research fields.
-
Cognitive Surrender: The Lineage of Handing Judgment to AI, What Generative AI Changed, and the Current Evidence on Countermeasures
2026-11-06
Cognitive surrender refers to the tendency to adopt AI output as one's own judgment without scrutiny and to stop deliberating altogether. Shaw and Nave (2026, not yet peer reviewed) named the phenomenon. When the AI was wrong, participants followed it on 79.8% of the trials in which they consulted it, accuracy fell from 45.8% without AI to 31.5%, and confidence rose rather than fell. The same pattern has been studied since aviation incident reporting in the 1970s under the names automation complacency and automation bias. Generative AI widened the problem from narrow alarms used by trained professionals to fluent, general-purpose advice used by everyone, and it shifted the concern from following errors to losing opportunities to learn and losing skill (a 17% drop on an exam taken without AI; a fall in the adenoma detection rate of non-AI colonoscopy from 28.4% to 22.4% after AI was introduced). The center of research moved accordingly, from human factors to HCI, the learning sciences, and cognitive psychology. On countermeasures, adding explanations and showing general warnings have weak effects. Local cues about where errors are likely, showing sources and inconsistencies, first-person expressions of uncertainty, outcome feedback with incentives, and assessment without AI work under conditions. Procedures in which people record their own judgment first reduce reliance on correct AI as well, and designs in which the AI argues back still have little evidence. Every effective design carries costs: people like it less, and it helps mostly those who already enjoy thinking.
-
Does a Dream Die When AI Fulfills It? Checking "Instant Dream Death," "Dream Deflation," the "Obsession Gap," "Passion Capitalism," and "Life Prototypes" Against the Academic Literature
2026-10-11
A post on X by paji_a on October 10, 2026 called it the "instant death of dreams" when generative AI fulfills in 20 minutes a dream pursued for 20 years, and went on to propose "dream deflation," an "obsession gap," "passion capitalism," and an "age of life prototypes." This note checks the five claims against 47 academic sources and research reports from psychology, labor economics, HCI, and sociology. Dream deflation is supported when limited to the price of making things and to the value of what made things signal. After an image-generation model leaked, the average price of anime-style commissions on an outsourcing platform fell by 64% while orders rose by 121% (Zhang et al.), and the tendency of employers to pay more for workers with applications tailored to the posting disappeared after LLMs arrived (Galdin and Silbert, not peer reviewed). For instant dream death, neither the compression of time nor the loss of satisfaction has been measured directly. The measured time savings concern tasks of tens of minutes to a few hours (40% less time, Noy and Zhang 2023), and experienced developers took 19% longer on real tasks (Becker et al. 2025, not peer reviewed). That fulfilling a dream may not feel good is plausible given that anticipation itself carries value, that people overestimate their emotional reactions to events, that attaining extrinsic aspirations is not related to psychological health, and that people feel little ownership of text written with AI, but no study was found that measured the satisfaction of people who fulfilled a long-held dream with AI. The first half of the obsession gap (AI narrows ability gaps) depends on the task: in customer support, productivity rose by 36% for the least skilled quintile (Brynjolfsson et al. 2025), whereas among Kenyan entrepreneurs, initially low performers did about 10% worse (Otis et al., not peer reviewed). No study was found showing that differences in obsession become valuable; grit correlates only moderately with performance (Credé et al. 2017), and being "abnormally absorbed" resembles obsessive passion, which carries costs (Vallerand et al. 2003). Passion capitalism is already an object of criticism in sociology and social psychology: poor treatment of workers seen as passionate is judged more legitimate (Kim et al. 2020), and venture capitalists funded on preparedness rather than passion (Chen et al. 2009). The prospect of discovering one's true passion conflicts with the finding that people who think interests are found rather than developed lose interest more when things become difficult (O'Keefe et al. 2018). Life prototyping is supported as the process of trying out provisional selves in new roles (Ibarra 1999), but changeable decisions lowered satisfaction (Gilbert and Ebert 2002).
-
Why Do Men in Their 40s and 50s Push AI So Hard? A Comparison with the Famicom, Pagers, and the Internet
2026-10-03
This note examines why people in their 40s and 50s in 2026 (born 1976–1986) advocate generative AI so strongly, by comparing AI with the information revolutions this generation lived through from childhood. It draws on 22 sources: Japanese official statistics, the theory of technology generations, the sociology of expectations, empirical studies of generative AI use, and the social history of Japan's information society. First, it confirms that loud advocacy does not mean heavy use. In a 2025 survey by Japan's Ministry of Internal Affairs and Communications (MIC), 78.2% of people in their 20s had used generative AI, compared with 49.0% in their 40s and 44.2% in their 50s. Second, to the three revolutions in the request (the Famicom, pagers, the internet) it adds two more: the mobile internet (i-mode, 1999) and the smartphone (household ownership rising from 9.7% to 49.5% between 2010 and 2012). On the Famicom, outdoor play had already fallen below indoor play around 1965, and the Science Council of Japan attributes the main cause to television. The Famicom did not start the move indoors; it filled play that had already moved indoors. Laid side by side, the revolutions that arrived in the middle of this generation's formative years (roughly ages 15–25) were the internet and the mobile internet: when individual internet use passed 20% at the end of 1999, they were 13–23. Under the definition of technology generations, this places them at the entrance of the internet generation. From verified facts the note derives four hypotheses. (1) Having watched, in their formative years, a new information technology change everything gives them a template for reading AI. (2) That template works as recalling past futures and using them in present claims about the future (Brown & Michael 2003). (3) The memory of Japan leading with i-mode but becoming isolated from world markets (Kushida 2011), and of its electronics losing competitiveness (Vogel 2013), strengthens a sense of 'don't miss the boat this time'. (4) From the Famicom to the mobile internet they were the young people adults worried about; with AI, for the first time, they are on the deciding side at work. Respondents to MIC's corporate survey were managers or above, and 86.4% in Japan said their company used generative AI in at least one task. The male skew may be explained by boys having entered information technology through games and computers (Ito & Okabe 2005) and by higher generative AI use among men (7.5 points in the US, about 1.8 times in Japan), but no study was found that measures advocacy by age and gender.
-
Does Seeming Common Online Mean It Is Common?
2026-08-30
Is "I see it a lot" a reliable cue for "a lot of people do it"? This note gathers 303 academic sources bearing on that question (29 on frequency cues and prevalence estimation, 26 on high-volume output by few actors, 27 on the generative-AI-specific information environment, 26 on self-fulfillment and norm dynamics, 27 on the representativeness of digital trace data, 117 from an additional collection covering the last three years, and 51 from a differential check against echo chamber research). The phenomenon of frequency and headcount coming apart has been measured since well before generative AI. Weaver et al. (2007), across 6 experiments, showed that people who heard one person repeat the same opinion three times estimated "this opinion is widely supported" almost as strongly as people who heard three different people each say it once, and the effect did not disappear even when participants were explicitly told the speaker was a single person. In Pew Research Center's (2019) probability-sample survey, the top 10% of U.S. adult Twitter users produced 80% of all tweets, and van Mierlo (2014) measured that the top 1% of users wrote 74.7% of posts across four health social networks, while the bottom 90% wrote only 1.3%. Rao and Reiley (2012), noting that the marginal cost of spam is close to zero, formalized the structure by which send volume decouples from the number of senders, and Lerman, Yan, and Wu (2016) formalized how skew in a network's degree distribution alone can make a globally rare attribute look like a majority under local observation. Over the last three years, the picture has shifted substantially. The path by which exposure moves the judgment of "what is normal" has now been directly demonstrated. In the third experiment of Glickman and Sharot (2025, *Nature Human Behaviour*, N=1,401), merely showing three Stable-Diffusion-generated images of a "financial manager," 1.5 seconds each, raised the share of participants choosing a white man as the most manager-like from 32.36% to 38.20% (p=0.04; the control group showed no significant difference). AlDahoul et al. (2025) showed that exposure to AI-generated faces moves perceptions of race and gender, and does so regardless of whether the images are disclosed as AI-generated. Brady et al. (2026, a *Nature* preregistered report) followed 2,000 people on Bluesky for eight weeks and confirmed, on a real platform, that an engagement-optimizing feed lowers the accuracy of social-norm perception while a diversifying algorithm improves it. Geber and Stahel (2026, N=1,021), however, reported that the type of recommendation had no significant effect on norm perception, and Liu et al. (2025, *PNAS*), across four YouTube experiments (roughly 9,000 participants total), could not detect a short-term polarization effect from manipulating filter-bubble conditions. Measurement of volume has advanced as well. Liang et al. (2024) estimated that 6.5% to 16.9% of the body text of peer reviews at AI conferences had been substantially altered by an LLM; Allaham and Diakopoulos (2026) measured that about 16% of generative search's cited sources are already AI-generated content; and Lee et al. (2025, *PNAS Nexus*) showed that estimates of the share of harmful posters overstate the measured value by up to roughly 100-fold. The detecting side has not kept up. A watermark-removal attack succeeded against 7 schemes at near-100% (Cheng et al. 2025, at a cost of $0.88 per million tokens); C2PA was concluded, under independent evaluation, not to meet its stated security goals (Golaszewski et al. 2026); and bot detection loses up to 29.6% performance to LLM-based evasion (Feng et al. 2024). Naming has arrived first as well: Schroeder et al. (2026) put forward the term "synthetic consensus" in a *Science* policy forum (unaccompanied by empirical evidence). The measuring instruments themselves have been contaminated too: Westwood (2025, *PNAS*) showed that an autonomous LLM agent can evade a survey's quality controls in 99.8% of 6,000 trials. What remains as a gap has narrowed to a single point. Even when Daikeler et al. (2025) systematically reviewed 58 data-quality frameworks, no error type existed for the case where the sender lies outside the population, and it was not even named as a gap. Alsalti et al. (2026), in a section on the coverage error of the total-survey-error framework, put this premise into words, but it has not reached formalization as a subtype. The differential check against echo chamber research, the closest neighboring concept, was also carried out, adding 51 sources. Within echo chamber research, the causal locus itself is contested (Sunstein places it in the receiver's own selection, Pariser in the algorithm, Cinelli et al. 2021 in homophily, and Bakshy et al. 2015 separate the two, finding that individual clicking accounts for a roughly 70% reduction that exceeds the algorithm's roughly 15% reduction), yet the premise that a real human being stands behind each observed item is never made explicit across the 24 sources reviewed — this follows from the structure of the methodology itself, since a node being human is the starting point. The accumulation of disconfirming evidence is substantial as well: Guess (2021) measured the overlap in partisan media diets at 50% to 65%, and four studies from Meta's 2020 election research showed that even substantially moving exposure left attitude measures unmoved (Nyhan et al. 2023, *Nature*; two papers by Guess et al. 2023, *Science*; González-Bailón et al. 2023, *Science*; at a scale of 208 million people). The lineage of misestimated composition ratios carries the same structure: Ahler and Sood (2018) showed that respondents estimate the share of LGBT people among Democrats at 32% (the true figure is 6%) and the share of high earners among Republicans at 38% (the true figure is 2%), yet in both cases the estimated target is a group that actually exists, and the remedy is to present the true base rate (in Ahler 2014, an intervention moderated opinion by 8 to 13 percentage points). The effect size of selective exposure itself remains modest, at d=0.36 (Hart et al.'s 2009 meta-analysis). The difference sorts into three points: where the distortion is located (in the observer's sample, or in whether the supplied item corresponds to a human); what the remedy presupposes (a real distribution of opinion, or the existence of a true ratio); and the direction of the error (toward one's own side, toward a salient minority, or independent of either). A rebuttal holds, however, if Guay et al.'s (2025, *PNAS*) account of "rescaling under uncertainty" is correct: since ratio estimation regresses toward 0.5 regardless of input, there may be little room left for contaminated input to make it worse. This rebuttal bears on the task of ratio estimation, but it does not directly touch the task Glickman and Sharot (2025) moved — judging "who is typical."
-
Can Frequency Be Asked What It Closes? (A Novelty Reassessment)
2026-08-30
Reassessed the novelty of the handoff item "the decoupling of representation frequency from social prevalence by generative AI" against 303 items of literature (135 classic-through-roughly-2024 items, 117 items added from the most recent three years, and 51 items on the echo-chamber difference). The verdict is stricter than the first assessment. First, the decoupling itself is already established in the literature (Weaver et al. 2007, Rao & Reiley 2012, Pew 2019, van Mierlo 2014, Lerman et al. 2016). Second, the core that the first assessment had left as a gap has now been filled. Experiment 3 of Glickman & Sharot (2025, Nature Human Behaviour, N=1,401) shows that showing just three Stable Diffusion-generated images for 1.5 seconds each shifts judgments of "who looks like a financial manager" from 32.36% to 38.20% (p=0.04), and AlDahoul et al. (2025) showed in a pre-registered experiment that exposure to AI-generated faces shifts perceptions of population composition regardless of whether disclosure is given. The causal path from exposure to normative perception has already been empirically demonstrated. Third, the naming has also been pre-empted. Schroeder et al. (2026) propose "synthetic consensus" in a policy essay in Science. Fourth, the residual intent requirement has also weakened. The 7.9 million articles of pink slime (Horne & Gruppi 2024) and the 84 potential cases of agentic flooding of government services (Schmitz et al. 2026, most of them by users acting in good faith) already exist as case studies of mass generation. Only one point remains. Even though Daikeler et al. (2025) systematically reviewed 58 data-quality frameworks, no error type exists for the case where the sender lies outside the population, and it is not even named as a gap. Alsalti et al. (2026) put this premise into words in the section on coverage error within total survey error, but they do not go so far as to formulate it as a subtype. Novelty, therefore, can be placed only on the side of a "diagnostic apparatus," not a "discovery." Here the way of cutting the question changes. Applying a stance that judges by function and does not ask who carried out the work to frequency cues shifts the question from "is the estimate accurate" to "what does that frequency close off." The two existing lines of apparatus (manipulation research, which requires intent and persona, and measurement theory, which requires the real existence of a population-side term) both depend on authorship, and replacing that with a judgment by function is itself where the contribution sits. The provisional scoring of the novelty axes is PD=3, Mat=2, Met=4, Res=not yet scored, and RW=3 if submitted as a review paper. Met≥4 also touches the single-axis exception gate. Public Opinion Quarterly is the first-choice venue for submission, with Big Data & Society as the second. The difference from echo chamber research, the nearest neighboring concept, was also examined against 51 items of literature, and the verdict is that the difference holds. Across 24 items of echo chamber research, not one paper explicitly states the premise that a real human being stands behind each observed item, and the prescriptions all either change the distribution of opinions that actually exist or present a true base rate (Ahler 2014 found opinion moderated from 8% to 13%). For items with no corresponding human, neither premise is satisfied. The finding of the four Meta 2020 election studies (that substantially changing exposure does not move attitudes) is also a tailwind, since it confirms that the object of manipulation has always remained the visibility of real people who exist. One new threat has appeared, however. Guay et al. (2025, PNAS)'s "uncertainty-based rescaling" claims that proportion estimation is, in general, a domain-wide property that regresses toward 0.5 to begin with. If this is correct, it could be met with "there is little room for the input's contamination to make things worse." The response is to separate the tasks. What UBR explains is proportion estimation, whereas what Glickman & Sharot (2025) moved was typicality judgment, which does not ask for a proportion (AlDahoul et al. 2025 do ask for proportions, so only Experiment 3 of Glickman & Sharot supports this separation). Placing proportion estimation as the dependent variable therefore gets absorbed, while placing typicality judgment and the range of options derived from it steps outside UBR's reach. This separation directly determines the experimental design. The final scoring is left to paper-pipeline's novelty-auditor.
-
The History of Slop and Countermeasures: Why the Same Wager Keeps Failing
2026-08-28
This note traces the cycle in which low-quality generated content pollutes the environment and countermeasures for selection are built in response, across 140 sources (21 on prehistory, 28 on email spam, 25 on search and bots, 29 on academic publishing, 37 on AI countermeasures since 2022) spanning the era after print to 2026. Historically, countermeasures fall into roughly five types: the recipient classifies (Bayesian filters), cost is imposed on the sender (proof-of-work), provenance is certified (SPF/DKIM/DMARC, C2PA, watermarking), personhood is proven (CAPTCHA), and a gatekeeper is placed (peer review, blacklists, community norms). And there are five types in how they break down. (1) A design that presumes the attacker lacks a certain capability becomes invalid once that capability improves (CAPTCHA was a design that bet its security on AI's unsolved problems, and a GAN solver broke 33 schemes in 0.05 seconds). (2) A countermeasure that imposes cost does not work against an opponent who can pass that cost onto someone else (Laurie & Clayton 2004 demonstrated the failure of proof-of-work). (3) When provenance is voluntary, only honest participants use it (in Durumeric 2015, DMARC policy specification stood at 1.1%). (4) Gatekeeping errors concentrate on the weaker side (blacklists misclassifying Global South journals, discrimination against Tor users, false positives against non-native speakers). (5) Countermeasures that worked concentrated verification capability among large actors. Three things are new this time only: generated content and the genuine article cannot be told apart, the impossibility of strong watermarking has been proven (Zhang et al. ICML 2024), and no peer-reviewed research within the scope of this collection was found showing that any countermeasure actually reduced the volume or circulation of slop. What the countermeasures that worked historically shared was not "telling apart" but "changing the structure."
-
Is Capability Inside the Model? A Relational Ontology of AI Evaluation and a Test for Contextuality
2026-07-19
A single deep argument that inverts one asymmetry AI evaluation quietly assumes, namely that a model 'has' a capability while measurement merely 'reveals' it, using the same symmetry-completion move by which de Broglie derived matter waves. The inversion says there is no context-independent 'true capability': measurement does not reveal capability, it brings it into being. This is offered not as a settled fact but as a falsifiable hypothesis, testable by whether capability measurement exhibits non-classical contextuality. The sheaf-theoretic apparatus has already been applied at scale to LLM word meaning with non-classicality measured empirically (Lo, Sadrzadeh & Mansfield 2025, Proc. R. Soc. A); the core proposal is to move the target from meaning to capability. Novelty is strictly limited to three points: the ontological inversion that drops realism, the transfer of the contextuality test from meaning to capability, and the unification of scattered phenomena under one ontology. 22 references (19 with confirmed bibliography, 3 flagged for primary verification).
-
AI, Psychology, and Cognitive Science Research Trends: Machine Psychology and the Cognitive Modeling of LLMs (2024–2026)
2026-07-19
An integrative summary of research trends at the intersection of large language models and psychology/cognitive science in 2024–2026, drawn from 30 peer-reviewed papers and major preprints. It tracks four clusters: machine psychology, LLMs as cognitive models, the psychology of human-AI interaction, and the methodological turn in cognitive science. 30 references.
-
Research Currents in AI and Economics: Productivity, Labor Markets, and Methods (2024–2026)
2026-07-19
A synthesis of 16 studies at the intersection of generative AI and economics that gained depth in 2024–2026, drawn from peer-reviewed journals and NBER/SSRN working papers. Organized into five clusters: productivity field experiments, labor-market exposure and employment, macro growth theory, LLMs as a research tool, and algorithmic pricing. 16 references.
-
Research Trends in AI Law and Governance: Regulation, Accountability, and AI Governance (2024–2026)
2026-07-19
A literature review of AI law and governance research from 2024–2026, drawn from 33 primary legal sources and peer-reviewed law articles. Covers the staged application of the EU AI Act and its GPAI Code of Practice, the reversal of U.S. executive orders, the multi-layered governance of NIST/OECD/UN/Council of Europe, the institutionalization of algorithmic audits, gaps in AI liability, and the litigation and scholarship over copyright and LLM training. 33 references.
-
AI and the Humanities: Digital Humanities and Large Language Models (2024–2026)
2026-07-19
A literature review of the intersection of AI (LLMs and generative AI) and the humanities, drawn from 19 peer-reviewed articles and major preprints published between 2024 and mid-2026. It traces five currents: the LLM turn in digital humanities, the reversal in transcription and OCR accuracy, the philosophy of AI around meaning and understanding, the disputes over authorship and disclosure, and critical AI studies. 19 references.
-
AI and Social Science Research Trends: Computational Social Science and LLM Agents (2024–2026)
2026-07-19
An integrated scoping summary of research currents that emerged after 2023 at the intersection of AI (especially LLMs and generative AI) and the social sciences, drawn from peer-reviewed papers and highly cited preprints. It surveys four clusters: the LLM turn in computational social science, LLMs as research tools and the validity critiques of that use, silicon sampling and generative agents, and AI as a substantive object of study. 29 references.
-
Does AI Make Smart People Smarter and Everyone Else Shallower? — A Two-Lineage Debate Between Industry and Academia
2026-06-09
Debates the Zenn article (pdfractal) 'Does AI make smart people smarter and the less smart shallower?' across two lineages: the academic lineage (peer-reviewed research) and the industry lineage (the trend reports practitioners actually read, from public institutions, research firms, and consultancies). The opposition is not 'academia vs. industry' but 'the cognitive science camp (deterioration, widening) vs. the industry report camp (= relaying labor economics = compression, leveling-up)'. The article stands on the minority cognitive-science side. The key to reconciliation is the article's §6 'inside/outside the boundary' = metacognition, which shakes hands with the jagged frontier and with the warning signals industry reports themselves emit (trough of disillusionment, literacy deficits).