Large language models eroding science understanding: an experimental study

Referee: [Abstract / Methods] Abstract and Methods: The description of the model-modification procedure is insufficiently detailed to assess whether the observed behavior generalizes to publicly available LLMs. No information is given on whether the prioritization was achieved via fine-tuning, retrieval-augmented generation, heavy system-prompt weighting, or another internal mechanism; without this, it is unclear whether the setup simulates ordinary user exposure to fringe material or instead relies on non-replicable internal re-ranking.

Authors: We agree that the Methods section requires greater specificity to allow assessment of replicability and generalizability. The original manuscript described the use of custom LLMs modified to prioritize fringe papers but did not fully elaborate the technical implementation. In the revision, we will expand this to explicitly state that prioritization was implemented via fine-tuning on a dataset derived from the selected fringe papers augmented by system-prompt weighting; we will report the fine-tuning approach, dataset size, and hyperparameters used. This will clarify that the procedure is replicable in principle while acknowledging it does not directly replicate ordinary user interactions with unmodified public models. revision: yes

Referee: [Results] Results: The claim that altered-model outputs were 'difficult for non-experts to detect as misleading' is load-bearing for the central argument yet lacks supporting data. The manuscript does not report the design of any detection task, the number or demographics of non-expert participants, the exact instructions given, or any statistical comparison against expert or control conditions.

Authors: The referee is correct that the manuscript provides no formal experimental design, participant numbers, or statistical analysis for a non-expert detection task; the statement was based on qualitative assessment of output fluency by the authors rather than collected data. We will revise the Results and Discussion to qualify this claim explicitly as an observation of linguistic plausibility rather than an empirically validated finding from participant testing. No new data collection is possible at this stage, so the revision will remove the stronger phrasing and note it as a limitation requiring future work. revision: partial

Referee: [Discussion] Discussion: The conclusion that LLMs 'cannot replace expert judgment' follows from the specific manipulation performed but is not yet supported by evidence that standard, unmodified models exhibit comparable vulnerability under realistic prompting or retrieval conditions.

Authors: We accept that the conclusion should be scoped more narrowly to the manipulated models studied. The core finding is that LLMs can be made to generate consensus-contradicting yet fluent responses when trained or prompted to prioritize fringe sources, which demonstrates a vulnerability that could affect real-world use even if standard models require more effort to elicit similar outputs. In revision we will rephrase the Discussion to state that expert judgment remains necessary because models are susceptible to such influence, while adding explicit discussion of the limitation that our results do not directly test unmodified public LLMs under typical user prompting or RAG conditions. revision: yes