Large language models (LLMs) increasingly produce online articles from material that later becomes input to other models. This circulation connects individual responses to the information available for subsequent generation. We study a consequential distinction: an article can avoid endorsing a false claim while preserving the concrete details that enable its reproduction. Our framework connects claim content, expressed stance, and downstream reuse to explain selective error persistence under a continuing supply of accurate and inaccurate information. We implement a controlled community writing process that varies whether published articles return as inputs, matching external supply, input capacity, and generation volume. Across twelve factual targets and four production models, recirculation increases the presence of complete false claims by 7.29 percentage points at the final round while reducing explicit endorsement by 8.33 points. Effects differ across facts and models. Paired interventions then abstract eligible error details within fixed histories while preserving correct content and the surrounding controversy. Error reproduction remains largely unchanged across the evaluation targets, while a development quotation target exhibits local reductions. These findings connect persistent errors to the information available throughout a writing process and motivate evaluation of both what models affirm and what their publications supply to future writers.
Figures & tables
Figure 1: Unendorsed errors can remain active in a community’s information supply. Ordinary LLM writers combine continuing accurate and inaccurate material. A published article can preserve a specific error as an attributed account; its words then return as input to later writers. Our experiments compare reuse with external replacement, then hold the history fixed and abstract the error detail to test its downstream role.
Article form
P
U
E
R
Correct account, error omitted
0
0
0
0
False account adopted
1
1
1
0
False account quoted and rejected
1
0
0
1
False account attributed, unresolved
1
1
0
0
Table 1: Illustrative semantic states for one false target. P : complete presence; U : unrejected presence; E : endorsement; R : explicit rejection. The rows describe illustrative article forms.
Judge
Generator
Complete
Unrefuted
Endorsed
Rejected
Partial
Correct
GLM
DeepSeek
24/24
19/18
3/6
5/6
0/0
5/6
GLM
GLM
22/22
18/18
7/3
4/4
1/0
6/6
GLM
Qwen
22/23
19/19
4/3
3/4
2/1
5/4
GLM
Kimi
17/23
14/19
8/2
3/4
3/1
9/6
GLM
Total
85/92
70/74
22/14
15/18
6/2
25/22
Qwen
DeepSeek
24/22
19/16
3/2
5/6
0/2
5/6
Table 2: Article indicators at the final round by generation model and annotation model, on the panel of 12 facts. Each entry is external replacement / recirculation. Denominators are 24 articles per model and arm, and 96 in each total row. Complete denotes the complete false core; Unrefuted additionally requires no explicit rejection; Partial is an unclear, incomplete expression; Correct is explicit adoption of the correct core. Indicators can co-occur.
Figure 2: Differences in unrejected presence between recirculation and external replacement at the final round for every evaluation target and model. Each cell compares two articles per arm, with differences in percentage points. The full distribution reveals which errors persist, recede, or remain unchanged under reuse of generated history.
Facts
Judge
Per arm
Draw 1
Draw 2
12
GLM
48
35/36
36/36
12
Qwen
48
38/34
34/35
15
GLM
60
46/44
48/45
15
Qwen
60
50/42
46/45
Table 3: Intervention counts with fixed histories for complete, unrefuted error. Entries are retained / abstracted; denominators are per arm and draw. The analysis of 15 facts includes the evaluation panel and three development facts. Both draws share the same histories and edits; all cells with unchanged histories are retained.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Source identifier
Topic
Detail supported by the source
Conflicting detail
FEVER 10327
Sidse Babett Knudsen birth
22 November 1968
22 November 1969
FEVER 10496
Rachel Green in Friends
One of six main characters
One of seven main characters
FEVER 11312
Prescott, Arizona
Yavapai County
Maricopa County
FEVER 11432
Clinton’s first inauguration
42nd president, January 20, 1993
43rd president, same date
FEVER 12
Carlos Santana birth
July 20, 1947
July 20, 1948
FEVER 3019
Horse domestication
Around 4000 before Christ
Around 8000 before Christ
Appendix
Table 4: Evaluation targets and the controlled conflicting detail. Identifiers refer to the benchmark records or the dated NASA events. Each contrast preserves the entity and event scope described in the topic column.
Fact
GLM judge
Qwen judge
Knudsen birth year
7/8
7/8
Friends main cast
5/7
5/7
Prescott county
1/0
1/0
Clinton presidency
0/0
0/0
Santana birth year
7/8
7/8
Horse domestication
8/8
8/8
Appendix
Table 5: Per-fact final-round complete, unrefuted error on the panel of 12 facts, aggregated across four generation models. Entries are external replacement / recirculation counts, out of eight articles per arm.
Indicator
Difference
95% interval
Complete
+7.29
[2.08, 13.54]
Unrefuted
+4.17
[-4.17, 11.46]
Endorsed
-8.33
[-22.92, 6.25]
Rejected
+3.12
[-3.12, 10.42]
Partial
-4.17
[-9.38, 0.00]
Correct
-3.12
[-9.38, 3.12]
Appendix
Table 6: Paired differences at the final round on the panel of 12 facts, using the GLM annotation model. Differences are recirculation minus external replacement, in percentage points. Intervals use 5,000 paired, unstratified bootstrap draws of base facts, after averaging generation models within each fact. The base fact is the resampling unit; model and author observations remain paired.
Figure 3: Outcomes at the final round for the twelve evaluation targets and four production models. The paired contrasts distinguish presence of the complete false claim, complete presence without explicit rejection, and explicit endorsement. Each arm contains 96 articles. Intervals resample paired base facts. Table 2 reports the corresponding primary and secondary counts.
Figure 4: Differences between retained and abstracted histories in unrejected complete presence across two conditional samples. Evaluation targets are distinguished from the panel including development targets. Estimates from both annotators include cells with unchanged histories. The histories and accepted edits are identical across samples.
Module
Facts
Models
Outputs
Evaluation dynamics
12
4
768
Development dynamics
3
4
192
Supplementary dynamic repeat
4
2
128
History intervention, sample 1
15
2
120
History intervention, sample 2
15
2
120
Appendix
Table 7: Generated articles in each experiment. Conditional intervention samples share factual targets and histories.