{"id":"b5edd0ed-580b-4c1a-bba4-643bb80dcaf5","arxiv_id":"2607.24808","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"EEG from 52 viewers let a neural network detect 'awe' responses to AI-generated biodigital architecture images at 77% accuracy in the gamma band, though two of the three emotion classes were excluded post hoc.","lead":"Researchers used a low-cost EEG headset on 52 volunteers viewing AI-generated images of 'biodigital' architecture and report that a neural network detected 'awe' responses with about 77% accuracy in the gamma frequency band. A generalist might read this as a step toward letting architects test public preferences for futuristic buildings without questionnaires — if the shaky statistics hold up.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Preprocessing's 15 Hz low-pass filter removes the 30-45 Hz gamma band that the headline 77% accuracy depends on, making the gamma-band result internally impossible.","rationale":"The reader's verdict (REJECT) and my own read align — the paper does not support its headline result. However, the reader's weakest_assumption focused on image-identity/emotion confounds, which is a serious external validity threat. I believe the more load-bearing, internal inconsistency is the 15 Hz low-pass filter used in preprocessing (§2.7) that removes the gamma band (30-45 Hz) before classification. This alone invalidates the central gamma-band accuracy claim, rendering further questions about what the classifier learned moot. The reader did mention this filter contradiction in their rationale, so there is partial agreement, but it was not elevated to the weakest assumption. The paper also concedes its own contamination and artifact problems in the limitations, which further corroborates the rejection. My concrete test — checking whether gamma-band power survives the low-pass — would settle the issue directly.","tokens_in":7817,"tokens_out":2683,"duration_ms":30169,"concrete_test":"Apply the exact preprocessing of §2.7 to a sample of raw EEG epochs (including the 15 Hz 4th-order Butterworth low-pass), then band-pass filter the cleaned data from 30-45 Hz and compute signal power/amplitude. If the gamma-band power is effectively zero or negligible relative to the pre-low-pass data, the reported 77% gamma accuracy cannot be reproduced and the result should be withdrawn. Alternatively, if the authors provide code or a data-pipeline trace, run it and confirm whether the gamma-band features are nonzero for the 'awe' classification; a simple check of the trained EEGNet weights or input spectrograms would validate whether the classifier used real gamma-band content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim — gamma-band (30-45 Hz) accuracy of 77.07% ± 13.80% for 'awe' — is contradicted by the paper's own preprocessing pipeline. Section 2.7 states that after downsampling, spike artifact correction, and blink removal, \"The data were then low-pass filtered with a 4th-order 15 Hz Butterworth filter...\" This low-pass filter attenuates frequencies above 15 Hz; the gamma band (30-45 Hz) is therefore essentially removed before any analysis. Section 3.2 then reports applying EEGNet to \"the ABCD cleaned data\" across five frequency bands, including gamma, but provides no mechanism by which gamma-band information could survive the 15 Hz low-pass. The cleaned data fed into the band-specific filters for gamma would be near-zero or purely numerical noise. Accuracy near 77% on such input cannot reflect genuine neural gamma activity; it would have to be an artifact of the filtering cascade, label leakage, or an unreported deviation from the stated pipeline. This is an internal inconsistency, not a matter of interpretation. The abstract repeats the gamma claim, underscoring that the entire contribution rests on a result the methodology excludes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an EEG study of emotional responses to AI-generated biodigital architecture images. In a pre-experiment, 336 online participants rated 600 images with emojis, and 60 extreme-scoring images were selected for three emotion categories (awe, disgust, content). Fifty-two volunteers then viewed these images in an EEG session while providing their own emoji ratings, which served as classification labels. Using EEGNet on eight channels, the authors report best classification accuracy for awe in the gamma band (77.07% ± 13.80%) and delta band (70.65% ± 10.96%), and use SHAP values to rank visual features such as greenery, granularity, and dampness. The paper concludes that gamma-band EEG and SHAP can objectively guide biodigital architecture design.","tokens_in":7988,"tokens_out":4629,"duration_ms":49669,"significance":"If the reported result were valid, it would be a useful demonstration that low-cost, eight-channel consumer EEG combined with EEGNet can decode aesthetic preference for architectural imagery, and that SHAP can rank design-relevant visual features. The study has some methodological strengths: it reports an explicit a priori sample-size estimation based on an external dataset, uses a published blink-correction algorithm (ABCD), and describes channel selection from source localization. However, the central empirical claim is internally inconsistent with the stated preprocessing pipeline, and the experimental design does not disentangle emotion decoding from image-identity or response-preparation artifacts. These issues undermine the paper's main conclusion as written.","major_comments":[{"comment":"Section 2.7 states that after spike artifact correction the data were low-pass filtered with a 4th-order 15 Hz Butterworth filter. Section 3.2 then reports EEGNet classification in the gamma band (30-45 Hz). A 4th-order Butterworth low-pass at 15 Hz attenuates 30 Hz by roughly 24 dB and 45 Hz by roughly 38 dB, so the gamma band is effectively absent from the data that would be fed to the gamma-band filter. The reported 77.07% ± 13.80% gamma-band accuracy therefore cannot reflect genuine neural gamma activity unless the pipeline was not followed as stated or label leakage occurred. Because the abstract and discussion both rest on this number, the central empirical claim is internally inconsistent.","section":"§2.7 vs §3.2"},{"comment":"The classification task conflates the emotion construct with stimulus identity and response preparation. In §2.1, images were deliberately preselected because they scored at the extremes of the same emoji rating scale later used in §2.5 as the EEG classification label. All subjects viewed the same 60 images (20 per class), and labels are the participants' own emoji responses made after each image. A classifier can therefore succeed by learning image-specific evoked potentials or motor/preparation activity before the emoji click. No shuffled-label control, image-identity control, or leave-one-image-out cross-validation is reported in §3.2. The reported accuracy does not by itself establish emotion decoding.","section":"§2.1, §2.5, §3.2"},{"comment":"The SHAP feature-importance analysis is derived from gamma-band classification, which is invalidated by the preprocessing contradiction in §2.7. Additionally, the six image characteristics (dryness, greening, curvature, smoothness, brightness, granularity, texture) are scored by ChatGPT-4 with no validation, inter-rater reliability, or evidence that these attributes are independent of the preselection criteria. The qualitative conclusions in §4.1 about greenery, granularity, and dampness therefore lack a sound empirical basis as presented.","section":"§3.3"}],"minor_comments":[{"comment":"The text cites 'Yang et al., 2026' for neural responses to emotion, but the reference list contains 'Yang, H., Liu, S., & Xie, Z. (2024). CommuniWave...', which appears unrelated to neuroscience. Please correct or replace this citation.","section":"§1 / References"},{"comment":"The caption contains a typo: 'classification accuray' should be 'classification accuracy.'","section":"Figure 3"},{"comment":"The a priori sample size calculation is based on a different dataset (imagined emotion study) with a different electrode montage and task. The translation of Cohen's d values into the chosen protocol (52 subjects, 20 trials per emotion) is not fully described; please clarify how the conservative estimate of 2,585 trials relates to the final design.","section":"§2.3"},{"comment":"For reproducibility, please specify the exact threshold criteria used to select 'awe,' 'disgust,' and 'content' images from the emoji ratings, and state the number of images in each final category.","section":"§2.1 / §2.5"}],"recommendation":"reject","confidential_remarks":"The topic is timely, and the paper attempts to combine a priori power analysis, low-cost EEG, and interpretable machine learning. However, the internal contradiction between the 15 Hz low-pass filter and the gamma-band accuracy, together with the lack of controls for image identity and label leakage, invalidates the central claim. These are not merely presentation issues; they require re-analysis or additional experiments. If the raw data are available, a corrected pipeline with proper control analyses could potentially salvage the delta-band result, but as submitted the manuscript does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nQuick take: the paper's central result—77% gamma-band decoding of 'awe'—is not supported as written. Section 2.7 says the data were low-pass filtered at 15 Hz after blink removal; Section 3.2 then reports best accuracy in the gamma band (30–45 Hz). That is an internal contradiction. Unless the authors can show a deviation from the stated pipeline, the gamma numbers describe filtered-out information. I checked the full text and this is not a typo in the abstract alone; the methods explicitly describe the 15 Hz filter and the results specifically reference gamma.\n\nThat said, the paper isn't without merit. The stimulus domain—AI-generated biodigital architecture—is genuinely new, and the authors put real effort into grounding their design choices externally: eLORETA channel selection and FDMC sample-size estimation from the Onton & Makeig dataset. The limitations section is refreshingly honest, conceding WiFi/USB interference, an uncontrolled lab environment, and algorithm limits. Those concessions align with what you'd expect from a careful group working under constraints.\n\nThe soft spots go beyond the filter. The classification task is ambiguous between binary and ternary; disgust is dropped for being below chance; and the same 20 images per class were shown to all participants, with labels coming from their own emoji responses. That means the classifier could be exploiting image-evoked potentials rather than emotion. A shuffled-label or image-identity control is missing. The SHAP feature-importance conclusions—greenery good, dampness bad—mostly re-describe how images were pre-selected, so they aren't an independent discovery.\n\nSo the reader's reject verdict is about right, but not because the idea is bad. A corrected re-analysis with subject-independent cross-validation, an image-identity control, a clear class structure, and a clarified filter cascade could salvage the core question. The external benchmark work and the honest limitations suggest the authors are capable of that.\n\nWho should read this? Anyone working on low-cost EEG for design preference will find the experimental setup instructive, but only as a cautionary example. The paper deserves a serious referee despite the flaw—the questions it raises are worth answering, and the methodological backbone (channel selection, sample sizing) is better than most. I'd send it for review, but expect a major revision before it's publishable.","headline":"The gamma-band result is internally contradicted by the paper's own 15 Hz low-pass filter, so the headline accuracy is unsupported as written, but the study has salvageable methodological bones.","tokens_in":8603,"tokens_out":2537,"would_cite":false,"duration_ms":25939,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that gamma-band EEG from eight scalp channels can classify awe responses to AI-generated biodigital architecture images with 77% accuracy.","keywords":["EEG emotion recognition","biodigital architecture","AI-generated images","gamma band","EEGNet","SHAP","architectural aesthetics","affective computing"],"falsifier":"Train the classifier on some images and test on unseen images of the same emotion categories; if accuracy does not stay above chance (or near 77%), the reported number is stimulus memorization rather than emotion decoding. Equally decisive: shuffle the response labels and confirm accuracy drops to chance; if it does not, the pipeline has learned label-independent structure.","tokens_in":7531,"feed_emoji":"🧠","tokens_out":5768,"duration_ms":58367,"temperature":0.7,"pith_summary":"The paper seeks to show that emotional responses to AI-generated biodigital architecture can be read objectively from EEG, and specifically that the gamma band (30-45 Hz) from eight scalp electrodes carries enough signal to distinguish awe from other reactions with 77.07% ± 13.80% accuracy. It also argues the same signal can be turned into design guidance: greenery and non-uniform granularity are linked to positive preference, while damp or sticky appearances trigger negative responses. If these claims hold, architects and urban designers could test aesthetic reactions without relying only on self-report, and consumer-grade EEG could serve as an objective preference measure. The paper frames this as a methodology study, using 60 images preselected from 600 AI-generated candidates and 52 EEG participants.","feed_headline":"EEG reads awe in AI architecture images at 77%","feed_subtitle":"Low-cost EEG picks out awe, and greenery plus texture explains why; designers get an objective preference test.","key_machinery":"The pipeline is the machinery: eight scalp electrodes (Fp1, Cz, CPz, CP1, CP2, Pz, O1, O2) chosen from prior source localization of an imagined-emotion dataset; artifact cleaning with adaptive blink correction and drift removal; band filtering; and EEGNet, a compact convolutional network with temporal, depthwise, and separable convolution layers, trained with binary cross-entropy. SHAP values on the gamma-band classifier are then used to rank image features. The gamma-band classifier is the load-bearing component: it produces the 77% accuracy and the feature-importance ordering. The image side uses 4x4 segmentation, K-means color extraction, and six perceptual scores (dryness, greening, curv","core_discovery":"The central claim is that a low-cost eight-channel EEG setup, cleaned with adaptive blink and drift correction and classified by a compact convolutional network, can decode the 'awe' emotion from brain responses to AI-generated biodigital architecture images above chance. In the gamma band this reaches 77.07% ± 13.80% accuracy; delta follows at 70.65% ± 10.96%. The same gamma-band model, interpreted with a game-theoretic feature attribution, yields a ranked explanation: presence of trees and plants and non-uniform granularity predict positive preference, while perceived dampness or stickiness predicts negative reactions. Disgust did not exceed the chance threshold, which the paper attributes","pith_inferences":["Because all participants viewed the same 60 images and labeled their own responses, the reported accuracy could reflect image-identity learning or response-preparation activity rather than a general emotion state; a held-out-image test would settle this.","The feature-importance links are correlational with image statistics; generating synthetic images that vary greenery, granularity, and dampness independently would test whether these are causal design drivers.","The gamma-band effect might be partly driven by low-level visual properties (contrast, texture density) rather than by emotion per se; comparing responses to matched non-architectural textures would isolate the emotional component.","If replicated, the same pipeline could extend from static images to virtual-reality walkthroughs or real buildings, where scale and movement may change the emotional response."],"forward_implications":["A compact neural net on eight EEG channels can classify one aesthetic emotion (awe) at about 77% accuracy, making objective preference testing with consumer-grade hardware a realistic target.","Gamma and delta bands are the informative frequencies for this kind of visual aesthetic judgment, so future studies can concentrate acquisition and analysis there.","Greenery and non-uniform granularity emerge as positive design levers; damp/sticky appearance emerges as a negative one, giving architects concrete, data-backed directions.","Disgust responses to unfamiliar biodigital forms did not classify above chance, so extreme or unfamiliar shapes may need different stimulus design or emotion definitions.","The a priori sample-size method (about 43 subjects for conservative power) offers a template for planning EEG emotion experiments, even though it was calibrated on an external dataset."],"fun_headline_variants":["EEG sees awe in AI architecture at 77% accuracy","Gamma band EEG detects awe in AI architecture images","EEG decodes awe from AI architecture, not disgust","Greenery and texture in AI buildings drive EEG awe","77%: EEG spots awe in AI-generated architecture"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The 77% figure is interpreted as decoding 'awe,' but every participant saw the same fixed set of 60 images and gave their own emoji response, so without an image-identity or shuffled-label control the classifier might be learning the stimuli or the preparation for the response rather than the emotion.","fun_headline_variants_meta":{"raw":{"variants":["EEG sees awe in AI architecture at 77% accuracy","Gamma band EEG detects awe in AI architecture images","EEG decodes awe from AI architecture, not disgust","Greenery and texture in AI buildings drive EEG awe","77%: EEG spots awe in AI-generated architecture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000646,"raw_usage":{"total_tokens":2776,"prompt_tokens":686,"completion_tokens":2090,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":2012}},"tokens_in":430,"tokens_out":2090,"duration_ms":14301,"temperature":1.0,"reasoning_tokens":2012,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:35:26.869870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the classifier on some images and test on unseen images of the same emotion categories; if accuracy does not stay above chance (or near 77%), the reported number is stimulus memorization rather than emotion decoding. Equally decisive: shuffle the response labels and confirm accuracy drops to chance; if it does not, the pipeline has learned label-independent structure.","supporting_citations":[],"review_version":1}