# How Cognitive Surrender Becomes AI Slop: Why No One Closes the Gap in Completeness

- Author: Shuichiro Ogawa (Design Researcher / Design Consultant), https://jamjet-iqbal.com/
- Published: 2026-11-05; updated: 2026-11-06
- Canonical (HTML): https://jamjet-iqbal.com/en/notes/cognitive-surrender-ai-slop/
- Japanese version: https://jamjet-iqbal.com/notes/cognitive-surrender-ai-slop/ (Markdown: https://jamjet-iqbal.com/notes/cognitive-surrender-ai-slop.md)
- Language: en
- Keywords: cognitive-surrender, ai-slop, evaluative-judgement, automation-bias, workslop, botshit, verification-cost, metacognition, processing-fluency, signal-detection

> This note links cognitive surrender (the tendency to adopt AI output as one's own judgment without scrutiny) and AI slop (AI-generated material passed on to others without being checked) through mechanisms on the human side. The requester's hypothesis is as follows: in situations where an expert would be expected to deliver fully finished work, AI produces output of low completeness, and slop arises when a person judges that output sufficient or does not evaluate it at all. We checked 46 sources against their full texts (3 against abstracts only) and examined the hypothesis by dividing it into three routes. In the first route, the output is not evaluated (people who consulted a wrong AI followed it on 79.8% of trials, Shaw and Nave 2026, not peer reviewed). In the second route, the output is evaluated but passed on the basis of how polished it looks (ChatGPT answers were chosen 35% of the time although 52% of them contained errors, and 39% of the errors were missed, Kabir et al. 2024). In the third route, the evaluator lacks the ability to see the gap (with AI, the Dunning-Kruger effect disappeared and participants overestimated their performance regardless of their scores, Fernandes et al. 2026; 71% did not notice sycophancy, Bo et al. 2026). The three routes also correspond to a qualitative typology from education research (3 of the 6 types of judgment in Walton et al. 2025). Once the gap leaves the producer, the recipient fills it (pull requests by less experienced developers received 4.52 times as many review comments, Asdaque et al. 2026). In addition, there is a fourth route in which the evaluative standard itself erodes because practice is handed over to AI. The hypothesis is broadly supported, but three points need correction. No one has measured the percentage of completeness, and no method for measuring it exists (Shaib et al. 2025). The 100% is not an individual's ability but the level that society expects of a role. The fluency and overconfidence of the models also hide the gap, so the cause cannot be placed on humans alone. The closest prior work is a survey on workslop (Liebscher et al. 2026, not peer reviewed; 52.7% of 962 respondents reported having sent it, more often among those with higher trust in AI), which shows an association between trust and sending but does not separate the mechanisms. A paper would add 'not evaluating' as a third component to the framework that views AI oversight through signal detection theory (Langer et al. 2024; Guo et al. 2024), and would connect the acceptance decision on generated output and the recipient's burden in a single experiment. At the time of the handoff, the novelty assessment was a provisional pass on exactly 2 axes, but paper-pipeline judged it FAIL (the full texts of the nearest prior work already stated the core).

## The same output can leave by two routes

Suppose that someone asks an AI to draft a report.
The text that comes back has consistent headings, well-formed paragraphs, and even numbers.
If an expert wrote the report, work would still remain at this point: checking where the numbers came from, cutting points to fit the reader's situation, and correcting overstatements.
That remaining work takes longer than the draft.

The requester's hypothesis is that AI slop is what results when no one does this remaining work.
In a situation where an expert would finish the work to 100% before delivering it, AI produces something at roughly 10% completeness.
If a person judges it sufficient, or does not evaluate it at all, it leaves at 10%.

The hypothesis tries to connect two bodies of literature that developed separately.
One is research on **cognitive surrender**, that is, adopting AI output as one's own judgment without scrutiny; it deals with what happens inside an individual's head ([Cognitive Surrender: The Lineage of Handing Judgment to AI, What Generative AI Changed, and the Current Evidence on Countermeasures](https://jamjet-iqbal.com/en/notes/cognitive-surrender-literature/)).
The other is research on **AI slop**, that is, AI-generated material passed on to others without being checked; it deals with what accumulates in the information environment ([AI Slop: Reading It as Outsourced Verification, Not Low Quality](https://jamjet-iqbal.com/en/notes/ai-slop-design-creativity/)).
The former measures whether people follow wrong answers, and the latter measures the volume in circulation and the burden on recipients.

We found only one empirical study that connects the two.
Liebscher et al. (2026, not peer reviewed) surveyed 962 full-time desk workers in the United States, and 52.7% reported having sent AI-generated workslop (described below) to others.
Sending was more common among people with high trust in AI, people who could not decide for themselves how to use AI, people whose organizations encouraged AI use, and people with low psychological safety at work.
On the receiving side, 38% reported having received workslop from colleagues, and they spent about 3.4 hours per month on average dealing with it.
High trust in AI is also a condition that Shaw and Nave identified as making cognitive surrender more likely.
However, this survey is based on self-report, and it does not distinguish whether senders failed to check the output, checked it and let it through, or checked it but could not see the gap.
This note examines how far the available evidence supports that distinction.

## Where the two terms overlap and where they diverge

Shaw and Nave (2026, not peer reviewed) defined cognitive surrender as the tendency to hand over judgment, effort, and responsibility to AI output, which occurs especially when the output arrives fluently, confidently, and without friction.
On trials in which the AI was wrong, participants who consulted it followed it 79.8% of the time.
Across the three experiments, 73.2% of the trials in which participants consulted the AI and the AI was wrong were cases of surrender, and only 19.7% ended with participants overriding the AI and answering correctly.

The definition on the slop side goes back to Willison (2024), who popularized the term.
He wrote that sharing artificially generated content with others without reviewing it is rude, and he called such content slop.
The same post also states that "not all AI-generated content is slop," and it reserves the term for content that is generated mindlessly and imposed on people who did not ask for it.
The center of the definition is not poor quality but the absence of review and the imposition on others.

The two terms overlap in the absence of review.
Cognitive surrender is adopting output as one's own judgment without reviewing it, and slop is unreviewed output passed on to others.
If surrendered output stays with the person, only that person bears the cost.
Once it is passed to someone else, the same output becomes slop.
This is the one point where the two diverge: surrender refers to a pattern of individual judgment, whereas slop refers to circulating material and the burden it creates.

There are three other related terms.
Hicks, Humphries, and Slater (2024) argued that LLM output is bullshit in Frankfurt's sense, because it is indifferent to truth.
This is an argument about the nature of the output, and it does not ask how people handle it.
Hannigan, McCarthy, and Spicer (2024) state in their abstract that plausible but inaccurate output becomes botshit when people use it uncritically, and they divide chatbot use into four types along two axes: whether the truth of the output can be verified and how much its truth matters.
de Fine Licht (2026) discussed the moral status of **workslop** (AI-generated material that looks adequate but lacks the substance the task requires), and showed that it can impose burdens on others even without deliberate shirking and that responsibility is determined by the distribution of control.
In this division, managers set the conditions (workload, deadlines, evaluation metrics, AI policy, and training), and workers decide whether to use AI, how far to check its output, and how much burden to pass on to others.

Botshit concerns truth, and workslop concerns missing substance.
The requester's hypothesis asks about the cognitive step one level before these two, namely, why people fail to notice what is missing, or notice it and still let it through.
Table 1 of de Fine Licht's paper (p.8) also contains rows close to the hypothesis: metrics that reward volume or polish invite surface-level output, and insufficient competence produces unintended workslop through an inability to assess AI output.
The text on the same page states that workers may accept AI output "largely as it stands."
However, the paper is about the allocation of responsibility; it lists these items as conditions under managerial control and does not separate the cognitive mechanisms.

## Dividing the hypothesis into three routes

The wording of the hypothesis, "judging it sufficient, or not evaluating it," actually contains three different events.

The first is the route of **not evaluating**.
The person adopts the output without looking at it, or while only acting as if they had looked.
This is what the definition of cognitive surrender points to most directly.

The second is the route of **evaluating, but with a lowered acceptance threshold**.
The person reads the output and notices to some extent what is missing, but lets it through anyway.
The person accepts from AI a level that they would not accept in their own work.

The third is the route of **being unable to evaluate, so that the gap is invisible**.
The person evaluates in earnest but lacks the ability to tell 10% from 100%, so 10% looks like 100%.
Kruger and Dunning (1999) state in their abstract that people with low skill in a domain bear a dual burden: they reach mistaken conclusions, and they also lack the metacognitive ability to recognize those mistakes.

The three-way distinction has already appeared once in the literature we collected.
Walton, Bearman, and colleagues traced how 26 university students in Australia worked on assignments with generative AI, and divided the judgment events into six types (Walton et al. 2025).
Three of these types roughly match the three routes above.
"Submitting AI output for the task without judging it" corresponds to the first, "adopting ideas with a low level of criticality" to the second, and "relying on AI for what one cannot do oneself" to the third.
For the third type, Walton et al. report both cases in which students relied on AI while aware of their own lack of ability and cases in which they did so without that awareness.
This study asks about the effects on learning, and it does not address whose burden the submitted work becomes later.

The three routes also call for different countermeasures.
For people who do not evaluate, inserting an evaluation step is enough; for people whose threshold has dropped, an inserted step only lets the output pass at the same low threshold.
For people who cannot tell the difference, neither a step nor a threshold works, and the evaluation has to be passed to someone else.
Treating the hypothesis as a single mechanism hides these differences.

## Not evaluating: the route of surrender

The evidence for this route is the research on cognitive surrender itself.
Sperber et al. (2010) argued that people rely heavily on information from others but also have a suite of mechanisms that scrutinize the source and content of that information (epistemic vigilance).
This route can be read as a state in which that vigilance stops working in front of AI output.
In Shaw and Nave's experiments, participants who had access to AI were more confident in their answers than those without access, even after following a wrong AI.
The participants did not notice that they had not evaluated.

Marcoccia et al. (2026, not peer reviewed) always gave participants the option to answer "I don't know" to difficult questions in five experiments (N = 3,132).
Merely seeing AI advice made participants stop withholding their answers, and on problems where the advice was mostly wrong, accuracy fell to about one third of that without AI while confidence nearly doubled.
Being able to say "I don't know" is a minimum condition for recognizing 10% output as 10%.
That condition disappears as soon as the advice is visible.

Lee et al. (2025) showed, in self-reports from 319 knowledge workers, that higher confidence in generative AI was associated with less critical thinking, and higher self-confidence with more.
The same study reports that the content of critical thinking shifts from producing to verifying, integrating, and stewardship.
Evaluation remained in the work and moved to its center.
People with higher confidence in AI do less of the work that moved there.

The conditions under which evaluation is skipped are also becoming visible.
Si et al. (2024) reported that, in a task of verifying the truth of claims, people who read LLM explanations were faster than those who searched and equally accurate, but relied too much on the explanations when the explanations were wrong.
The faster checking is completed, the more easily it counts as having checked.
Faas et al. (2026, not peer reviewed) reported, in an experiment simulating drone monitoring (N = 274), that participants in the condition where AI narrowed the options down to one felt less moral responsibility for accidents.
When no room for choice remains, both the reason to evaluate and the responsibility for not evaluating weaken.

## Evaluating, but passing on the basis of appearance

The key to the second route is that AI output has a high apparent completeness from the start, separately from the completeness of its content.
In expert work, the polish of the surface and the working-out of the content generally advance together.
In AI output, only the surface is close to 100% from the beginning.

Kabir et al. (2024) examined ChatGPT answers to 517 Stack Overflow questions and found that 52% contained errors and 77% were more verbose than human answers.
Nevertheless, in a user study with 12 participants, participants chose the ChatGPT answer 35% of the time, citing its comprehensiveness and well-organized language style, and overlooked the errors in the answers 39% of the time.
The participants did evaluate.
Their criterion was placed on form rather than content.

Liu et al. (2023) audited the answers of four generative search engines and reported that, although the answers looked fluent and useful, only 51.5% of sentences on average were fully supported by citations, and 74.5% of citations supported their sentences.
Moreover, the perceived utility of the answers was inversely correlated with citation accuracy.
The more useful an answer looks, the weaker its support.
Clark et al. (2021) examined the reasons people give when evaluating generated text and reported that reasons referring to form were about twice as frequent as reasons referring to content (form 47%, content 25%).

The mechanism by which appearance moves judgment was known before AI.
Fazio et al. (2015) state in their abstract that repeated statements become easier to process and are more easily felt to be true, and that this illusion occurs even when people know the correct answer; they called the reliance on fluency without using one's existing knowledge **knowledge neglect**.
Pennycook et al. (2015) showed that there are individual differences in the tendency to rate grammatically well-formed but meaningless statements as "profound," that this tendency is linked to an intuitive cognitive style, and that a bias toward accepting statements as true is part of this receptivity.

Generative AI returns output in exactly the form that fits this mechanism.
Steyvers et al. (2025) reported that people who read a model's default explanations overestimated the accuracy of its answers, and that longer explanations increased their confidence even when accuracy did not change.
Zhou et al. (2024) reported that on average 47% of the answers in which language models expressed confidence were wrong, and that users relied heavily on generated answers regardless of whether expressions of confidence were present.
Akpinar et al. (2026) reported that even readers with machine learning expertise could not reliably distinguish human-written abstracts from LLM-written ones, and that they rated abstracts edited by an LLM more highly than abstracts written only by humans or only by LLMs.
Even for experts, a polished appearance works as a cue to quality.

This route supports the hypothesis in a different way from surrender.
People have not abandoned evaluation.
The scale of their evaluation has shifted toward form.

## Unable to evaluate: when 10% does not look like 10%

The third route concerns the ability of the person who evaluates.
To see 10% as 10%, one must know what the remaining 90% consists of.

Prather et al. (2024) followed novice programmers using generative AI through observation and eye tracking.
Students who progressed well used AI as a tool to speed up their own plans and ignored unsuitable suggestions.
Among students who struggled, known metacognitive difficulties were made worse by AI, and they finished believing they had done better than they had, holding an illusion of competence.
The same tool widens the gap between those who can tell the difference and those who cannot.

Fernandes et al. (2026) showed that participants who solved logic problems with ChatGPT scored higher than the norm but estimated their scores as even higher.
The Dunning-Kruger effect usually seen in this task (the tendency of low performers to overestimate themselves more) disappeared with AI, and participants overestimated their performance to the same extent regardless of whether their scores were high or low.
The more participants knew about AI, the more overconfident and inaccurate their self-assessment was.
Bo et al. (2026) reported that novices who solved problems with a sycophantic chatbot could not correct their misconceptions and performed worse, while 71% did not notice the difference in sycophancy between the two chatbots.

A lack of discrimination also affects experienced workers.
In the field experiment by Dell'Acqua et al. (2026), when consultants used AI on tasks selected to lie outside the AI's capability, correct answers fell to 60% and 70.6%, compared with about 84.5% in the group without AI.
Because the boundary between what AI does well and badly does not coincide with apparent difficulty, even experienced workers find it hard to tell from the outside which output is the 10%.

Expertise weakens this route but does not close it.
In Dratsch et al. (2023), when radiologists received incorrect suggestions presented as coming from AI, the proportion of correct ratings fell from 79.7% to 19.8% in the inexperienced group, and from 82.3% to 45.5% even in the very experienced group.
In Gaube et al. (2021), incorrect advice lowered diagnostic accuracy in both groups, radiologists and physicians without that specialty.
Parasuraman and Manzey (2010) concluded in their review that automation bias occurs in both novices and experts and cannot be prevented by training or instructions.
In Perry et al. (2023), participants who used an AI assistant wrote less secure code and were also more likely to believe that their code was secure.

From the person's own side, this route cannot be distinguished from not having evaluated.
The person believes they have checked.
From the outside, it has the same consequence as the first route: something at 10% leaves without being checked.

## Once the output leaves, closing the gap falls to the recipient

Whichever of the three routes the output takes, the remaining 90% does not disappear when the output leaves the producer's hands.
Either someone fills it, or the output is used with no one filling it.

Software development is one of the rare settings where this transfer can be counted.
Asdaque et al. (2026) analyzed 22,953 pull requests (PRs) involving AI agents and reported that, compared with PRs from experienced developers, PRs from less experienced developers had 2.15 times as many commits, 1.47 times as many changed files, and 4.52 times as many review comments, a 31% lower acceptance rate, and 5.16 times longer time to resolution.
The difference in the ability to evaluate becomes, as it is, a difference in reviewers' workload.
Afroz et al. (2026, not peer reviewed) carried out a time-series analysis of more than 2 million PRs and issues in 294 repositories and reported that in 2025 the volume of PRs increased while the merge rate fell, and that for one-time contributors the merge rate fell by 18.18% relative to the counterfactual.
Recipients are becoming more cautious.
In a survey of 239 practitioners by Rahman et al. (2026, not peer reviewed), 61.1% required stronger evidence for AI-authored PRs than for human-authored ones, and the proportion who withheld approval because of remaining uncertainty rose from 19.4% for human PRs to 43.9% for AI PRs.

Stenberg, the maintainer of curl, wrote that about 20% of the vulnerability reports in 2025 were AI slop, and that as of early July only about 5% had been genuine vulnerabilities (Stenberg 2025); in January 2026 he ended the bug bounty program (Stenberg 2026).
These are the maintainer's own counts and have not been verified by a third party.
Even so, as a case in which the amount of work left to recipients shut down a program itself, it points in the same direction as the other numbers.

Economics can explain why this transfer is hard to stop.
Zhang and Zhang (2025, not peer reviewed) proposed an asymmetric model in which AI substitutes for labor in producing low-quality content but only complements human expertise in producing high-quality content.
The cost of producing 10% output falls, but the cost of filling the remaining 90% does not.
When the producer skips evaluation, that cost does not disappear but moves to the recipient.

Responsibility shifts in the same direction.
Elish (2019) called the structure in which humans with little control are made responsible for the failures of automated systems a **moral crumple zone**, and argued that it protects the technical system rather than the people.
In the case of slop, the person who finally bears the burden is not the producer who used AI but the recipient who receives and checks the output.
If the recipient does not check it, the 10% passes on to the next person.

## The standard itself erodes

The three routes so far concern judgment at one point in time.
There is one more route that works over time.
The standard of evaluation develops through accumulated experience of making things oneself, receiving others' evaluations, and revising.
If that experience is handed to AI, the standard does not develop.

Education research has a name for this ability.
Tai et al. (2018) defined **evaluative judgement** as the capability to make decisions about the quality of one's own work and that of others, and argued that it should be a goal of higher education.
Bearman et al. (2024) argued that this capability is needed even more in the era of generative AI.
They named three aspects (the quality of AI output, the quality of the process of using AI, and AI evaluating students' evaluative judgement) and described a responsibility to ensure that people do not give up the role of arbiter of quality.
The "100%" in the requester's hypothesis is a level visible only to people who have this evaluative judgement.

A lack of practice erodes both skill and standards.
In a field experiment in high school mathematics by Bastani et al. (2025), the group with access to plain GPT-4 scored 17% lower than the group without it on an exam taken without AI.
In an observational study by Budzyń et al. (2025), after AI assistance was introduced for colonoscopy, the adenoma detection rate in examinations without AI fell from 28.4% to 22.4% (as an observational study, it cannot establish causation).
Melumad and Yun (2025) reported, in seven experiments (N = 10,462), that people who learned from LLM summaries gained shallower knowledge than those who learned the same facts through web search, and that the advice they wrote on that basis was thinner, less original, and less likely to be adopted by recipients.

This route feeds back into the three earlier routes.
If the ability to evaluate does not develop, more people enter the third route (being unable to tell the difference).
As more 10% output circulates, more people also learn by using it as a model.
Slop can produce evaluators who cannot recognize slop.
However, no study has yet measured this cycle itself.
What is available are studies that measured each step of the cycle separately.

## What in the hypothesis should be corrected

Broadly, the hypothesis is supported by the evidence.
Each of the three routes has empirical support, and the transfer of the gap to recipients' work has been counted.
At the same time, correcting three points would make the argument stronger.

The first concerns the numbers "10%" and "100%."
We found no study that measured the percentage of completeness.
No method for measuring slop exists yet.
Shaib et al. (2025, not peer reviewed) started from the premise that AI slop has no agreed definition or means of measurement, and built evaluation dimensions from interviews with experts; however, binary judgments of "is it slop" varied with subjectivity and only correlated with dimensions such as coherence and relevance.
Completeness is not a property of the object but a relational quantity determined by what the output is used for and to whom it is passed.
This is also why Willison's definition did not call all AI-generated material slop.
As one's own draft, 10% can be quite useful.
The hypothesis would be more accurate if rewritten from "it is bad because it is 10%" to "something at 10% is passed, with the face of 100%, to a place that presupposes 100%."

The second concerns the phrase "an expert would deliver 100%."
The 100% is not an individual's ability but the level that society expects of a role.
When non-experts use AI, the result is often better than what they could have produced themselves.
Brynjolfsson et al. (2025) state in the abstract of the working paper version that, with generative AI assistance, less experienced agents improved in both speed and quality, while the most skilled agents gained little speed and saw a slight decline in quality.
What is an improvement by the person's own standard can still fall short of the role's standard.
Slop arises from the gap between the producer's standard and the standard of the role in which the output is placed.
This rewriting also explains why slop increases as AI use spreads.
People who feel their output is sufficient by their own standard have more opportunities to deliver it to places that require the role's standard.

The third concerns placing the cause only on the human side.
The requester focused on human-side causes, but the second route works only in combination with properties of the models: they finish the surface before the content, express excessive confidence (Zhou et al. 2024), and make people more confident as explanations become longer (Steyvers et al. 2025).
People lower their threshold partly because there is output in a form that leads them to lower it.
Asking about human-side mechanisms is right, but the evidence fits better if the claim is written not as "humans are to blame" but as "the gap disappears where the mechanisms of human evaluation and the form of model output mesh."
In addition, Vaccaro et al. (2024), who synthesized the performance of human-AI combinations, showed a tendency for combinations to lose on decision tasks and gain on content creation tasks.
Gains on content creation tasks appear in the numbers even when the scale of evaluation has shifted toward form.
It follows that slop arises most easily precisely on the side of these creation tasks.

## If this were a paper

At the time of the handoff, we judged that this analysis could be the seed of a peer-reviewed paper.
The reason was that the measures of individual judgment built up by research on cognitive surrender and automation bias, and the measures of circulation and burden built up by research on slop and workslop, had not yet been connected within a single experiment.

**Question**: In the acceptance decision made before AI output is passed to others, how much of the gap does each of the following let out: the absence of evaluation, a lowered acceptance threshold, and a lack of discrimination?

**Core of the measurement**: The acceptance decision is decomposed using signal detection theory.
AI outputs for the same task are prepared in versions of different completeness based on expert scoring, and participants decide whether to "pass it on as is or revise it."
Low discrimination (d′) indicates the third route; adequate discrimination with a criterion (c) shifted toward leniency indicates the second route; and behavioral records such as not opening the output or spending almost no time viewing it separate the absence of evaluation as the first route.
The same output is labeled either "made by AI" or "made by a colleague," and the threshold is also compared with the one participants apply to their own work.
If the time the recipient spends revising is measured, the gap that left the producer can be counted as a burden.

**Nearest prior work and what remains**:
The idea of decomposition through signal detection theory is not new in itself.
Langer, Baum, and Schlicker (2024) argued that human oversight, in which people detect inaccurate and unfair AI outputs, should be understood by separating the sensitivity of discrimination from the tendency to report an error (response bias).
They also state that for judgments of unfairness, where the ground truth is not readily available, the same decomposition can be used by choosing a standard to serve as the ground truth (p.13).
Guo et al. (2024) presented a framework that uses statistical decision theory to measure separately reliance, as the probability of following AI recommendations, and the loss caused by being unable to discriminate cues.
Neither addresses the situation of judging whether to "pass it on as is" for generated material with graded completeness, nor a third component of not evaluating at all.

Liebscher et al. (2026, not peer reviewed) showed an association between trust in AI and the behavior of sending workslop, and in their discussion they described its link with automation complacency (reduced attention to reviewing and correcting output) (p.16).
However, the study is a self-report survey and does not measure the mechanisms separately.
de Fine Licht (2026) discussed the allocation of responsibility for workslop from a normative standpoint and listed causes close to the three routes in a table, but did not measure the cognitive mechanisms separately.
Hannigan et al. (2024) divided the uses of botshit by how easily truth can be verified and how much it matters, but did not address the psychology of whether people verify.
Walton et al. (2025) found types close to the three routes qualitatively, but asked about the effects on learning and did not address circulation.
Bearman et al. (2024) made an educational argument for developing evaluative judgement and did not present a measurement framework.

**Candidate contributions**:
Three points remained.
The first was to add "not evaluating" as a third component to the two components of signal detection theory and to measure the three routes separately in the acceptance of generated material.
The second was to connect the acceptance decision and the burden of revision borne by the recipient within a single experiment.
The third was to follow the route of "the standard eroding" longitudinally.
The first is an extension of the frameworks of Langer et al. and Guo et al. and is not an entirely new method.

**Provisional scores on the five novelty axes**: RW 3, PD 3, Mat 2, Met 2, Res 2 (a provisional pass on exactly 2 axes, RW and PD, at 3).
The method axis was lowered to 2 because of Langer et al. and Guo et al.
Connecting the experiment to the recipient's burden was expected to raise Mat and Met.
The final scoring was left to the novelty-auditor of paper-pipeline.

**Judgment by paper-pipeline (2026-11-06)**: The outline as handed off was judged FAIL (PD 2, Mat 2, Met 2, Res 0).
Because it is an empirical research paper, RW was not scored.
All 31 reworked proposals were also judged FAIL.
The deciding evidence lay outside the pages cited in the handoff, in the very sources that the handoff had named as nearest prior work.
Table 1 of de Fine Licht (2026) placed causes corresponding to the three routes and the recipient's burden in one table.
Liebscher et al. (2026) had already stated, on pp. 4, 15, and 18 of their text, the core of the claims that could be built from this handoff.
Treating the neglect to verify automation separately from bias in weighting information had also been written earlier by Parasuraman and Manzey (2010) (pp. 396 and 398, according to the check by paper-pipeline).
The proposal to add a third component to the two components of signal detection theory was judged to be a case of renaming a known decomposition and moving it to another context.

## Related notes

- [Cognitive Surrender: The Lineage of Handing Judgment to AI, What Generative AI Changed, and the Current Evidence on Countermeasures](https://jamjet-iqbal.com/en/notes/cognitive-surrender-literature/): the lineage of cognitive surrender, what generative AI changed, and the evidence on countermeasures
- [AI Slop: Reading It as Outsourced Verification, Not Low Quality](https://jamjet-iqbal.com/en/notes/ai-slop-design-creativity/): reading slop as the externalization of verification
- [An Information Environment That Presumes Slop: Design That Doesn't Bet on the Ability to Discern](https://jamjet-iqbal.com/en/notes/slop-era-information-design/): designing information environments that assume slop
- [Why Do LLMs Become Sycophantic? How Preference Learning, Internal Circuits, and Input Framing Produce the Behavior](https://jamjet-iqbal.com/en/notes/llm-sycophancy-mechanisms-literature/): the mechanisms of sycophancy
- [Does Early Use of Generative AI Inhibit the Formation of Thought? A Literature Map of Cognitive Offloading and Learning](https://jamjet-iqbal.com/en/notes/ai-cognitive-offloading-learning/): a literature map of generative AI and learning

## Unverified items

- The full texts of Kruger and Dunning (1999), Fazio et al. (2015), and Hannigan et al. (2024) could not be retrieved, and they are described within the scope of their abstracts (`[要一次検証: 本文未到達 (full text not reached, abstract only; tried APA and Elsevier subscriptions, author versions, CiteSeerX, and Europe PMC)]`). All three are used only for peripheral positioning, and no numbers are cited from them.
- Brynjolfsson et al. (2025) was checked against the working paper version (arXiv 2304.11771), not the QJE text. No numbers are cited; only the direction is described.
- A correction notice is attached to the publisher's version of Walton et al. (2025). In the version used for checking (the OA version on figshare), the description of the six types cited in this note was confirmed.

## References

All accessed on 2026-11-06. See `source/review/cognitive-surrender-ai-slop/papers.md` for the ledger and verification records.

### Concepts

- Shaw, S. D., & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. SSRN Working Paper 6097646. https://doi.org/10.2139/ssrn.6097646 (preprint, not peer reviewed)
- Willison, S. (2024). Slop is the new name for unwanted AI-generated content. https://simonwillison.net/2024/May/8/slop/
- Hannigan, T. R., McCarthy, I. P., & Spicer, A. (2024). Beware of botshit: How to manage the epistemic risks of generative chatbots. *Business Horizons*, 67(5), 471–486. https://doi.org/10.1016/j.bushor.2024.03.001
- Hicks, M. T., Humphries, J., & Slater, J. (2024). ChatGPT is bullshit. *Ethics and Information Technology*, 26, 38. https://doi.org/10.1007/s10676-024-09775-5
- de Fine Licht, K. (2026). AI Workslop: The Moral Significance of Withholding Effort. *Philosophy & Technology*, 39(3). https://doi.org/10.1007/s13347-026-01164-8
- Shaib, C., Chakrabarty, T., Garcia-Olano, D., & Wallace, B. C. (2025). Measuring AI "Slop" in Text. arXiv:2509.19163. https://arxiv.org/abs/2509.19163
- Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one's own incompetence lead to inflated self-assessments. *Journal of Personality and Social Psychology*, 77(6), 1121–1134. https://doi.org/10.1037/0022-3514.77.6.1121
- Tai, J., Ajjawi, R., Boud, D., Dawson, P., & Panadero, E. (2018). Developing evaluative judgement: enabling students to make decisions about the quality of work. *Higher Education*, 76(3), 467–481. https://doi.org/10.1007/s10734-017-0220-3
- Bearman, M., Tai, J., Dawson, P., Boud, D., & Ajjawi, R. (2024). Developing evaluative judgement for a time of generative artificial intelligence. *Assessment & Evaluation in Higher Education*, 49(6), 893–905. https://doi.org/10.1080/02602938.2024.2335321
- Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. *Human Factors*, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
- Sperber, D., Clément, F., Heintz, C., Mascaro, O., Mercier, H., Origgi, G., & Wilson, D. (2010). Epistemic vigilance. *Mind & Language*, 25(4), 359–393. https://doi.org/10.1111/j.1468-0017.2010.01394.x
- Elish, M. C. (2019). Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. *Engaging Science, Technology, and Society*, 5, 40–60. https://doi.org/10.17351/ests2019.260

### Not evaluating

- Marcoccia, C., Quattrociocchi, W., & Capraro, V. (2026). AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized. arXiv:2607.13562. https://arxiv.org/abs/2607.13562 (preprint)
- Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. *CHI '25*. https://doi.org/10.1145/3706598.3713778
- Walton, J., Bearman, M., & Crawford, N. (2025). How university students work on assessment tasks with generative artificial intelligence: matters of judgement. *Assessment & Evaluation in Higher Education*. https://doi.org/10.1080/02602938.2025.2570328
- Si, C., Goyal, N., Wu, S. T., Zhao, C., Feng, S., Daumé III, H., & Boyd-Graber, J. (2024). Large Language Models Help Humans Verify Truthfulness – Except When They Are Convincingly Wrong. *NAACL 2024*, 1459–1474. https://doi.org/10.18653/v1/2024.naacl-long.81
- Faas, C., Uth, R., Sterz, S., Langer, M., & Feit, A. M. (2026). Don't blame me: How Intelligent Support Affects Moral Responsibility in Human Oversight. arXiv:2602.10701. https://arxiv.org/abs/2602.10701 (preprint)

### Evaluating with a lowered threshold

- Steyvers, M., Tejeda, H., Kumar, A., et al. (2025). What large language models know and what people think they know. *Nature Machine Intelligence*, 7, 221–231. https://doi.org/10.1038/s42256-024-00976-7
- Zhou, K., Hwang, J. D., Ren, X., & Sap, M. (2024). Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty. *ACL 2024*, 3623–3643. https://doi.org/10.18653/v1/2024.acl-long.198
- Fazio, L. K., Brashier, N. M., Payne, B. K., & Marsh, E. J. (2015). Knowledge does not protect against illusory truth. *Journal of Experimental Psychology: General*, 144(5), 993–1002. https://doi.org/10.1037/xge0000098
- Kabir, S., Udo-Imeh, D. N., Kou, B., & Zhang, T. (2024). Is Stack Overflow Obsolete? An Empirical Study of the Characteristics of ChatGPT Answers to Stack Overflow Questions. *CHI '24*. https://doi.org/10.1145/3613904.3642596
- Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. *Findings of EMNLP 2023*, 7001–7025. https://doi.org/10.18653/v1/2023.findings-emnlp.467
- Akpinar, N.-J., Avula, S., Lee, C.-J., et al. (2026). LLM or Human? Perceptions of Trust and Quality in Research Summaries. *CHI '26*. https://doi.org/10.1145/3772318.3793386
- Pennycook, G., Cheyne, J. A., Barr, N., Koehler, D. J., & Fugelsang, J. A. (2015). On the reception and detection of pseudo-profound bullshit. *Judgment and Decision Making*, 10(6), 549–563. https://doi.org/10.1017/s1930297500006999

### Unable to evaluate

- Prather, J., Reeves, B. N., Leinonen, J., MacNeil, S., et al. (2024). The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers. *ICER '24*, 469–486. https://doi.org/10.1145/3632620.3671116
- Fernandes, D., Villa, S., Nicholls, S., et al. (2026). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. *Computers in Human Behavior*, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779
- Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021). All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text. *ACL-IJCNLP 2021*, 7282–7296. https://doi.org/10.18653/v1/2021.acl-long.565
- Dratsch, T., Chen, X., Rezazade Mehrizi, M., et al. (2023). Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. *Radiology*, 307(4), e222176. https://doi.org/10.1148/radiol.222176
- Gaube, S., Suresh, H., Raue, M., et al. (2021). Do as AI say: susceptibility in deployment of clinical decision-aids. *npj Digital Medicine*, 4, 31. https://doi.org/10.1038/s41746-021-00385-9
- Perry, N., Srivastava, M., Kumar, D., & Boneh, D. (2023). Do Users Write More Insecure Code with AI Assistants? *CCS '23*, 2785–2799. https://doi.org/10.1145/3576915.3623157
- Bo, J. Y., Kazemitabaar, M., Deng, M., et al. (2026). Invisible Saboteurs: Sycophantic LLMs Mislead Novices in Problem-Solving Tasks. *CHI '26*. https://doi.org/10.1145/3772318.3791365
- Dell'Acqua, F., McFowland III, E., Mollick, E., et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. *Organization Science*, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. *Quarterly Journal of Economics*, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

### The standard itself erodes

- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. *PNAS*, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Budzyń, K., Romańczyk, M., Kitala, D., et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. *The Lancet Gastroenterology & Hepatology*, 10(10), 896–903. https://doi.org/10.1016/S2468-1253%2825%2900133-5
- Melumad, S., & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. *PNAS Nexus*. https://doi.org/10.1093/pnasnexus/pgaf316

### The burden passed to recipients

- Asdaque, S. A., Haider, I., Malik, M. U., et al. (2026). Novice Developers Produce Larger Review Overhead for Project Maintainers while Vibe Coding. *MSR '26*, 898–902. https://doi.org/10.1145/3793302.3793597
- Afroz, S., Miller, C., Menezes, T., et al. (2026). "AI Slop is DDoSing Open Source": Understanding the Impact of AI-Generated Contributions on Open Source Sustainability. arXiv:2607.04003. https://arxiv.org/abs/2607.04003 (preprint)
- Rahman, M. S., Alam, K., Roy, B., & Roy, C. K. (2026). Plausible Is Not Enough: Trust and Evidence in AI-Authored Pull Request Review. arXiv:2610.21144. https://arxiv.org/abs/2610.21144 (preprint)
- Stenberg, D. (2025). Death by a thousand slops. https://daniel.haxx.se/blog/2025/07/14/death-by-a-thousand-slops/
- Stenberg, D. (2026). The end of the curl bug-bounty. https://daniel.haxx.se/blog/2026/01/26/the-end-of-the-curl-bug-bounty/
- Zhang, Y., & Zhang, T. (2025). The Economics of Information Pollution in the Age of AI. arXiv:2509.13729. https://arxiv.org/abs/2509.13729
- Liebscher, A., Lee, A. Y., Rapuano, K. M., Kellerman, G. R., & Niederhoffer, K. G. (2026). Workslop: Examining the prevalence, antecedents and consequences of low-quality AI-generated content at work. PsyArXiv. https://doi.org/10.31234/osf.io/5f78h_v1 (preprint)
- Langer, M., Baum, K., & Schlicker, N. (2024). Effective Human Oversight of AI-Based Systems: A Signal Detection Perspective on the Detection of Inaccurate and Unfair Outputs. *Minds and Machines*, 35, 1. https://doi.org/10.1007/s11023-024-09701-0
- Guo, Z., Wu, Y., Hartline, J. D., & Hullman, J. R. (2024). A Decision Theoretic Framework for Measuring AI Reliance. *FAccT '24*. https://doi.org/10.1145/3630106.3658901
- Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. *Nature Human Behaviour*, 8(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1