An Information Environment That Presumes Slop: Design That Doesn't Bet on the Ability to Discern

Gone After Forty-Eight Hours

Show someone a video that builds resistance to misinformation. Test them on the spot, and the effect shows up. Wait two days without testing, then measure, and the effect is nowhere to be found.

That is the result Capewell and colleagues reported in 2024. The video intervention lost significant difference from the control group after 48 hours (d=0.010, p=.854), and the game-based intervention vanished within 10 days. The effect survived only in the group given an immediate posttest (d=0.416, p=.007).

In other words, the act of recalling memory itself was producing part of the effect. Maertens and colleagues’ five longitudinal, preregistered experiments (N=11,759) point in the same direction. Text- and video-based inoculation lasts about a month, but what predicts persistence is not motivation but memory, and without an interposed recall it fades.

What is at issue here is not whether the intervention is meaningless. It is whether society’s design can afford to bet on “people becoming able to tell apart.”

Where Does Raw Discernment Ability Actually Stand

According to Diel and colleagues’ systematic review (56 papers, N=86,155), human accuracy at detecting deepfakes is 55.54% (95% CI [48.87, 62.10]), a confidence interval that straddles the 50% chance level. By modality: audio 62.08%, video 57.31%, image 53.16%, text 52.00%. In logs from a Microsoft Research online game as well (over 12,500 participants, about 287,000 ratings), accuracy at detecting AI-generated images stayed at 62%, higher for images of people but dropping sharply for natural and urban landscapes.

Experts fare no better. In an experiment where 63 German university instructors judged excerpts (200–300 words) from thesis writing, accuracy was 57% for AI text and 64% for human text, not significantly different from a machine detector. The more expert the AI-written text, the harder it becomes to tell apart.

What about training? An experiment with 254 Czech speakers reports that advice alone did not raise accuracy, and that only training accompanied by explicit feedback improved both accuracy and the calibration of self-confidence. The assumption participants initially held (that “AI-written text has this kind of style”) was mistaken to begin with.

What Improves Is the Ability to Doubt

Meta-analyses of intervention studies report, on average, that there is an effect. Huang and colleagues’ meta-analysis (k=49, N=81,155) puts the overall effect at d=0.60. But broken down, the reduction in sharing intent is d=1.04 and discernment is d=0.76, while the reduction in actually believing misinformation is the smallest, at d=0.27.

Pfänder and Altay’s systematic review (67 papers, 303 effect sizes, N=194,438, 40 countries) measures this asymmetry directly. Discernment itself is high (d=1.12). But the ability to believe the truth is lower than the ability to doubt falsehood, and the difference, a skepticism bias, comes to d=0.32. When the news is politically discordant, the gap widens to 0.78.

Sometimes this shows up as a side effect. A health media-literacy intervention targeting cancer news lowered support for inaccurate headlines by 5.6 points, but also lowered support for accurate headlines by 7.6 points, and discernment did not improve. In a classroom RCT with 2,288 high-school students in northern Italy, three interventions (Civic Online Reasoning, cognitive-bias education, and inoculation theory) all failed to significantly improve discernment, and the inoculation intervention strengthened general skepticism and lowered trust in scientifically valid content. Three interventions given to 459 upper-secondary students likewise showed no lasting improvement at a three-month follow-up.

The accuracy-nudge line of work is the most thoroughly tested intervention in this field. A meta-analysis of 20 experiments (N=26,863) reports an increase in sharing discernment (b=0.038, z=7.10) and a 10% reduction in intent to share false headlines, but the effect on true news was non-significant (b=0.006, p=.150). In a preregistered direct replication, the effect size in the treatment group was comparable, but the baseline in the control group rose, shrinking the relative effect to about half. A reanalysis claiming the effect does not work on conservatives followed, and a preregistered adversarial collaboration settled the matter as a small but real effect (a 3.3–14.6% reduction for Republicans, 7.6–19.1% for Democrats).

The debate continues, but the number design should take away from it is clear. What training mainly moves is the doubting side, the effect is small, and left alone it fades.

The Designs That Worked Placed the Verifier Somewhere Else

So what did work?

In 1998, Los Angeles County mandated the posting of restaurant hygiene grade cards. Jin and Leslie’s natural experiment shows that after the mandate, hygiene scores rose, demand became sensitive to quality, and hospitalizations from food poisoning fell. It did not teach users how to tell hygiene apart. It made invisible quality visible at the storefront.

The work of verifying did not disappear. An inspector verifies, and the result is posted on the storefront. What moved is the place where the duty to verify is borne.

This is exactly the structure Akerlof formalized in 1970. In a market where quality cannot be observed, low quality drives out high quality. What Spence showed two years later was a solution in which a costly signal creates a separating equilibrium. Read in the context of slop, this becomes the question of whether an institution can supply a mark that the side able to produce cheaply cannot imitate.

But certification does not function automatically. Lizzeri showed that a monopolistic certification intermediary optimally discloses only pass/fail and captures most of the information rent. In the practice of rating, competition under the issuer-pays model induces rating inflation. Reputation systems degrade too. Pressure on raters to avoid harming the counterparty structurally inflates scores, and the system undermines its own effectiveness.

What about collective verification? What Birdwatch, the predecessor of Community Notes, introduced was not a majority vote but a bridging algorithm that prioritizes cross-partisan agreement. Chuai and colleagues analyzed 237,180 cascades and over 431 million reposts, and report that after a note was displayed, the spread of misleading posts fell 61.2%, and the odds of the poster deleting the post themselves rose 94.3%. Slaughter and colleagues’ PNAS paper, using different data, produced estimates of −46.1% for reposts, −44.1% for likes, −21.9% for replies, and −13.5% for views. That lay crowd ratings correlate with professional fact-checker ratings to a comparable degree has also been shown in a comparison of 207 articles and 1,128 raters.

Where Has Society Been Producing Trust

The grade card’s approach has a long prehistory.

What Shapin and Schaffer (1985) described was how 17th-century experimental philosophy assembled trust in facts. Boyle’s air-pump experiments had a hard limit on how many people could witness them in person. What was adopted instead was virtual witnessing, a written substitute for being present. Writing down the apparatus’s appearance, the procedure, even the number of failures, in fine detail, reconstructed the laboratory inside the reader’s head. What produced trust was not the reader’s judgment but the form of the account.

Whose account was worth believing was decided by a separate apparatus. Shapin (1994) argues that apparatus was the honor of a gentleman. The premise that someone of economically independent standing had no motive to lie underwrote the credibility of testimony. The grounds lay not in examining the content but in the speaker’s social position.

Standing as a guarantee was eventually replaced by something else. What Porter (1995) traced was the process by which experts’ personal judgment yielded to quantification and standardization. The reason for the replacement was not accuracy. It was a demand for accountability from outside. Once experts could no longer say “trust my judgment,” numbers took over the role.

Daston and Galison (2007), drawing on the history of scientific atlases, laid out objectivity as passing through three regimes. Truth-to-nature, mechanical objectivity, and trained judgment. Mechanical objectivity tried to eliminate the observer’s judgment through the apparatus, but could not eliminate it entirely, and trained judgment returned as a third regime.

What the four have in common is that trust was produced outside the individual’s ability to discern. The form of the account, standing, numbers, apparatus. None of them require the recipient to become able to tell things apart.

This framework has already been redeployed for AI. Christin (2016) mapped the dispute over algorithmic objectivity onto Daston and Galison’s three regimes. Rieder and Simon (2016) placed trust in evidence derived from big data as an extension of Porter’s framework. Anichini and Kotras (2024) examined radiology and prehistoric archaeology in practice and show that “trained judgment” resurfaces at the moment of AI adoption. In Schenk (2025)‘s experiment (793 Swiss students), judgments about the acceptability of an AI judge appeared as a conflict between mechanical objectivity and empathic understanding. The fact that the same framework has been independently invoked four times is evidence that this framework is still in use now.

Even so, the apparatus has a failure mode of its own. What Power (1997) called the audit society is a state in which the rituals of verification multiply without substantive verification accompanying them. An audit functions as a form that answers the demand for accountability, and performing the form becomes the goal in itself. The reason assessments of Community Notes are split (individual notes work while the system as a whole shows no significant effect) has this shape. That the apparatus exists and that the apparatus is working have to be measured separately.

Under What Conditions Does It Still Break Down

Assessments of the same mechanism are split.

A quasi-experiment covering the entire rollout reported that the introduction of the Community Notes feature itself produced no statistically significant reduction in engagement with misinformation. The main cause the authors cite is the delay before a note is displayed. Most spread is over within the first few hours, but reaching consensus takes time.

Another constraint lies on the side of the participants. A study analyzing 2,250 vaccine-related posts and 18 months of data showed that claims imposing greater cognitive load to fact-check are less likely to receive a note that reaches a “helpful” rating. A survey of 382 people shows the same tendency. The more effort a lie takes to check, the more it slips through collective verification.

Label design has structural side effects too. Attaching a warning to only some fake news raises the perceived accuracy of the fake news left unwarned. A field experiment that distributed in-browser credibility labels found no average effect on either the quality of news consumption or belief in misinformation. An AI-generated label lowers trust equally whether it is attached to true information or to information a human made.

Collective intelligence itself has conditions for holding. Showing people others’ estimates destroys independence among individuals, shrinks the diversity of the distribution, and erases the benefit of averaging. The design of verification needs to keep the order of taking independent first-pass judgments first and placing aggregation and disclosure after.

The Bottleneck Is the Allocation of the Burden of Verification

A sister note, AI Slop: Reading It as Outsourced Verification, Not Low Quality, read slop not as a property of the object but as the allocation of the burden of verification. The institutional numbers are consistent with that reading.

Peer review is not in short supply. By Kovanis and colleagues’ estimate, as of 2015 the total supply of reviewers exceeded demand by 15–249%. It is skewed. The top 20% of reviewers carry 69–94% of the review labor.

Strengthen the gatekeeper, and the supply side dies. As a result of Wikipedia automating quality control, the rate of new contributors being immediately reverted rose, and editor retention fell from 40.0% at the end of 2006 to 13.2% at the start of 2007. Simply adding more editors does not raise quality on its own. Quality rises only when explicit or implicit coordination accompanies the increase; a growth in headcount without coordination was actually harmful.

Lower the cost of verification, and overreliance falls. Vasconcelos and colleagues explained overreliance on AI not as a cognitive bias but as cost-benefit reasoning about verification, and showed across five studies (N=731) that overreliance falls when the cost of verification is low. Conversely, assistance does not erase effort; it displaces it. In an AI-assisted coding experiment (N=60), subjective load fell by an average of 18.2 TLX points and time-on-task shortened by 22%, while a measure of verification load partially mediated the rise in stress and fatigue that came with task complexity. Cognitive forcing functions reduce overreliance, but subjective satisfaction ends up lowest.

The irony of automation that Bainbridge wrote in 1983 applies unchanged. Automation leaves humans only the difficult residue, and a human turned into a monitor loses the opportunity to maintain skill. The decline in detection performance on monitoring tasks comes not from carelessness but from the limits of an attentional resource that requires effort.

The Pathways of Reach Are Narrowing

Separate from the question of telling things apart, there is the question of reaching them at all.

In scholarly information, the more digitization advanced, the more citations concentrated on a small number of recent papers (34 million papers, 1945–2005). In fields with a higher annual output of papers, referencing already-highly-cited papers becomes obligatory, and the canon stiffens (241 fields, 90 million papers, 1.8 billion citations). The more volume increases, the narrower the destinations reached become.

Generative search reinforces this tendency from a different direction. In a benchmark of 11,500 queries, an AI Overview was generated for 51.5% of them, and the source overlap between conventional and generative search had an average Jaccard similarity below 0.2. An analysis of over 200,000 real-world ChatGPT usage logs found that, even restricted to searchable queries, AI responses show lower diversity than Google search results across most topics. A study auditing the cited sources of generative search across 712 queries reports that about 16% were already AI-generated content.

Citations are not being verified either. An analysis of 56,381 papers and 2.2 million citations found an invalid-citation rate of 1.07% (up 80.9% year over year in 2025), and in a user survey (N=94), of the 86.7% who said they “always verify,” 41.5% had in fact copy-pasted BibTeX entries without verifying them. A qualitative study of expert users (N=21) casts doubt on the very verifiability that source-cited answers advertise. Learning grows shallower too. Across seven experiments totaling N≈10,462, an RCT found that learning with ChatGPT shortened time spent, lowered self-rated “I learned something new,” and produced shallower learning even after controlling for information content.

The same thing is happening outside scholarly information too.

Roughly a quarter of U.S. local newspapers have disappeared since 2004 (Abernathy 2020). In regions where newsroom staff have shrunk, the political competitiveness of mayoral elections falls (Rubado & Jennings 2020). A study that cross-referenced 15 years of content analysis across 200 newspapers with election outcomes showed that the contraction of local reporting is linked to lower name recognition of legislators and lower voter turnout (Hayes & Lawless 2021). In an age when the total volume of information has grown, the pathway for reaching local information has narrowed.

Something else moves into the narrowed space. What Bengani (2019) identified was a network of sites that pose as local newspapers while distributing algorithmically generated articles. What initially numbered over 450 was later counted at over 1,200. Making one is not hard either. Fine-tuning a model on just 40,000 Italian-language articles is enough to mass-produce “news-like” text at a level humans cannot distinguish from the real thing (Puccetti et al. 2024).

On the recipient’s side, people are starting to opt out. In a survey of over 93,000 people across 46 countries, the share of people who deliberately avoid the news rose from 29% in 2017 to 38% (Newman et al. 2025).

The assumption that a citation, once attached, makes something checkable is not holding up either.

In a study that manually audited generative-search output, only 51.5% of generated sentences were fully supported by their citations, and the citations themselves backed their corresponding sentence only 74.5% of the time (Liu, Zhang & Liang 2023). In real browsing data from 900 U.S. adults, the click-through rate on links fell from 15% to 8% on pages where an AI summary was shown, and clicks through to the cited source stayed at about 1% of visits (Chapekis et al. 2025). And a preregistered experiment with a sample representative of US adults (N=4,927) shows that presenting a citation or reference link significantly raises trust even when its content is wrong or hallucinated (Li & Aral 2025). A think-aloud study (N=22) likewise found that chat-format answers, by their appearance of completeness, produce misplaced trust, raising confidence even in wrong answers (Mayerhofer et al. 2025).

Displaying a source does not prompt verification. It hands out the feeling of having verified.

As Simon wrote in 1971, a wealth of information creates a poverty of attention. What is different is that what is growing poor now is not only attention but the diversity of pathways as well.

What Becomes the Object of Design

No peer-reviewed research directly answering RQ1 was found within the scope of this collection. The closest, C35 (a conceptualization of epistemic security), is a 2020 institutional report that does not presume the mass circulation of generated content. So what follows is assembled from adjacent fields.

Who bears the burden of verification, and how is it allocated: the skew in peer review (69–94% falling to the top 20%) and the side effect of strengthening the gatekeeper (retention 40.0% → 13.2%) take the same shape as a problem of allocation design.

Can judgment be moved away from the point of consumption: rather than teaching recipients how to tell hygiene apart, the grade card put the inspector’s judgment on the storefront. The bridging algorithm made the object of aggregation the quality of agreement, not a majority vote. Neither eliminated the labor of discernment. Both changed who bears it and where.

Can delay and load be treated as design variables: the effect of a note disappears with delay in its display, and claims that take more effort go unverified. Speed and labor are first-order design variables alongside accuracy.

Can a layperson choose whom to depend on: the epistemic dependence Hardwig defined holds that even someone without an expert’s direct evidence can rationally believe. Goldman offered selection criteria for that, but in the cognitive islands Nguyen described, no evaluation test exists for the layperson, and a mistaken choice of expert reinforces itself. Slop works to widen this island.

Can independence be secured before aggregation: making others’ estimates visible destroys the diversity that collective intelligence depends on.

Can we design the stopping point rather than the total volume: A meta-analysis of information overload (117 studies, N=133,011) confirms that overload correlates positively with information avoidance and stress (Graf & Antoni 2022). But a comprehensive review points out that objective information volume and subjective feelings of overload do not line up (Arnold et al. 2023). People stop searching not on exhaustive coverage but on a judgment of “enough,” and that stopping rule can be identified (Browne et al. 2007, N=115). What can be reduced is not the total volume but the grounds for judging that it is fine to stop.

Can the intermediary be switched by condition: Human editing and algorithmic recommendation are not uniformly better or worse. In a field experiment, the algorithm outperforms on clicks on average, but human editing wins out when personal data is scarce and on days with high news salience (Peukert et al. 2023). Which one to use becomes a design variable that switches with the condition.

How far can the labor of screening be shifted to a machine: Active-learning literature screening reaches 95% recall in simulation after checking just 8–33% of the corpus (van de Schoot et al. 2021). But a retrospective evaluation of a different tool left a 12.6% false-negative rate (Gates et al. 2018). The burden can be shifted, but what gets missed at the new location has to be measured.

Conditions Under Which This Reading Fails

First, if evidence emerges that discernment ability rises structurally. A 30-minute training given to 32 intelligence analysts improved overall accuracy from 72% by 9 points (95% CI [2.7, 15.4]). This is a small sample of professionally trained people, and whether it replicates in a general population is untested, but if it does replicate, the premise of “not betting on the ability to discern” weakens.

Second, if the premise of misinformation research itself collapses. Budak and colleagues argue that average exposure to problematic content is limited and concentrated among a small segment, and that the algorithm’s contribution is smaller than the conventional view holds. A meta-analysis on the effect of corrections remains unsettled, with rebuttal and reply continuing even after peer review. The backfire effect was not found even once, in a test covering 52 issues and over 10,100 participants. Both the danger and the effectiveness of intervention may be smaller than assumed.

Third, if measurement specific to slop itself becomes established. Most of the empirical work found in this collection targets misinformation, not unverified generated content in general. If the two are different things, the reach of the numbers cited here narrows.

Outline for Turning This Into a Review Paper

A design for restructuring this note into a submission-ready paper. The body text is not yet written.

Title Candidates

1. Designing for Slop: Why Detection Is the Wrong Bet and What to Design Instead: an option that foregrounds the claim.
2. From Discernment to Distribution: A Review of Information Environment Design Under Unverified Generative Content: an option that announces the axis of verification-burden allocation.
3. Who Verifies? A Review of Epistemic Infrastructure in the Age of AI Slop: an option that poses it as a question.

RQ

• RQ1: Taking as given that the generation of AI slop cannot be stopped, what does society (institutions, infrastructure, platforms, education) target for design?
• RQ2: How do recipients discern thin information? Does that ability rise with training, how large is the effect, and where does it fail?
• RQ3: How should the pathways for reaching important information be designed and maintained?

Contribution Claims

Slop research (definition, detection, countermeasures), intervention research on information environments (literacy, collective verification, nudges), and the social history of trust (the changing regimes of objectivity) have so far accumulated as separate bodies of literature. This review realigns the three along a single axis: who the burden of verification is allocated to.

The order follows the distribution of novelty. Points 1 and 2 constitute this review’s own contribution, while points 4 and 5 reposition what prior reviews have already shown individually.

1. Pathways of reach are narrowing on both the scholarly-information and news sides (RQ3). And displaying a source does not prompt verification; it hands out the feeling of having verified (a citation raises trust even when its content is wrong, and clicks through to the cited source stay at about 1% of visits). No systematic review integrating the effect of generative search on pathways of reach was found within the scope of this collection.
2. The regime of trust production has been independently invoked for AI four times (Christin 2016, Rieder & Simon 2016, Anichini & Kotras 2024, Schenk 2025). This is empirical evidence that the framework of mechanical objectivity and trained judgment is actively in use. No review integrates these; only the accumulation of individual cases has progressed.
3. This translates the allocation of the burden of verification into measurable design variables: delay, cognitive load, and allocation rate. But the axis itself is not a new discovery. Rittel and Webber’s no stopping rule (property 2) and the absence of a right to be wrong (property 10), Simon’s bounded rationality, and Costanza-Chock’s separation of benefit and burden each occupy the content of this axis in their own vocabulary. The novelty lies only in the three variables, which appear in measurable form in none of them.
4. The designs that worked shifted the labor of discernment from the point of consumption to the point of production or mediation, and made public who operates at the destination of that shift (RQ1). Saying “discernment is not asked of the individual” is wrong. It is an individual who writes a Community Notes note, and an individual (an inspector) who conducts a hygiene inspection. What disappeared was not the labor of discernment but the obligation to bear it at the point of consumption. This shift also appears across the history of trust production since 17th-century virtual witnessing, but the history does not run in one direction. What Daston and Galison showed was that mechanical objectivity could not eliminate judgment entirely, and trained judgment returned as a third regime.
5. A design that bets on the recipient’s discernment ability is unsupported on three counts: effect size, decay, and side effects (RQ2). But discernment is of two kinds, and the reach of the evidence differs between them. Identifying provenance (whether something is generated) has a raw accuracy of 55.54%, straddling chance level (Diel et al. 2024). Judging the truth or falsity of content has a training effect that vanishes within 48 hours to 10 days without an immediate posttest, and what mainly rises is the ability to doubt (skepticism bias d=0.32, Pfänder & Altay 2025). Under the sister note’s definition, content counts as slop even when accurate if it is imposed without verification, so it is the former that reaches slop directly, while the latter reaches it only partially. This diagnosis itself, moreover, has already been shown individually by prior reviews (Kozyreva et al. 2024’s catalog, Bak-Coleman et al. 2022’s quantification, Tay et al. 2023’s identification of side effects). This review’s contribution is limited to separating this into provenance identification and veracity judgment, and making explicit the reach each has to the definition of slop.
6. No peer-reviewed research directly answers RQ1. This review pinpoints the location of that gap and presents it as a research agenda.

Chapter Outline

• 1 Introduction: Opens from the fact that countermeasures betting on discernment have repeatedly failed, and reframes the question from “can it be told apart” to “who verifies it, and how is that burden allocated.” Places here a table contrasting prior concepts. The rows: Rittel & Webber (1973)‘s no stopping rule and the absence of a right to be wrong, Simon (1969/1996)‘s bounded rationality and satisficing, Simon (1971)‘s scarcity of attention, Costanza-Chock (2020)‘s separation of benefit and burden, and the fourth of Buchanan (1992)‘s four orders. The columns: “what the concept states” and “what this review adds (the measurable variables of delay, cognitive load, and allocation rate).” Without this table, the axis of integration reads as a paraphrase of existing concepts.
• 2 Background: Sets three side by side. (a) The definition of slop and the outsourcing of verification (the corpus of the sister note AI Slop: Reading It as Outsourced Verification, Not Low Quality), (b) the history of the regimes of trust production (virtual witnessing, the honor of a gentleman, quantification, mechanical objectivity, trained judgment, the audit society) and their redeployment in the age of AI, (c) the gap from prior design principles. In (c), two points are written, one paragraph each. That Norman (1988/2013)‘s knowledge in the world and forcing function are self-contained within a single artifact, whereas this review’s object is a transfer that crosses multiple agents and carries failure modes specific to transfer (the effect vanishes with delay in display; claims with higher cognitive load go unverified). And that, in light of Krippendorff (2006)‘s second-order understanding, what counts as “having been verified” at the destination of the transfer depends on shared meaning, and the meaning of a rating symbol can differ across groups.
• 3 Method: The lightweight scoping procedure, search terms, inclusion and exclusion criteria, charting items (intervention type, effect size, sample, decay, side effects, where the burden is allocated). Aiming for PRISMA-ScR compliance would require additional collection (below). Makes the relationship to prior reviews explicit in Method. This review treats existing systematic reviews (Kozyreva et al. 2024, Pfänder & Altay 2025, Bak-Coleman et al. 2022, and others) not as primary evidence but as objects of integration. Without stating this treatment, the review reads as trailing behind existing reviews.
• 4 Results / RQ3: Pathways of reach. The stiffening of the canon, the low source overlap of generative search, news deserts, the hollow spin of source display. Placed first because its novelty is highest.
• 5 Results / RQ1: The transfer of the labor of discernment. Mandatory disclosure, the limits of signaling and certification, the bridging algorithm, the side effects of strengthening the gatekeeper, the skew in peer-review labor. Makes explicit, for each case, who operates at the destination of the transfer and who bears the cost of a mistaken judgment.
• 6 Results / RQ2: Where the recipient’s discernment ability currently stands. Splits into two tables: provenance identification and content veracity judgment. The former covers Diel et al. (2024), Roca et al. (2025), Fiedler & Döpke (2025), Milička et al. (2025), Kamali et al. (2026). The latter covers Pfänder & Altay (2025), Huang et al. (2024), Lyons et al. (2024), Martini et al. (2025), and the accuracy-nudge line of work. Immediately after each table, writes in one sentence which of the two carries reach against the definition of slop.
• 7 Discussion: Presents three asymmetries. (a) Provenance cannot be discerned, but correction works (meta-analysis d=0.29; the backfire effect did not replicate in a large-scale test), (b) the target is not the volume of information but the allocation of verification (subjective feelings of overload and objective volume do not line up), (c) apparatus can become ritualized (the audit-society critique has the same shape as the Community Notes finding, where the effect disappears at the level of the whole system).
• 8 Research agenda: Study designs to fill the gap in RQ1. Measurement scales for burden allocation, experiments that set delay and load as independent variables, non-English-language contexts.
• 9 Limitations: Overrepresentation of the English-speaking world, the proportion of preprints, the absence of measurement criteria specific to slop, extrapolation from misinformation research, and survivorship bias among the success cases of disclosure institutions (the population of disclosure institutions that did not work has not been presented).

Novelty Verification (Conducted 2026-08-28)

Both a search of prior reviews and a critique from design theory were applied independently. The result is conditionally novel.

Threats to Novelty (Prior Reviews)

• Kozyreva et al. (2024)‘s Toolbox (Nature Human Behaviour, 81 papers, nine kinds of individual-level intervention) almost entirely preempts the scope of contribution 5. Differentiation is limited to the difference between a catalog and a negative summary.
• The boosting lineage of Kozyreva et al. (2020) / Herzog & Hertwig (2025) preempts the “design the environment side” skeleton of contribution 4. Weaken the connection to the historical lineage, and this looks like an extension of that lineage.
• Bak-Coleman et al. (2022) / Tay et al. (2023) have already pointed out, quantitatively and methodologically, the insufficient effect of individual interventions and the oversight of side effects.

Threats to Novelty (Design Theory)

• The axis of integration sits at the intersection of Rittel & Webber’s property 2 and property 10, Simon’s bounded rationality, and Costanza-Chock’s separation of benefit and burden. This note cites none of the three.
• Contribution 4 can be read as a redescription of Norman’s knowledge in the world. If the difference cannot be written, withdraw it.
• This review observes neither the act of design nor designers. Because this does not match the object domain Cross (1982) defined for Design Studies, that journal does not stand as a first choice.
• The two sister notes have already placed “slop is the allocation of the burden of verification” (the definition) and “the countermeasures that worked changed the structure rather than trying to discern” (the conclusion). This needs to answer the suspicion of circularity, namely that it is merely reconfirming a consequence already built into the definition. Making explicit the observational conditions under which contribution 4 could be falsified is the minimal response.

Gaps That Still Hold

• A general theory of the allocation of the burden of verification. Peer-review cost theory is domain-specific, delegation theory predates peer review, and no prior review was found that stands as a society-wide allocation theory. The clearest gap of the seven items.
• A review integrating the application of mechanical objectivity to AI. Four individual cases exist, but no integration. Accumulation is progressing, so the risk of someone else writing it first is moderate.
• A systematic review of the effect of generative search on pathways of reach.
• This subject within design research journals. Zero hits across Design Studies / Design Issues / She Ji.

Candidate Venues (Restricted to Diamond OA)

The conditions are that the author pays no charge, the reader can read it for free, and it is peer-reviewed. The standing of the journal is not a criterion.

Under these conditions every original candidate falls away. Design Issues costs the author nothing if published under subscription, but then readers cannot access it freely; choosing OA costs $1,800. The OA option for Design Studies is $4,690, Big Data & Society is $2,250, Design Science is £2,610, and the International Journal of Design is $1,000. The Bulletin of JSSD (Japanese Society for the Science of Design) has a “Review Articles” submission category and a close subject fit, but charges a flat ¥40,000 publication fee regardless of page count, which falls outside the author-pays-nothing condition.

• First choice: Engaging Science, Technology, and Society (ESTS, published by the Society for Social Studies of Science). The subject fit is the highest. The flood of AI-generated content, the allocation of verification labor, and knowledge infrastructure are central concerns of STS. From Volume 12 in 2026 the submission categories are unified into one, and a non-standard synthesis can be submitted if the cover letter explains its purpose. Limit 9,000 words (including notes, excluding references), double-anonymous review. The evidence for no APC comes only from search-engine-cached pages, so confirm with the editors before submitting.
• Second choice: a Perspective Article (展望論文) in Cognitive Studies, the journal of the Japanese Cognitive Science Society. If the first author is a member, neither review fees nor publication charges apply, and the article is published on J-STAGE. The Perspective Article category corresponds to a review article, making the fit of submission category the most certain on this list. It can be written in Japanese. However, RQ1 (the design of social institutions) falls outside the reach of cognitive science, so the paper would need to be rebuilt around RQ2 (the limits of judgment) and RQ3 (search and reach), with RQ1 demoted to the discussion. Non-members pay a ¥10,000 review fee and ¥15,000 per page.
• Third choice: Journal of Human-Technology Relations (JHTR, TU Delft Open). This is the only journal for which the absence of both submission fees and APC was confirmed directly on the official page. It is an interdisciplinary journal spanning design research, HCI, and STS, and the subject fit is high. However, its “reviews” category covers book and exhibition reviews rather than literature reviews, so whether a theoretical synthesis can go into the Research article category (limit 10,000 words) must be asked in advance.
• Fourth choice: Artifact: Journal of Design Practice (Intellect, funded by the Royal Danish Academy) and FORMakademisk (Oslo Metropolitan University). Both are diamond OA design journals, but the former leans toward practice research and the latter toward design education, leaving a distance from this subject.

The excluded candidates are worth recording. Information Research is the most certain diamond OA venue but was not accepting submissions as of August 2026. First Monday ceased publication with its May 2026 issue. tripleC peer-reviews its Papers category, but the Reflections category that includes reviews is explicitly not peer-reviewed, which does not serve the purpose of building a publication record.

Scope of Additional Collection Needed

1. Primary literature in design theory (judged the top priority in the novelty verification). The original text of Rittel & Webber (1973)‘s property 2 and property 10, Simon (1969/1996) chapter 5, Norman (1988/2013) chapters 3 and 4, Krippendorff (1989, 2006), Buchanan (1992)‘s fourth order, Cross (1982), Costanza-Chock (2020)‘s separation of benefit and burden, Schön (1983) chapter 2. If reading these does not yield a difference that can be written, withdraw contributions 3 and 4.
2. A systematic search of Design Studies and Design Issues. Confirms, with the search scope stated explicitly, that no prior paper exists that sets “the allocation of the burden of verification” as its object of design. The current “zero hits” is the result of a cross-search, not a systematic search.
3. The gap in RQ1: epistemic security, information ecosystem design, primary sources on the public-policy side (the closest existing literature at present is a 2020 institutional report that does not presume the mass circulation of generated content).
4. The population of disclosure institutions that did not work: The hygiene grade card and Community Notes are presented as success cases, but the success rate across all cases of mandatory disclosure and collective verification is not shown. Needed to answer the survivorship-bias concern.
5. Non-English-speaking contexts: empirical research in Japanese and Chinese. The current corpus is heavily skewed toward the English-speaking world.
6. Measurement specific to slop: much of the current evidence targets misinformation, not unverified generated content in general. If the two are different things, the reach of this review’s figures narrows.
7. A cross-search of retractions: cross-checking against the Retraction Watch database has not been performed.

Unverified Items

Among the figures used in the body text, the following have not yet had primary-source verification completed.

• Slaughter et al. (2025) has an official correction dated January 30, 2026 (10.1073/pnas.2600635123). The correction adds a competing-interest statement and does not change any figures (confirmed in the full text of the correction on Europe PMC, 2026-11-06). The figures in the body (reposts −46.1% and others) were confirmed in the abstract.
• Diel et al. (2024)‘s 55.54%, its confidence interval (48.87 to 62.10), the 56 papers, and N=86,155 were confirmed in the abstract (2026-11-06). The full text has not been read.
• The exact sample size of Roozenbeek et al. (2022)‘s YouTube field experiment.
• The exact sample sizes of Guess et al. (2020)‘s U.S. and India surveys.
• The author listing for Wack et al. (2026) and Cheng et al. (2026).
• A cross-search for retractions and expressions of concern was not performed (zero hits within the scope of Crossref’s update field).

The 21 unverified items for the full corpus are listed at the end of source/review/slop-era-information-design/papers.md.

For the additional collection (groups G–K, 117 items), the following are unconfirmed.

• Li & Aral (2025) and Chapekis et al. (2025) are both preprints whose peer-reviewed publication status is unconfirmed. The figures cited in the body (N=4,927; reference links raise trust even when wrong; click rates falling from 15% to 8%; clicks to cited sources in about 1% of visits) were confirmed in the arXiv full texts (2026-11-06). The latter appears to share data with a Pew Research Center short read, but the identity of the authors and the data is unconfirmed.
• Which edition of Abernathy’s continuing news desert series the “roughly a quarter” figure in the body comes from.
• The DOI notation for Kuhlthau (1991) and Metzger & Flanagin (2013).

The unverified items for the corpus as a whole are at the end of source/review/slop-era-information-design/papers.md (22 items) and at the end of each of groups G–K.

The following remain unconfirmed from the novelty verification (2026-08-28).

• Whether a prior paper in Design Studies or Design Issues has posited the allocation of the burden of verification as an object of design. The current “zero hits” is the result of a cross-search, not a systematic search.

• The specific locations in the primary design-theory literature. The pages for properties 2 and 10 in Rittel & Webber (1973), the chapter numbers across editions of Simon (1969/1996), the chapter numbers across editions of Norman (1988/2013), and the relevant pages in Krippendorff (2006) and Costanza-Chock (2020) are all cited without confirmation.

• The full texts of Kozyreva et al. (2024), Herzog & Hertwig (2025), and Ecker et al. (2022) were behind authentication walls; the judgment rests on bibliographic records and abstracts only.

• The APCs of the candidate venues. The official pages of ESTS and JCMC were blocked by bot protection, so the evidence that they charge nothing rests on search-engine-cached pages. Whether JHTR, Artifact, and FORMakademisk accept literature reviews is also unconfirmed. Ask the editors before submitting.

Related Notes

• AI Slop: Not Low Quality but the Outsourcing of Verification: The sister note that defined slop as the allocation of the burden of verification. This note uses that definition from the side of institutional design.
• The History of Slop and Countermeasures: Why the Same Wager Keeps Failing: The five types of countermeasure, and how each breaks down. This note’s finding that “the designs that worked moved the labor of discernment away from the point of consumption” is the present-day form of that historical conclusion.
• Slop Seen From Design History: The Profession Began as a Countermeasure to Shoddy Mass Production: The history of how selection by taste judgment was also an apparatus of class and gender. The prehistory of the problem that gatekeeping errors concentrate on the weaker side.
• How Cognitive Surrender Becomes AI Slop: Why No One Closes the Gap in Completeness: A discussion that separates the pathways by which verification is skipped on the maker’s side (not evaluating, the standard drifting toward form, not seeing the difference, the standard thinning). It is the source of the burden that this note treats from the side of recipients and institutions.
• Recent Currents in LLM-as-a-Judge: A Literature Map of 53 Core Studies and a Standalone Chapter on Creativity Evaluation: The question of whether verification can be handed back to a machine. Treats a non-human option as a destination for the allocation of the burden of verification.
• From Maker to Editor: A Structural Analysis of the Designer Role Transition in the Age of AI: The shift in role from maker to editor. A description of the change in burden allocation from the side of professional practice.
• Does Early Use of Generative AI Inhibit the Formation of Thought? A Literature Map of Cognitive Offloading and Learning: Cognitive offloading and learning. Sits in the same stratum as the finding that generative search makes learning shallower (Melumad & Yun 2025).
• 90% Adoption x 10% Approval — The Paradox of AI Tool Diffusion and Evaluative Divergence: The divergence between adoption rate and approval rate. Connects to the finding that labels change attitudes but not behavior.
• The Vanishing of Knowledge: Undiscovery as the Dual of Discovery: The disappearance of knowledge that goes unreached. Where the stiffening of the canon and the low source overlap both lead.
• What Do Data Visualization and Data Art Share, and Where Do They Differ?: A note that organizes a procedure for separating information displays from artworks by what a figure asks of its reader. It maps this note’s question of where to place the burden of verification onto a three-layer question seen from the side of the displayed object.

References

Abernathy, P. M. (2020). The Expanding News Desert. UNC Hussman School of Journalism and Media. https://www.usnewsdeserts.com/

Adams, Z., Osman, M., Bechlivanidis, C., & Meder, B. (2023). (Why) Is Misinformation a Problem? Perspectives on Psychological Science. https://doi.org/10.1177/17456916221141344

Akerlof, G. A. (1970). The Market for “Lemons”: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431

Allaham, M., & Diakopoulos, N. (2026). Synthetic Sources? Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources. arXiv:2605.23684. https://arxiv.org/abs/2605.23684

Allen, J., Arechar, A. A., Pennycook, G., & Rand, D. G. (2021). Scaling up fact-checking using the wisdom of crowds. Science Advances, 7(36). https://doi.org/10.1126/sciadv.abf4393

Altay, S. (2026). Rethinking the problem of misinformation and its solutions. New Media & Society. https://doi.org/10.1177/14614448261428635

Anichini, G., & Kotras, B. (2024). Trained Judgements: Artificial Intelligence, Epistemic Tensions and the Production of Scientific Objectivity. Science, Technology, & Human Values. https://doi.org/10.1177/01622439241262854

Arnold, M., Goldschmitt, M., & Rigotti, T. (2023). Dealing with information overload: a comprehensive review. Frontiers in Psychology, 14, 1122200. https://doi.org/10.3389/fpsyg.2023.1122200

Aslett, K., Guess, A. M., Bonneau, R., Nagler, J., & Tucker, J. A. (2022). News credibility labels have limited average effects on news diet quality and fail to reduce misperceptions. Science Advances. https://doi.org/10.1126/sciadv.abl3844

Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6). https://doi.org/10.1016/0005-1098(83)90046-8

Bak-Coleman, J. B., Kennedy, I., Wack, M., Beers, A., Schafer, J. S., Spiro, E. S., Starbird, K., & West, J. D. (2022). Combining interventions to reduce the spread of viral misinformation. Nature Human Behaviour, 6, 1372–1380. https://doi.org/10.1038/s41562-022-01388-6

Bengani, P. (2019). “Pink Slime”: Partisan Journalism and the Future of Local News. Columbia Journalism Review / Tow Center for Digital Journalism. https://www.cjr.org/tow_center/pink-slime-partisan-journalism-and-the-future-of-local-news.php

Bolton, P., Freixas, X., & Shapiro, J. (2012). The Credit Ratings Game. Journal of Finance, 67(1). https://doi.org/10.1111/j.1540-6261.2011.01708.x

Browne, G. J., Pitts, M. G., & Wetherbe, J. C. (2007). Cognitive Stopping Rules for Terminating Information Search in Online Tasks. MIS Quarterly, 31(1), 89–104. https://aisel.aisnet.org/misq/vol31/iss1/7/

Buchanan, R. (1992). Wicked Problems in Design Thinking. Design Issues, 8(2), 5–21. https://doi.org/10.2307/1511637

Budak, C., Nyhan, B., Rothschild, D. M., Thorson, E., & Watts, D. J. (2024). Misunderstanding the harms of online misinformation. Nature, 630(8015). https://doi.org/10.1038/s41586-024-07417-w

Butler, L. H., DeGutis, J., Tay, L. Q., Ecker, U. K. H., & Swire-Thompson, B. (2025). Corrections are effective for science misinformation. Nature Human Behaviour, 9(12), 2458–2460. https://doi.org/10.1038/s41562-025-02245-y

Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. PACM HCI, 5(CSCW1), Art.188. https://doi.org/10.1145/3449287

Capewell, G., Maertens, R., Remshard, M., van der Linden, S., Compton, J., Lewandowsky, S., & Roozenbeek, J. (2024). Misinformation interventions decay rapidly without an immediate posttest. Journal of Applied Social Psychology. https://doi.org/10.1111/jasp.13049

Chan, M. S., & Albarracín, D. (2023). A meta-analysis of correction effects in science-relevant misinformation. Nature Human Behaviour. https://doi.org/10.1038/s41562-023-01623-8

Chan, M. S., & Albarracín, D. (2025). Reply to: Corrections are effective for science misinformation. Nature Human Behaviour, 9(12), 2461–2470. https://doi.org/10.1038/s41562-025-02265-8

Chapekis, A., Lieb, A., Shah, S., & Smith, A. (2025). Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview. arXiv:2608.04831. https://arxiv.org/abs/2608.04831

Christin, A. (2016). From Daguerreotypes to Algorithms: Machines, Expertise, and Three Forms of Objectivity. ACM SIGCAS Computers and Society, 46(1). https://doi.org/10.1145/2908216.2908220

Chu, J. S. G., & Evans, J. A. (2021). Slowed Canonical Progress in Large Fields of Science. PNAS, 118(41), e2021636118. https://doi.org/10.1073/pnas.2021636118

Chuai, Y., Pilarski, M., Renault, T., Restrepo-Amariles, D., Troussel-Clément, A., Lenzini, G., & Pröllochs, N. (2026). Community-based fact-checking reduces the spread of misleading posts on X (formerly Twitter). Nature Communications, 17, 4070. https://doi.org/10.1038/s41467-026-72597-0

Chuai, Y., Tian, H., Pröllochs, N., & Lenzini, G. (2024). Did the Roll-Out of Community Notes Reduce Engagement With Misinformation on X/Twitter? PACM HCI, 8(CSCW2). https://doi.org/10.1145/3686967

Costanza-Chock, S. (2020). Design Justice: Community-Led Practices to Build the Worlds We Need. MIT Press.

Cross, N. (1982). Designerly Ways of Knowing. Design Studies, 3(4), 221–227. https://doi.org/10.1016/0142-694X(82)90040-0

Daston, L., & Galison, P. (2007). Objectivity. Zone Books.

Diel, A., Lalgi, T., Schröter, I. C., MacDorman, K. F., Teufel, M., & Bäuerle, A. (2024). Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers. Computers in Human Behavior Reports, 16, 100538. https://doi.org/10.1016/j.chbr.2024.100538

Dorst, K., & Cross, N. (2001). Creativity in the design process: co-evolution of problem–solution. Design Studies, 22(5), 425–437. https://doi.org/10.1016/S0142-694X(01)00009-6

Ecker, U. K. H., Lewandowsky, S., Cook, J., Schmid, P., Fazio, L. K., Brashier, N., Kendeou, P., Vraga, E. K., & Amazeen, M. A. (2022). The psychological drivers of misinformation belief and its resistance to correction. Nature Reviews Psychology, 1, 13–29. https://doi.org/10.1038/s44159-021-00006-y

Escobar, A. (2018). Designs for the Pluriverse: Radical Interdependence, Autonomy, and the Making of Worlds. Duke University Press.

Evans, J. A. (2008). Electronic Publication and the Narrowing of Science and Scholarship. Science, 321(5887), 395–399. https://doi.org/10.1126/science.1150473

Fan, G., Liu, D., Pan, L., & Zhang, R. (2026). When Help Hurts: Verification Load and Fatigue with AI Coding Assistants. CHI 2026. https://doi.org/10.1145/3772318.3791176

Fiedler, A., & Döpke, J. (2025). Do humans identify AI-generated text better than machines? International Review of Economics Education, 49, 100321. https://doi.org/10.1016/j.iree.2025.100321

Filippas, A., Horton, J. J., & Golden, J. (2018). Reputation Inflation. EC ‘18. https://doi.org/10.1145/3219166.3219222

Gates, A., Johnson, C., & Hartling, L. (2018). Technology-assisted title and abstract screening for systematic reviews: A retrospective evaluation of the Abstrackr machine learning tool. Systematic Reviews, 7(1), 45. https://doi.org/10.1186/s13643-018-0707-8

Goldman, A. I. (2001). Experts: Which Ones Should You Trust? Philosophy and Phenomenological Research, 63(1). https://doi.org/10.1111/j.1933-1592.2001.tb00093.x

Graf, B., & Antoni, C. H. (2022). Drowning in the flood of information: a meta-analysis on the relation between information overload, behaviour, experience, and health and moderating factors. European Journal of Work and Organizational Psychology, 32(2). https://doi.org/10.1080/1359432X.2022.2118051

Grier, R. A., Warm, J. S., Dember, W. N., Matthews, G., Galinsky, T. L., Szalma, J. L., & Parasuraman, R. (2003). The Vigilance Decrement Reflects Limitations in Effortful Attention, Not Mindlessness. Human Factors, 45(3). https://doi.org/10.1518/hfes.45.3.349.27253

Grossman, R., Liu, S., Chen, M. K., Smith, M., Borcea, C., & Chen, Y. (2026). How Generative AI Disrupts Search. arXiv:2604.27790. https://arxiv.org/abs/2604.27790

Guess, A. M., Lerner, M., Lyons, B., Montgomery, J. M., Nyhan, B., Reifler, J., & Sircar, N. (2020). A digital media literacy intervention increases discernment between mainstream and false news in the United States and India. PNAS, 117(27), 15536–15545. https://doi.org/10.1073/pnas.1920498117

Halfaker, A., Geiger, R. S., Morgan, J. T., & Riedl, J. (2013). The Rise and Decline of an Open Collaboration System. American Behavioral Scientist, 57(5). https://doi.org/10.1177/0002764212469365

Hardwig, J. (1985). Epistemic Dependence. The Journal of Philosophy, 82(7). https://www.jstor.org/stable/2026523

Hayes, D., & Lawless, J. L. (2021). News Hole: The Demise of Local Journalism and Political Engagement. Cambridge University Press.

Herzog, S. M., & Hertwig, R. (2025). Boosting: Empowering Citizens with Behavioral Science. Annual Review of Psychology, 76, 851–881. https://doi.org/10.1146/annurev-psych-020924-124753

Huang, G., Jia, W., & Yu, W. (2024). Media Literacy Interventions Improve Resilience to Misinformation. Communication Research. https://doi.org/10.1177/00936502241288103

Jin, G. Z., & Leslie, P. (2003). The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards. Quarterly Journal of Economics, 118(2). https://doi.org/10.1162/003355303321675428

Kamali, N., Gerstner, C. R., Hullman, J., & Groh, M. (2026). Generative AI Literacy Training Improves Intelligence Analysts’ Discrimination of Real and AI-Generated Images. arXiv:2606.28510. https://arxiv.org/abs/2606.28510

Kittur, A., & Kraut, R. E. (2008). Harnessing the wisdom of crowds in Wikipedia: quality through coordination. CSCW ‘08. https://doi.org/10.1145/1460563.1460572

Kovanis, M., Porcher, R., Ravaud, P., & Trinquart, L. (2016). The Global Burden of Journal Peer Review in the Biomedical Literature. PLOS ONE, 11(11). https://doi.org/10.1371/journal.pone.0166387

Kozyreva, A., Lewandowsky, S., & Hertwig, R. (2020). Citizens Versus the Internet: Confronting Digital Challenges With Cognitive Tools. Psychological Science in the Public Interest, 21(3), 103–156. https://doi.org/10.1177/1529100620946707

Kozyreva, A., Lorenz-Spreen, P., Herzog, S. M., Ecker, U. K. H., Lewandowsky, S., Hertwig, R., et al. (2024). Toolbox of individual-level interventions against online misinformation. Nature Human Behaviour, 8, 1044–1052. https://doi.org/10.1038/s41562-024-01881-0

Krippendorff, K. (1989). On the Essential Contexts of Artifacts or on the Proposition that Design Is Making Sense (of Things). Design Issues, 5(2), 9–39. https://doi.org/10.2307/1511512

Krippendorff, K. (2006). The Semantic Turn: A New Foundation for Design. CRC Press.

Li, H., & Aral, S. (2025). Human Trust in AI Search: A Large-Scale Experiment. arXiv:2504.06435. https://arxiv.org/abs/2504.06435

Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of the Association for Computational Linguistics: EMNLP 2023. https://aclanthology.org/2023.findings-emnlp.467/

Lizzeri, A. (1999). Information Revelation and Certification Intermediaries. RAND Journal of Economics, 30(2). https://doi.org/10.2307/2556078

Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. PNAS, 108(22), 9020–9025. https://doi.org/10.1073/pnas.1008636108

Lyons, B., King, A. J., & Kaphingst, K. A. (2024). A Health Media Literacy Intervention Increases Skepticism of Both Inaccurate and Accurate Cancer News Among U.S. Adults. Annals of Behavioral Medicine, 58(12). https://doi.org/10.1093/abm/kaae054

Maertens, R., Roozenbeek, J., Simons, J. S., Lewandowsky, S., Maturo, V., Goldberg, B., et al. (2025). Psychological booster shots targeting memory increase long-term resistance against misinformation. Nature Communications. https://doi.org/10.1038/s41467-025-57205-x

Martel, C., Rathje, S., Clark, C. J., Pennycook, G., Van Bavel, J. J., Rand, D. G., & van der Linden, S. (2024). On the Efficacy of Accuracy Prompts Across Partisan Lines: An Adversarial Collaboration. Psychological Science. https://doi.org/10.1177/09567976241232905

Martini, C., Floris, M., Ronzani, P., Ausili, L., Pennacchioni, G., Adorno, G., & Panizza, F. (2025). The impact of interventions against science disinformation in high school students. Scientific Reports. https://doi.org/10.1038/s41598-025-16565-6

Mayerhofer, K., Capra, R., & Elsweiler, D. (2025). Blending Queries and Conversations: Understanding Tactics, Trust, Verification, and System Choice in Web Search and Chat Interactions. CHIIR ‘25. https://doi.org/10.1145/3698204.3716454

Melumad, S., & Yun, J. H. (2025). Experimental Evidence of the Effects of Large Language Models Versus Web Search on Depth of Learning. PNAS Nexus, 4(10), pgaf316. https://doi.org/10.1093/pnasnexus/pgaf316

Milička, J., Marklová, A., Drobil, O., & Pospíšilová, E. (2025). Learning to detect AI texts and learning the limits. PLOS ONE, 20(10), e0333007. https://doi.org/10.1371/journal.pone.0333007

Newman, N., Fletcher, R., Robertson, C. T., Arguedas, A. R., & Nielsen, R. K. (2025). Digital News Report 2025. Reuters Institute for the Study of Journalism, University of Oxford. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2025

Nguyen, C. T. (2020). Cognitive Islands and Runaway Echo Chambers: Problems for Epistemic Dependence on Experts. Synthese, 197(7). https://doi.org/10.1007/s11229-018-1692-0

Norman, D. A. (2013). The Design of Everyday Things: Revised and Expanded Edition. Basic Books.

Nygren, T., Efimova, E., & Gesser-Edelsburg, A. (2025). Investigating the long-term impact of misinformation interventions in upper secondary education. PLOS ONE. https://doi.org/10.1371/journal.pone.0326928

Pennycook, G., & Rand, D. G. (2022). Accuracy prompts are a replicable and generalizable approach for reducing the spread of misinformation. Nature Communications. https://doi.org/10.1038/s41467-022-30073-5

Pennycook, G., Bear, A., Collins, E. T., & Rand, D. G. (2020). The Implied Truth Effect. Management Science. https://doi.org/10.1287/mnsc.2019.3478

Peukert, C., Sen, A., & Claussen, J. (2023). The Editor and the Algorithm: Recommendation Technology in Online News. Management Science, 70(9). https://doi.org/10.1287/mnsc.2023.4954

Pfänder, J., & Altay, S. (2025). Spotting false news and doubting true news: a systematic review and meta-analysis of news judgements. Nature Human Behaviour. https://doi.org/10.1038/s41562-024-02086-1

Porter, T. M. (1995). Trust in Numbers: The Pursuit of Objectivity in Science and Public Life. Princeton University Press.

Power, M. (1997). The Audit Society: Rituals of Verification. Oxford University Press.

Puccetti, G., Rogers, A., Alzetta, C., Dell’Orletta, F., & Esuli, A. (2024). AI “News” Content Farms Are Easy to Make and Hard to Detect: A Case Study in Italian. Proceedings of ACL 2024. https://aclanthology.org/2024.acl-long.817/

Rieder, G., & Simon, J. (2016). Datatrust: Or, the Political Quest for Numerical Evidence and the Epistemologies of Big Data. Big Data & Society, 3(1). https://doi.org/10.1177/2053951716649398

Rittel, H. W. J., & Webber, M. M. (1973). Dilemmas in a General Theory of Planning. Policy Sciences, 4(2), 155–169. https://doi.org/10.1007/BF01405730

Roca, T., Cintron Roman, A., Torres Vega, J., Duarte, M., Wang, P., White, K., Misra, A., & Lavista Ferres, J. (2025). How good are humans at detecting AI-generated images? arXiv:2507.18640. https://arxiv.org/abs/2507.18640

Roozenbeek, J., Freeman, A. L. J., & van der Linden, S. (2021). How Accurate Are Accuracy-Nudge Interventions? A Preregistered Direct Replication. Psychological Science. https://doi.org/10.1177/09567976211024535

Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., & Lewandowsky, S. (2022). Psychological inoculation improves resilience against misinformation on social media. Science Advances, 8(34), eabo6254. https://doi.org/10.1126/sciadv.abo6254

Rubado, M. E., & Jennings, J. T. (2020). Political Consequences of the Endangered Local Watchdog: Newspaper Decline and Mayoral Elections in the United States. Urban Affairs Review, 56(5). https://doi.org/10.1177/1078087419838058

Schenk, P. (2025). Heartless Rulez! Mechanical Objectivity, Empathic Understanding, and the Permissibility of AI. Frontiers in Sociology, 10. https://doi.org/10.3389/fsoc.2025.1604709

Schön, D. A. (1983). The Reflective Practitioner: How Professionals Think in Action. Basic Books.

Seger, E., Avin, S., Pearson, G., Briers, M., Ó hÉigeartaigh, S., & Bacon, H., et al. (2020). Tackling Threats to Informed Decision-Making in Democratic Societies: Promoting Epistemic Security in a Technologically-Advanced World. The Alan Turing Institute. https://www.turing.ac.uk/sites/default/files/2020-10/epistemic-security-report_final.pdf

Shapin, S. (1994). A Social History of Truth: Civility and Science in Seventeenth-Century England. University of Chicago Press. https://doi.org/10.7208/9780226148847

Shapin, S., & Schaffer, S. (1985). Leviathan and the Air-Pump: Hobbes, Boyle, and the Experimental Life. Princeton University Press.

Simon, H. A. (1971). Designing Organizations for an Information-Rich World. In M. Greenberger (Ed.), Computers, Communications, and the Public Interest (pp. 37–52). Johns Hopkins Press.

Simon, H. A. (1996). The Sciences of the Artificial (3rd ed.). MIT Press.

Slaughter, I., Peytavin, A., Ugander, J., & Saveski, M. (2025). Community notes reduce engagement with and diffusion of false information online. PNAS, 122(38). https://doi.org/10.1073/pnas.2503413122

Spence, M. (1973). Job Market Signaling. Quarterly Journal of Economics, 87(3). https://doi.org/10.2307/1882010

Tay, L. Q., Lewandowsky, S., Hurlstone, M. J., Kurz, T., & Ecker, U. K. H. (2023). A focus shift in the evaluation of misinformation interventions. Harvard Kennedy School Misinformation Review. https://doi.org/10.37016/mr-2020-124

van de Schoot, R., de Bruin, J., Schram, R., Zahedi, P., de Boer, J., Weijdema, F., … Oberski, D. L. (2021). An open source machine learning framework for efficient and transparent systematic reviews. Nature Machine Intelligence, 3, 125–133. https://doi.org/10.1038/s42256-020-00287-7

Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R. (2023). Explanations Can Reduce Overreliance on AI Systems During Decision-Making. PACM HCI, 7(CSCW). https://doi.org/10.1145/3579605

Venkit, P. N., Laban, P., Zhou, Y., Mao, Y., & Wu, C.-S. (2025). Search Engines in the AI Era: A Qualitative Understanding to the False Promise of Factual and Verifiable Source-Cited Responses in LLM-based Search. FAccT 2025. https://doi.org/10.1145/3715275.3732089

Wack, M., Warren, A., & Alam, M. (2026). The Laziness of the Crowd. arXiv:2603.11120. https://arxiv.org/pdf/2603.11120

Wojcik, S., Hilgard, S., Judd, N., Mocanu, D., Ragain, S., Hunzaker, M. B. F., Coleman, K., & Baxter, J. (2022). Birdwatch: Crowd Wisdom and Bridging Algorithms Can Inform Understanding and Reduce the Spread of Misinformation. arXiv:2210.15723. https://arxiv.org/abs/2210.15723

Wood, T., & Porter, E. (2019). The Elusive Backfire Effect. Political Behavior. https://doi.org/10.1007/s11109-018-9443-y

Xu, Z., et al. (2026). GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models. arXiv:2602.06718. https://arxiv.org/abs/2602.06718

Yu, Y., Li, Y., Suri, S., & Counts, S. (2026). From Searchable to Non-Searchable: Generative AI and Information Diversity in Online Information Seeking. arXiv:2604.10258. https://arxiv.org/abs/2604.10258