laigai_data_card

Authorship¶

Publishers¶

Publishing Organization: -Adam Mickiewicz University (AMU)¶

Industry Type(s)¶

Academic - Tech
Academic - Non-Tech (Psychology/Cognitive Science)

Contact Detail(s)¶

Publishing POC: Maciej Behnke
Affiliation: Adam Mickiewicz University
Contact: macbeh@amu.edu.pl, macbehnke@gmail.com
Website: https://www.affectdatabases.amu.edu.pl/
Repo: https://doi.org/10.17605/OSF.IO/V8DKM

Dataset Owners¶

Team¶

LAI‑GAI Research Team (AMU + international collaborators)

Contact Detail(s)¶

Dataset Owner(s): Maciej Behnke (PI)
Affiliation: Adam Mickiewicz University
Contact: macbeh@amu.edu.pl, macbehnke@gmail.com

Author(s)¶

Maciej Kłoskowski — Faculty of Psychology and Cognitive Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Michał Klichowski — Cognitive Neuroscience Center, Adam Mickiewicz University, Poznań, Poland; Faculty of Educational Studies, Adam Mickiewicz University, Poznań, Poland: Contributor
Wadim Krzyżaniak — Faculty of Psychology and Cognitive Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Kacper Szymański — Faculty of Psychology and Cognitive Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Patryk Maciejewski — Faculty of Psychology and Cognitive Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Patrycja Chwiłkowska — Faculty of Psychology and Cognitive Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Marta Kowal — IDN Being Human Lab – Institute of Psychology, University of Wrocław, Wrocław, Poland: Contributor
Rafał Jończyk — Cognitive Neuroscience Center, Adam Mickiewicz University, Poznań, Poland; Faculty of English, Adam Mickiewicz University, Poznań, Poland: Contributor
Jan Nowak — Network Services Division, Poznan Supercomputing and Networking Center, Poznań, Poland: Contributor
Szymon Kupiński — Network Services Division, Poznan Supercomputing and Networking Center, Poznań, Poland: Contributor
Dominika Kunc — Department of Artificial Intelligence, Wroclaw University of Science and Technology, Wrocław, Poland: Contributor
Stanisław Saganowski — Department of Artificial Intelligence, Wroclaw University of Science and Technology, Wrocław, Poland: Contributor
Aakash Chowkase — Yale Center for Emotional Intelligence, Child Study Center, Yale University, New Haven, CT, USA: Contributor
Farida Guemaz — Department of Psychology and Educational Department, Mohamed Lamine Debaghine University Setif2 Setif, Algeria: Contributor
Kevin S. Kertechian — Organization, Management and Human Resource, ESSCA School of Management, Boulogne-Billancourt, France: Contributor
Ameer I.M.T. Maadal — Department of Psychology, La Trobe University, Melbourne, Australia: Contributor
Leonardo A. Aguilar — School of Psychology, Central University of Venezuela, Caracas, Venezuela: Contributor
Barnabas T. Alayande — Center for Equity in Global Surgery, University of Global Health Equity, Kigali, Rwanda: Contributor
Vimala Balakrishnan — Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur, Malaysia; Department of Computer Science and Engineering, Korea University, Seoul 02841, Korea: Contributor
Dana M. Basnight-Brown — Department of Psychology, United States International University-Africa, Nairobi, Kenya: Contributor
Jordane Boudesseul — Laboratoire Parisien de Psychologie Sociale, University of Paris Nanterre, Paris, France; Instituto de Investigación Científica, Universidad de Lima, Peru: Contributor
Tomás A. D’Amelio — Centre de Recerca Matemàtica, Bellaterra, Spain: Contributor
Jovi C. Dacanay — School of Economics, University of Asia and the Pacific, Pasig City, Philippines: Contributor
Abhishek Dedhe — Psychology, Carnegie Mellon University, Pittsburgh, US; 22Center for the Neural Basis of Cognition, Carnegie Mellon University, Pittsburgh, US: Contributor
Shan Gao — School of Foreign Languages, University of Electronic Science and Technology of China, Chengdu, China: Contributor
Joao F. G. B. Takayanagi — Department of Psychology, University of Florida, Gainesville, United States: Contributor
Md. Rohmotul Islam — Department of Psychology, University of Chittagong, Chattogram, Bangladesh: Contributor
Alvaro Mailhos — Facultad de Psicología, Universidad de la República, Montevideo, Uruguay: Contributor
Christine M. Mpyangu — Department of Religion & Peace Studies, Makerere University, Kampala, Uganda: Contributor
Moises Mebarak — Departamento de Psicología, Universidad del Norte, Barranquilla, Colombia: Contributor
Arooj Najmussaqib — Health Department, MoPDSI, Islamabad, Pakistan: Contributor
Ju Hee Park — Department of Child and Family Studies, Yonsei University, Seoul, Korea: Contributor
Ekaterine Pirtskhalava — Department of Psychology, Ivane Javakhishvili Tbilisi State University, Tbilisi, Georgia: Contributor
Eli Rice — Department of Psychology, University of Pittsburgh, Pittsburgh, United States: Contributor
Sohrab Sami — Department of Psychology, University of Windsor, Windsor, Canada: Contributor
Yuki Yamada — Faculty of Arts and Science, Kyushu University, Fukuoka, Japan: Contributor
Jan Baczyński — Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Liliana Dera — Faculty of Anthropology and Cultural Studies, Adam Mickiewicz University, Poznań, Poland: Contributor
Szymon Jęśko-Białek — Faculty of Medicine, Prince Mieszko I Medical Academy in Poznan: Contributor
Jakub Łączkowski — Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Hubert Marciniak — Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Filip Nowicki — Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland: Contributor
Bartosz Wilczek — Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland: Contributor
James J. Gross — Stanford University, Stanford, USA: Contributor
Nicholas A. Coles — Department of Psychology, University of Florida, Gainesville, United States: Contributor

Authors' contributions: Roles and Responsibilities in the Project (CRediT taxonomy)¶

Formulating research goals and study design [Conceptualization, Methodology]: Maciej Behnke, Nicholas A. Coles.
Developing image library (image generation, quality evaluations, corrections): Maciej Behnke, Maciej Kłoskowski, Wadim Krzyżaniak, Kacper Szymański, Patryk Maciejewski, Stanisław Saganowski, Dominika Kunc, Marta Kowal, Michał Klichowski, Rafał Jończyk, Jan Baczyński, Liliana Dera, Szymon Jęśko-Białek, Jakub Łączkowski, Hubert Marciniak, Filip Nowicki, Bartosz Wilczek.
Developing study materials (e.g., rating scales, survey design) [Resources]: Maciej Behnke, Jan Nowak, Szymon Kupiński, Filip Nowicki.
Reviewing image adjustments for cultural appropriateness; offering suggestions and feedback on adjustments [Methodology, Validation]: Leonardo A. Aguilar, Barnabas T. Alayande, Vimala Balakrishnan, Dana M. Basnight-Brown, Jordane Boudesseul, Aakash Chowkase, Tomás A. D'Amelio, Jovi C. Dacanay, Abhishek Dedhe, Shan Gao, Joao F. Goes Braga Takayanagi, Farida Guemaz, Md. Rohmotul Islam, Kevin S. Kertechian, Ameer I.M.T. Maadal, Alvaro Mailhos, Christine M. Mbabazi, Moises R. Mebarak, Arooj Najmussaqib, Ju Hee Park, Ekaterine Pirtskhalava, Eli Rice, Sohrab Sami, Yuki Yamada.
Coordinating with cultural experts for consultations and approvals [Project administration]: Maciej Behnke.
Securing funding and handling financial matters [Funding acquisition]: Maciej Behnke, Michał Klichowski, Kacper Szymański.
Conducting Studies 1-6 [Investigation]: Maciej Behnke.
Analyzing data [Formal analysis]: Maciej Behnke, Stanisław Saganowski, Dominika Kunc.
Strategic decisions about manuscript structure, framing, and positioning: Maciej Behnke, Nicholas A. Coles, James J. Gross, Michał Klichowski.
Writing and revising the primary manuscript [Writing – original draft; Writing – review & editing]: Maciej Behnke, Marta Kowal, Michał Klichowski, Rafał Jończyk, Nicholas A. Coles, James J. Gross.
Reviewing and approving the final version of the manuscript [Writing – review & editing]: All authors.
Managing manuscript submission to journals [Project administration]: Maciej Behnke.
Final approval and accountability: All authors gave final approval for publication and agreed to be accountable for the work performed.

Funding Sources¶

National Science Centre in Poland (UMO-2020/39/B/HS6/00685; 2023/49/B/HS5/02139) and Excellence Initiative—Research University (ID-UB) program at Adam Mickiewicz University, Poznan (140/04/POB5/0001, 151/12/POB5/0005, 174/12/POB5/0001, 177/02/UAM/0012, 181/13/SNS/0003, 198/12/POB5/0007) supported preparing this article with a research grant. The funders had no role in study design, data collection, analysis, publishing decisions, or manuscript preparation.

Dataset Overview¶

LAI-GAI is the unified name for this project—methods, images, annotations, and app—released here as a single, versioned resource.

Data Subject(s)¶

✅ Synthetically generated data — all images are AI-generated (no real-world capture).
✅ Non-Sensitive data about people (depictions) — images portray people without identities (no PII).
✅ Non-Sensitive data about people (annotations) — human ratings collected under consent and released only in anonymized form.
✅ Data about places and objects — scenes, environments, and everyday objects used to elicit emotions.

Dataset Snapshot¶

Category | Data --- | --- Number of Instances | 847 images Target Emotions | 12 (amusement, awe, anger, attachment love, craving, disgust, excitement, fear, joy, neutral, nurturant love, sadness)
Cultural Contexts | 6 (African, Arabic, Asian, Indian, South/Central American, European/North American) Sex Variants | Sex variants consisted of matched male–female pairs for all applicable emotion categories (excluding Craving and Awe, which did not depict humans). Age Variants | Age variants consisted of matched triplets (minors, middle-aged adults, older adults) for all emotion categories,
except Nurturant Love, where depicting minors is inherent; for that category, we generated matched pairs (middle-aged adults, older adults) instead of triplets. Validation Participants | 2,470 across 58 countries (6 studies) Human Annotations | 12 discrete emotions (listed above) + six dimensional affect measures (positive, negative, arousal, calmness, approach, avoidance)

Content Description¶

Each data point is an AI-generated, photorealistic image with metadata (prompt, target emotion). Each image was rated by humans on 18 Likert-type scales (1–7): 12 discrete emotions (amusement, awe, anger, attachment love, craving, disgust, excitement, fear, joy, neutral, nurturant love, sadness) and 6 dimensional measures (positive, negative, arousal, calmness, approach, avoidance). For full details on the validation procedure and methodology, see our paper: [https://osf.io/v8dkm/files/8t34y].

Explorer app: We provide (or plan to provide) an online browser to filter and preview images by emotion, culture, sex/age variants, and success index. Access: [https://www.affectdatabases.amu.edu.pl/].

Note: The current/updated link is always listed on the OSF project page: [https://doi.org/10.17605/OSF.IO/V8DKM].

Creation Pipeline:

LAI-GAI size: 7.21 GB

Image name convention:

Culture codes: Study 3 & 4

| Code | Meaning | |---|---| | euna | Europe / North America | | af | Africa | | as | Asia (pan-Asian) | | ind | India | | arab | Arabic-speaking regions | | sa | Latin America |

Sex codes: Study 5

| Code | Meaning | |---|---| | sex1 | Female (depicted sex) | | sex2 | Male (depicted sex) |

Note. We use sex-adjusted visual depictions (male/female). These codes do not capture gender identity.

Age codes: Study 6

| Code | Meaning | |---|---| | age1 | Minors (childhood) | | age2 | Middle-aged adults | | age3 | Older adults |

Examples:

imagename_sa.png: image adapted to Latin America cultural context.
imagename_sex2.png: image adapted to depict a male individual.
imagename_age3.png: image adapted to depict older adult.

Motivations & Intentions¶

Motivations¶

Purpose(s)¶

Research - see Behnke et al., 2025 - https://osf.io/v8dkm/files/8t34y

Domain(s) of Application¶

Affective science, Emotion elicitation, Experimental psychology, Cross-cultural psychology, HCI, Behavioral science, Psychophysiology, Methods/measurement

Motivating Factor(s)¶

Address coverage gaps in legacy databases and reduce dependence on copyrighted media.
Create ethically shareable, license-clear affective image stimuli without issues common to real-photo sets.
Provide controllable variants (culture, sex, age) so researchers can test matched vs. unmatched conditions.
Supply a validated library with per-image human ratings (12 discrete emotions + 6 dimensions) to aid stimulus selection and power analysis.
Enable reproducibility via released prompts/metadata and transparent generation/curation workflow.

Intended Use¶

Dataset Use(s)¶

Safe for research use

Suitable Use Case(s)¶

Affect induction in lab/online studies (select images by target emotion or dimensional profile).
Cross-cultural investigations (compare culturally matched vs. unmatched stimuli).
Methodological work (estimating the smallest effect size of interest, SESOI, for differences in affect elicited by visual stimuli; benchmarking induction strength).
Evaluating machine learning models — testing how well models predict human ratings on 12 discrete emotions and 6 affect dimensions.
Training machine learning models images (situations) eliciting specific emotions, e.g., how to elicit disgust.

Additional Notes: Some stimuli are intentionally aversive (e.g., disgust) to ensure strong affect; use content warnings and follow IRB/ethics guidance.

Unsuitable Use Case(s)¶

Clinical or individual profiling/diagnosis (not validated for clinical decisions).
Targeted persuasion/behavioral manipulation or advertising optimization.
Production deployment where safety review is not feasible (e.g., consumer apps for minors).

Additional Notes: Images depict synthetic persons; still, abide by local content rules and avoid combining with real-person data to infer identity.

Initial Research and Problem Space(s)¶

The dataset addresses the need for modern, controllable, and transparent affective stimuli. It tests whether AI-generated images can match or exceed legacy databases in eliciting target emotions, quantifies the benefits of cultural matching, and provides sex/age variants with validation across large, international samples. Core questions include: (1) How well do AI-generated stimuli elicit specific emotions? (2) Do culturally matched images improve targeted affect? (3) How large is a practically meaningful difference between two affect-elicitation stimuli?

Detailed Methodology Description: [https://osf.io/v8dkm/files/8t34y]¶

Example of Data Points¶

Primary Data Modality¶

Image Data with Human Annotations

LAI-GAI Repository Map¶

📦 Library of AI-Generated Affective Images (LAI-GAI) [v8dkm]
└─ 🗄️  storage: osfstorage/
   ├─ Authors_Roles_preregistration.docx
   ├─ Ms_AI_Affective_Images_20062025.pdf
   ├─ Ms_AI_Affective_Images_30112025.docx
   ├─ Ms_AI_Affective_Images_Supp20062025.docx
   ├─ Ms_AI_Affective_Images_Supp_30112025.docx
   ├─ Random_order_preregistration.Rmd
   └─ Results Mock_preregistration.docx
├─ 🔗 component: k8xvh
│  📦 Images [k8xvh]
│  └─ 🗄️  storage: osfstorage/
│     ├─ 🔗 component: hcv6q
│     │  📦 All_Generated_Images [hcv6q]
│     │  └─ 🗄️  storage: osfstorage/
│     │     └─ all_Images_generated.zip
│     ├─ 🔗 component: qgney
│     │  📦 Images_Study2 [qgney]
│     │  └─ 🗄️  storage: osfstorage/
│     │     ├─ S2_scaled.zip
│     │     └─ S2_single.zip
│     ├─ 🔗 component: 6yc73
│     │  📦 Images_Study3 [6yc73]
│     │  └─ 🗄️  storage: osfstorage/
│     │     ├─ S3_4s.zip
│     │     ├─ S3_Pairs.zip
│     │     ├─ S3_Scaled.zip
│     │     └─ S3_Single.zip
│     ├─ 🔗 component: 4xbta
│     │  📦 Generated_Images_Not_Used [4xbta]
│     │  └─ 🗄️  storage: osfstorage/
│     │     └─ generated_but_not_used_in_studies.zip
│     ├─ 🔗 component: z6epq
│     │  📦 Images_Study4 [z6epq]
│     │  └─ 🗄️  storage: osfstorage/
│     │     ├─ S4_scaled.zip
│     │     └─ S4_single.zip
│     ├─ 🔗 component: 9m8ky
│     │  📦 Images_Study5 [9m8ky]
│     │  └─ 🗄️  storage: osfstorage/
│     │     ├─ S5_scaled.zip
│     │     └─ S5_single.zip
│     └─ 🔗 component: yj986
│        📦 Images_Study6 [yj986]
│        └─ 🗄️  storage: osfstorage/
│           ├─ S6_scaled.zip
│           └─ S6_single.zip
└─ 🔗 component: 8p572
   📦 Data, Analysis Code & Outputs [8p572]
   └─ 🗄️  storage: osfstorage/
      ├─ .RData
      ├─ .Rhistory
      ├─ 1README.pdf
      ├─ AIPS_preprocv3.ipynb
      ├─ AIPS_preprocv3_all.ipynb
      ├─ AIPS_preprocv3_s2.ipynb
      ├─ AIPS_preprocv3_s3.ipynb
      ├─ AIPS_preprocv3_s4.ipynb
      ├─ AIPS_preprocv3_s5.ipynb
      ├─ AIPS_preprocv3_s6.ipynb
      ├─ comparison_correlations.csv
      ├─ comparison_diff_stats.csv
      ├─ comparison_diff_stats_all.csv
      ├─ comparison_diff_stats_culture.csv
      ├─ comparison_response_distribution_summary.csv
      ├─ comparison_summary.csv
      ├─ image_emotion_means_S1.csv
      ├─ image_emotion_means_S123.csv
      ├─ image_emotion_means_S1233b45.csv
      ├─ image_emotion_means_S1233b45_2.csv
      ├─ image_emotion_means_S1233b45_emotion_corr_sorted.csv
      ├─ image_emotion_means_S123456.csv
      ├─ image_emotion_means_S123456_desc.csv
      ├─ image_emotion_means_S123456_emotion_corr_sorted.csv
      ├─ image_emotion_means_S123_desc.csv
      ├─ image_emotion_means_S123_emotion_corr_sorted.csv
      ├─ image_emotion_means_S2.csv
      ├─ image_emotion_means_S3.csv
      ├─ image_emotion_means_S3b.csv
      ├─ image_emotion_means_S4.csv
      ├─ image_emotion_means_S5.csv
      ├─ image_emotion_means_S6.csv
      ├─ image_emotion_means_updated.csv
      ├─ MMA_model_power.Rmd
      ├─ MMA_model_S1.Rmd
      ├─ MMA_model_S3.Rmd
      ├─ MMA_model_S34.Rmd
      ├─ MMA_model_S5.Rmd
      ├─ MMA_model_S6.Rmd
      ├─ osf_tree.ipynb
      ├─ prompt_dict.csv
      ├─ radar_by_category_with_image.zip
      ├─ S1233b45_data_out.csv
      ├─ S1233b4_data_out.csv
      ├─ S1233b_data_out.csv
      ├─ S12345_data_out.csv
      ├─ S1234_data_out.csv
      ├─ S123_data_out.csv
      ├─ S12_data_out.csv
      ├─ S13_data_with_cult_match_recoded.csv
      ├─ S1_1_620.txt
      ├─ S1_AI_position.csv
      ├─ S1_comparison_distribution.csv
      ├─ S1_comparison_means_adjusted.csv
      ├─ S1_data.csv
      ├─ S1_data_mlm.csv
      ├─ S1_data_mlm_2.csv
      ├─ S1_data_out.csv
      ├─ S1_data_out_recoded.csv
      ├─ S1_data_with_diff.csv
      ├─ S1_pilot.txt
      ├─ S1_pilot_data.csv
      ├─ S2_1_280.txt
      ├─ S2_281_300.txt
      ├─ S2_data.csv
      ├─ S2_data_1.csv
      ├─ S2_data_2.csv
      ├─ S2_data_out.csv
      ├─ S2_pilot0.txt
      ├─ S2_pilot01.txt
      ├─ S2_pilot1.txt
      ├─ S2_pilot_data.csv
      ├─ S2_pilot_datax.csv
      ├─ S34_data_mlm.csv
      ├─ S34_data_out.csv
      ├─ S3_AF_data.csv
      ├─ S3_AF_pilot.txt
      ├─ S3_AS_data.csv
      ├─ S3_AS_pilot.txt
      ├─ S3_comparison_distribution.csv
      ├─ S3_comparison_means_adjusted.csv
      ├─ S3_data.csv
      ├─ S3_data_mlm.csv
      ├─ S3_data_out.csv
      ├─ S3_data_out_recoded.csv
      ├─ S3_data_with_indiv_diff.csv
      ├─ S3_EUNA_data.csv
      ├─ S3_EUNA_pilot.txt
      ├─ S3_SA_data.csv
      ├─ S3_SA_pilot.txt
      ├─ S4_data.csv
      ├─ S4_data1.csv
      ├─ S4_data2.csv
      ├─ S4_data3.csv
      ├─ S4_data4.csv
      ├─ S4_data_out.csv
      ├─ S5_data.csv
      ├─ S5_data1.csv
      ├─ S5_data3.csv
      ├─ S5_data4.csv
      ├─ S5_data_mlm.csv
      ├─ S5_data_out.csv
      ├─ S6_data.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data1.csv
      ├─ S6_data2.csv
      ├─ S6_data_mlm.csv
      ├─ S6_data_out.csv
      ├─ study4_pilot.txt
      ├─ study4_pilot2.txt
      ├─ study5_pilot.txt
      ├─ study5_pilot2.txt
      ├─ study5_pilot3.txt
      ├─ study6_pilot.txt
      ├─ study6_pilot2.txt
      ├─ SupplementaryData_30112025.xlsx
      ├─ Table 2.csv
      ├─ Table 2_1.csv
      ├─ Table 3.xlsx
      ├─ Table_2_age.csv
      ├─ Table_2_culture_matched.csv
      ├─ Table_2_gender.csv
      ├─ Table_age_groups.csv
      ├─ Table_overall_stats.csv
      ├─ Table_sex1_vs_sex2.csv
      ├─ target_emotions.csv
      ├─ target_emotions0.csv
      ├─ target_emotions2.csv
      ├─ targeted_comparison_stats.csv
      ├─ targeted_comparison_stats_culture.csv
      ├─ targeted_comparison_stats_S1.csv
      ├─ targeted_comparison_stats_S3.csv
      ├─ targeted_comparison_stats_S33b.csv
      ├─ targeted_comparison_stats_S34.csv
      ├─ targeted_comparison_stats_S3b.csv
      ├─ targeted_comparison_stats_S4.csv
      ├─ targeted_emotion_with_max.csv
      └─ targeted_summary_by_emotion.csv

Participant level data¶

Example of Data Points¶

{
    "participantID": "P003",
    "set_number": 7,
    "consent": "YES",
    "page_number": 19,
    "Amusement": 2,
    "Awe": 5,
    "Anger": 1,
    "Attachment_love": 4,
    "Craving": 1,
    "Disgust": 1,
    "Excitement": 2,
    "Fear": 1,
    "Joy": 6,
    "Neutral": 3,
    "Nurturant_love": 5,
    "Sadness": 1,
    "Positive": 7,
    "Negative": 1,
    "Aroused": 1,
    "Calm": 5,
    "Approach": 4,
    "Avoid": 1,
    "Image_name": "posmidsoc_25.jpg",
    "Target_Name": null,
    "Comp_Target": null,
    "Comp_Pos": null,
    "Comp_Neg": null,
    "Comp_Aro": null,
    "Comp_Calm": null,
    "CompImage_name": null,
    "q1": 2,
    "q2": 4,
    "q3": 2,
    "q4": 2,
    "q5": 1,
    "age": 39,
    "gender": "Female",
    "country": "United Kingdom",
    "device": "laptop",
    "useData": "Yes",
    "prolific_id": "57e508cec3e5930001447391",
    "Duration": 2856552.019,
    "date_of_completion": "2025-03-17T19:26:21.449000",
    "target_emo_ind_name": "Nurturant love",
    "target_emo_ind_score": 5,
    "target_emo_comp_name": null,
    "is_AI": 0,
    "target_val": 7,
    "target_aro": 1,
    "target_mot": 4,
    "is_positive": 1,
    "is_arousing": 1,
    "is_approach": 1,
    "rating_cat": 0,
    "fail_att_check": 0,
    "fail_att_check_2": 0,
    "is_careless": 0,
    "is_careless_2": 0,
    "culture": "Europe and North America",
    "image_culture": "Other",
    "culture_matched": 0
  }

Participant-level data (field explanations)¶

Identifiers & session

participantID — anonymized respondent code (no PII).
set_number — stimulus set/block assigned to the participant (for counterbalancing).
consent — explicit consent flag (YES/NO).
page_number — page or trial index within the survey/task.
date_of_completion — ISO 8601 timestamp (UTC) for the submission.
Duration — time to complete in miliseconds.
device — participant’s device type (e.g., laptop, mobile, desktop).
useData — data-use permission flag (e.g., “Yes” to include in analyses/data release).
prolific_id — platform pseudonymous ID;

Stimulus info

Image_name — stimulus filename shown on the trial.
Target_Name — intended discrete emotion category for the target image (may be blank if not applicable on that trial).
is_AI — provenance of the specific image shown (1 = AI-generated, 0 = non-AI/legacy, if present in comparisons).
culture — participant’s cultural group label (derived from country or self-report, only for Study 3 & 4).
image_culture — cultural context label for the image (e.g., “Other”, “Europe and North America”, only for Study 3 & 4).
culture_matched — whether participant culture matches image culture (1 matched, 0 unmatched).

Ratings (per trial; 1–7 Likert)

Amusement, Awe, Anger, Attachment_love, Craving, Disgust, Excitement, Fear, Joy, Neutral, Nurturant_love, Sadness
Positive, Negative, Aroused, Calm, Approach, Avoid
Note: Higher values indicate stronger endorsement of that affect/dimension on the current image.

Target anchors & polarity flags (design/validation)

target_val, target_aro, target_mot — copy of target dimensional scale rating.
is_positive, is_arousing, is_approach — binary flags for the intended polarity of the stimulus on each dimension (1 = yes, 0 = no).

Comparison task (only on comparison trials)

Comp_Target — intended category of the comparison image.
Comp_Pos, Comp_Neg, Comp_Aro, Comp_Calm — target dimensional anchors for the comparison image.
CompImage_name — filename of the comparison image presented alongside the target.
Target_Name / target_emo_ind_name / target_emo_comp_name — name of the targeted emotion for given image (fields may be blank if that task was not shown on the trial).
target_emo_ind_score — strength/score for the chosen target category on that trial (often 1–7).

Quality control & meta

q1–q5 — attitudes toward emotional AI items (1-7 Likert): We used the following items: ”I believe emotional AI can effectively understand and respond to human emotions.", "I feel comfortable interacting with systems that use emotional AI to interpret my feelings.", "Emotional AI has the potential to improve human well-being by providing emotional support.", "I trust emotional AI to accurately identify and react to my emotional state.", "I am excited about the possibilities of integrating emotional AI into my daily life.".
fail_att_check, fail_att_check_2 — attention check outcomes (1 = failed, 0 = passed).
is_careless, is_careless_2 — careless responding indicators (algorithmic or self-report; 1 = flagged).
rating_cat — rating category: 0 - single image, 1 - image comparison, 2 - baseline measure, 3 - attention check (project-specific coding).

Demographics

age — participant age in years.
gender — self-reported gender (string).
country — country name.

Aggregated per image¶

Example of Data Points — Aggregated per image (means/SDs)¶

{
  "image_name": "amusement_new_10.jpg",
  "amusement_mean": 5.214285714,
  "awe_mean": 3.657142857,
  "anger_mean": 1.217391304,
  "attachment_love_mean": 4.142857143,
  "craving_mean": 2.428571429,
  "disgust_mean": 1.2,
  "excitement_mean": 3.871428571,
  "fear_mean": 1.285714286,
  "joy_mean": 5.042253521,
  "neutral_mean": 2.971428571,
  "nurturant_love_mean": 4.171428571,
  "sadness_mean": 1.142857143,
  "positive_mean": 5.85915493,
  "negative_mean": 1.314285714,
  "aroused_mean": 2.714285714,
  "calm_mean": 4.2,
  "approach_mean": 5.214285714,
  "avoid_mean": 1.514285714,
  "target_emo_ind_score_mean": 5.214285714,
  "amusement_std": 1.776602009,
  "awe_std": 2.166180866,
  "anger_std": 0.638692719,
  "attachment_love_std": 2.094061405,
  "craving_std": 2.06117617,
  "disgust_std": 0.67243878,
  "excitement_std": 1.962853798,
  "fear_std": 0.853685228,
  "joy_std": 1.710777766,
  "neutral_std": 1.970591235,
  "nurturant_love_std": 2.035706113,
  "sadness_std": 0.490066966,
  "positive_std": 1.245515093,
  "negative_std": 0.808341604,
  "aroused_std": 1.866119411,
  "calm_std": 1.938174847,
  "approach_std": 1.824891159,
  "avoid_std": 1.138812719,
  "target_emo_ind_score_std": 1.776602009,
  "n": 71,
  "target_emotion": "Amusement",
  "is_positive": true,
  "is_arousing": true,
  "is_approach": true,
  "target_val": 5.85915493,
  "target_aro": 2.714285714,
  "target_mot": 5.214285714,
  "is_efficient": false,
  "is_max": true,
  "is_ai": true,
  "used_in_study": 2,
  "culture": "",
  "prompt_gpt": "A photograph of a family pet doing something funny, like a dog wearing sunglasses and a hat. The background shows a cozy living room with a couch and family photos on the wall. Bright, warm lighting, playful details, hd quality, natural look"
}

Aggregated per image(field explanations)¶

*_mean, *_std — per-image mean/SD on 18 scales (12 discrete emotions + 6 dimensions).
n - number of human raters contributing to the per-image aggregates.
target_emotion — intended discrete category for the image.
is_positive, is_arousing, is_approach — binary flags for targeted valence/arousal/motivation.
target_val, target_aro, target_mot — copy of target dimensional scale rating.
is_efficient, is_max — selection flags (e.g., meets inclusion thresholds / top performer).
is_ai — image provenance flag (AI-generated).
used_in_study — study index/code if used in validation.
culture — cultural context label.
prompt_gpt — human-readable prompt/description used for generation.

Sensitivity of Data¶

Sensitivity Type(s)¶

Anonymous data
Children’s data (depictions only) — synthetic images of minors; no real minors or PII
None (S/PII) — no personally identifiable or sensitive personal data in the public release

Field(s) with Sensitive Data¶

Intentionally collected S/PII: None.
Unintentionally inferable (not labeled or released): - Possible inferences from visual context (e.g., religion via symbols, socio-economic status via settings).

Security and Privacy Handling¶

Participant-level files are anonymized.
Prompts and curation enforce age-appropriate, non-sexualized depictions of minors.
Some stimuli are aversive by design (e.g., disgust); use content warnings and follow IRB/ethics.

Risk Type(s)¶

Indirect risk: Misinterpretation of unlabeled, inferable attributes.
Residual risk: Exposure to aversive content.

Risk(s) and Mitigation(s)¶

Aversive content → Provide clear warnings; align selection with study aims and ethics approvals.

Sensitive Human Attribute(s)¶

✅ Gender
✅ Age
✅ Culture / Geography

Intentionality¶

Intentionally Collected Attributes¶

Human attributes (sex, age, and culture/geography) were designed into the stimuli to enable controlled comparisons.

Additional Notes: These are stimulus labels (what the image depicts), not participant attributes and not PII.

Unintentionally Collected Attributes (potentially inferable)¶

Human attributes not explicitly labeled but that could be inferred from visual cues.

Field Name (latent) | Description ---|--- Attractiveness / body type | Perceived attractiveness, weight, or body shape may be inferred from depictions. Socio-economic status | Implied via housing, clothing, objects (e.g., luxury goods). Religion or tradition | Implied via clothing, symbols, or settings. Disability status | Presence/absence of assistive devices or visible impairments (rare). Language/nationality | Implied by signage, scenery, or stereotypical cues.

Additional Notes: These attributes are not annotated and should not be treated as ground truth.

Rationale¶

We include sex, age, and culture labels to study (a) whether tailoring stimuli to participant demographics or cultural context improves targeted affect, and (b) whether demographic tailoring alters induction strength. Other sensitive attributes are excluded to minimize risks and unintended inferences.

Source(s)

Authoring and curation guidelines (prompt design).
Cultural expert reviews (iterative feedback and revisions).
Human ratings used to validate that tailoring did not unintentionally change targeted affect (except where intended).

Human Attribute Method (for stimulus labels)

Labels set by design at prompt time (e.g., “adult female… in [culture] context”), then verified during curation.
Cultural experts reviewed images for appropriateness; flagged stereotypes/mismatches → images revised or replaced.

Risk(s) and Mitigation(s)¶

Human Attribute: Culture

Risk type: Stereotyping / misrepresentation.
Mitigations: Cultural expert review and revision; avoid essentialist cues; document known gaps; allow user filtering; deprecate problematic items via changes.txt.

Human Attribute: Age (depictions of minors)

Risk type: Ethical concerns; accidental sexualization.
Mitigations: Strict prompt rules (age-appropriate contexts, clothing, activities); exclusion of any sexual content; additional human review before release.

Human Attribute: Sex presentation

Risk type: Reinforcing gender stereotypes.
Mitigations: Use varied roles/occupations/contexts; avoid gendered clichés; monitor with expert feedback.

Provenance¶

Collection¶

Method(s) Used

✅ Artificially Generated
✅ Crowdsourced – Paid (human ratings/validation)
✅ Taken from other existing datasets (as semantic seeds for prompts; originals not redistributed)
✅ Survey / forms (rating scales in online studies)

Methodology Detail(s)

Collection Type — Artificial Generation

Source: LLMs written prompts derived from (a) evocative images in prior affective databases (as conceptual seeds only) and (b) novel concepts brainstormed by authors. Images generated with image generative models.
Sensitive/high-risk? No PII; depictions are synthetic.
Dates of collection: Apr 2024 – Aug 2025
Primary modality: Image Data
Update frequency: Static (occasional hotfix replacements)

Collection Type — Crowdsourced Ratings

Source: Online participants rated images on 18 Likert scales (12 discrete emotions + 6 affect dimensions).
Platform: Prolific; web surveys.
Sensitive/high-risk? Low; standard consent and compensation.
Dates of collection: 2024–2025 (six studies)
Primary modality: Tabular Data (ratings)
Update frequency: Static (completed studies)

Collection Type — Prior Datasets (concept seeds)

Source: Existing affective image sets used to inspire prompts; no redistribution of originals.
Sensitive/high-risk? No, as only text concepts were used.
Primary modality: Text Data (descriptions)

Data Collection¶

Collection Method or Source:¶

Artificial generation + crowdsourced ratings

Description of Methods employed - Prompt drafting → image generation → human curation/edits → expert cultural review → re-edits (if needed) → rating studies → QC/exclusions → per-image aggregates.

Tools or libraries - Image platforms (e.g., Midjourney/Freepik/others as noted); survey platform; statistical scripts/notebooks for aggregation.

Additional Notes - Some images underwent artifact corrections (hands/faces/composition) while preserving the target affect.

Data Selection ((Images: high-level)¶

Cover 12 discrete emotions with adequate positive/neutral content (gaps in legacy sets). - Include cultural, sex, and age variants for comparisons.
Prefer photorealistic renders suitable for experimental presentation.

Data Inclusion (Images)¶

Images meeting target affect (per expert review) and passing basic quality checks.

Data Exclusion (Images)¶

Content violating model providers’ Terms of Service or ethics (e.g., sexualized minors).
Uncorrectable visual artifacts; severe mismatch with target affect.
Any item flagged by cultural experts as inappropriate after review.

Data Exclusion (Ratings)¶

Attention checks; careless-responding screens; exclusion criteria applied prior to aggregation.

Relationship to Source: Use & Utility(ies)¶

New, license-flexible stimuli designed for emotion induction, cross-cultural comparisons, and methodological benchmarks.

Benefit and Value(s)¶

Transparent prompts and metadata; controllable variants; validated aggregates to guide selection and power planning.

Limitation(s) and Trade-Off(s)¶

Model drift & moving targets. Generator capabilities, interfaces, and policies change rapidly; results validated today may shift as models evolve. This can reduce the longevity of benchmarks and may require periodic re-validation or a pivot toward validating pipelines rather than static sets.
Content filters constrain negative stimuli. Platform safety filters limited the creation of disturbing content (e.g., injuries, violence, controversial symbols), making it harder to elicit strong sadness/anger/negative affect compared to legacy datasets—though newer, less restrictive models improved feasibility for moderate negatives (e.g., disgust), which still warrant further validation.
Generation artifacts & realism gaps. Common issues included text errors in images, inconsistent context (e.g., anatomy like hands), low-resolution outputs, and difficulty rendering natural-looking faces in larger groups; additional prompt iterations and edits were often needed.
Attractiveness bias. Models tended to render uniformly attractive faces, making it challenging to generate less-attractive faces without introducing other artifacts—an ecological validity concern for some studies.
Anger is hard with still images. Anger stimuli often co-activated sadness, fear, or disgust; static images may be suboptimal for reliably isolating anger compared to dynamic media.
Cultural coverage is necessarily partial. Six broad cultural clusters (e.g., “Asian”) cannot capture within-region diversity; some cues risk stereotyping or over-emphasizing traditional elements. Continuous, local expert review was essential; within-culture disagreements were documented.
Language scope. Validation studies were conducted in English, which can attenuate responses in non-native readers; future multi-language replications are encouraged.

Mitigations recommended.
Document prompts and model/provider versions (an “immortal” textual layer), keep a light change log (e.g., changes.txt), pre-register selection/analysis where feasible, include cultural expert review rounds, and re-validate subsets when swapping generators or policies change.

Transformations¶

Synopsis¶

Transformation(s) Applied

Anomaly detection (attention checks, careless responding, outliers)
Cleaning mismatched values (labels, file paths, enums)
Cleaning missing values (drop or impute as specified)
Converting data types (ints/floats/booleans/timestamps)
Data aggregation (per-image means/SDs; success index)
Joining input sources (images ↔ prompts ↔ metadata ↔ ratings)

Field(s) Transformed — see notebooks (single source of truth)¶

To avoid repetition and keep everything reproducible, all field mappings and transformations are documented in code. Please see the OSF “Analysis Code & Outputs” component.

Where to find field definitions:
Human-readable descriptions of labels and scales are provided in the paper and its supplementary materials. This Data Card summarizes them; the canonical/authoritative definitions and mappings should be taken from the manuscript and supplement.

Libraries and Methods Used¶

All transformations (anomaly detection, cleaning, typing, joins, aggregations, and success-index computation) are fully documented and implemented in the project’s Python notebooks and scripts; treat those notebooks as the single source of truth.

Breakdown of Transformations¶

Cleaning Missing Value(s)¶

Summary: Drop trials with missing consent or failed attention checks.

Cleaning Mismatched Value(s)¶

Summary: Harmonized label spelling (e.g., Nurturant love), culture codes (euna, af, arab, as, ind, sa), and boolean fields.

Method(s) Used: Lookup tables; regex normalizers; validation against controlled vocab.

Anomalies¶

Summary: Removed trials/participants flagged by attention checks and outlier criteria (per paper).

Dimensionality Reduction (not applied to release)¶

Summary: No DR (e.g., PCA) applied to published aggregates. Researchers may apply DR downstream; report methods if used.

Joining Input Sources¶

Summary: Joined images ↔ prompts ↔ metadata ↔ ratings on image_name. Ensured one row per image in aggregate tables.
Risk & Mitigation: Key mismatches → quarantined.

Human Oversight Measure(s)¶

Multiple curation passes (artifact fixes, composition checks).
Cultural expert review and feedback with documented revisions.
Analyst review of exclusions and aggregates prior to release.

Additional Considerations¶

Provide changes.txt at the project root to record any post-release replacements/removals (date, OSF version, reason, replacement).

Annotations & Labeling¶

Annotation Workforce Type¶

Human Annotations (Crowdsourcing) — participant ratings via online surveys (Prolific).
Human Annotations (Expert) — cultural experts reviewing images for appropriateness; feedback used to revise/replace stimuli.
Machine-Generated (assistive) — model-generated prompts/descriptions (then human-edited); labels are human-authored/validated.

To avoid repetition: details of scales, procedures, and QC are expalined in detail in the paper + supplement. [https://osf.io/v8dkm/overview]

Annotation Description for Human Annotations (Crowdsourcing)¶

[edit!!!] add the script from the study + record the video of filling the survey

Validation Types¶

Method(s)¶

✅ Data Type Validation — schema/type checks for all fields.
✅ Range and Constraint Validation — e.g., ratings within [1-7]
✅ Code / Cross-reference Validation — filenames ↔ image_id ↔ metadata joins; one aggregate row per image.
✅ Structured Validation — preregistered analyses (Studies 1–3) testing induction strength, discreteness, and cultural match effects.
✅ Consistency Validation — attention checks, careless-responding screens prior to aggregation.

To avoid repetition: details of validation procedures are in the paper + supplement. [https://osf.io/v8dkm/overview]

Sampling Methods¶

Method(s) Used¶

Unsampled — the public release presents the entire dataset (no probabilistic subsetting).

Characteristic(s)¶

Sampling Type | Value ---|--- Upstream Source | Complete LAI-GAI image library (847 images) Total data sampled | Not applicable (full release) Sampling rate | Not applicable Notes | Any sampling occurs downstream (by users) for specific studies or train/test splits.

Sampling Criteria (for downstream users)¶

If you create subsets for experiments or ML splits, we recommend:

Emotion-balanced quotas (equal per the 12 emotions).
Report the OSF DOI + version and publish the exact stimulus list (image IDs).
Optional: marginal balancing for culture/sex/age (do not force all joint combos).

Extended Use¶

Use with Other Data¶

Safety Level

Conditionally safe to use with other data

Known Safe Dataset(s) or Data Type(s)

Affective image databases — e.g., legacy stimulus sets used for benchmarking at the image-level
Psychophysiology signals (task-level summaries) — e.g., ECG, SCL meausres per stimulus, with anonymized participant IDs.
Experimental metadata — trial order, timing, manipulation checks, counterbalancing keys.

Best Practices

Join on image IDs (e.g., image_name).
Cite the OSF DOI and version of LAI-GAI that you used.

Known Unsafe Dataset(s) or Data Type(s)

Face recognition / identity datasets — could enable unintended identity inference on synthetic faces.
Sensitive personal data (medical, financial, precise geolocation) — unnecessary and increases risk.
Attractiveness or demographic classification labels from automated pipelines — risk reinforcing biases.

Limitation(s) and Recommendation(s)

Limit: Synthetic photorealism may bias perceived attractiveness.
Rec: Balance selections; consider analysis covariates or robustness checks.
Limit: Cultural cues can co-vary with other signals.
Rec: Use balanced designs; pretest subsets for your population.
Limit: Negative content is constrained by generator policies.
Rec: Use multiple exemplars; pilot for effect size and ceiling/floor.

Forking & Sampling¶

Safety Level

Safe to fork and/or sample (with documentation)

Acceptable Sampling Method(s)

Random Sampling
Stratified Sampling (by emotion/culture/sex/age)
Weighted Sampling (to equalize categories)
Systematic Sampling

Best Practice(s)

Prefer stratified sampling to keep emotion and culture balanced.
Keep a record of sampled filenames/IDs and OSF version used.
For ML benchmarks, publish the stimulus list and selection script for reproducibility.

Risk(s) and Mitigation(s)

Risk: Skewed sampling inflates/deflates induction strength.
Mitigation: Predefine quotas per class; run pilot checks.
Risk: Overfitting to specific prompts/models.
Mitigation: Use cross-emotion/cross-culture validation; hold out families of stimuli.

Limitation(s) and Recommendation(s)

Limitation: Small N per niche combination. Recommendation: Pool across nearby categories or increase pilot size.

Known Correlations¶

Any known relationships (e.g., between sex_variant, age_variant, and culture, or co-varying visual cues) are documented in the manuscript. Please refer to the paper for details. [https://osf.io/v8dkm/files/8t34y]

Additional Notes:

Further analyses will appear in subsequent papers; we will update this section with summaries and links when those are published.
Correlation ceiling (upper bound): Because human ratings include noise and subjectivity, split-half human-to-human correlations provide an upper bound for model–human correlations. For this release, the macro-average split-half across all 18 scales is r = 0.935.

| Scale | Split-half r | |-------------------|:------------:| | Amusement | 0.9377 | | Anger | 0.9685 | | Attachment love | 0.9284 | | Awe | 0.8859 | | Craving | 0.9412 | | Disgust | 0.9775 | | Excitement | 0.9388 | | Fear | 0.9696 | | Joy | 0.9733 | | Neutral | 0.8649 | | Nurturant love | 0.9235 | | Sadness | 0.9769 | | Negative (valence)| 0.9859 | | Positive (valence)| 0.9829 | | Aroused | 0.6922 | | Calm | 0.9593 | | Approach | 0.9494 | | Avoid | 0.9672 |

Notes: Values are split-half human reliabilities computed from your validation ratings; the table shows the per-scale ceilings against which model–human correlations should be interpreted.

Use in ML or AI Systems¶

Dataset Use(s)

Training /Testing / Validation / Benchmarking (primary)
Fine-tuning: Not recommended for identity or demographic prediction. Acceptable for affect prediction tasks using labels/aggregates.

Notable Feature(s)

Human ratings on 18 scales per image (12 discrete + 6 dimensions) enable supervised evaluation of affect prediction models.
Participant demographics — e.g., age, gender, and country.
Cultural/sex/age variants allow fairness/robustness analyses across controlled factors.

Usage Guideline(s)

When training or evaluating models, never treat synthetic faces as identities.
Use per-image aggregates as targets for benchmarking; if using participant-level ratings, comply with ethics approvals and share only anonymized samples.
Report OSF DOI + version, the exact data split, and any preprocessing.
Include content warnings if releasing demo apps.

ML Application(s)¶

Affect prediction (regression) — predict 18 human rating means per image (12 discrete emotions + 6 dimensions).
Affect classification (multi-label / multi-class) — predict the target discrete emotion (12-way) and/or dimensional polarity flags (positive/arousing/approach).
Fairness/robustness analyses — compare errors across cultural/sex/age variants.

Evaluation Result(s)¶

No official benchmarks have been released yet.
To aid comparability, users can report the following (suggested):

Model Name: (fill in)
Typical model families: Random Forest regression; CNN finetuning (e.g., ResNet/EfficientNet); transformer/hybrid (e.g., ViT or CLIP embeddings + shallow regressor).

Hyperparameter tuning (“hypertuning”) parameters: (fill in; e.g., search strategy, search space, trials, selection metric, seed)

Evaluation Results (suggested metrics)

MAE (↓) per scale (18 values) and macro-averaged
RMSE (↓) per scale and macro-averaged
Spearman ρ (↑) with human means per scale
F1 score (↑) for 12-way target emotion (optional)

Caption: Report metrics overall. Include OSF DOI + version and the exact image list.

Additional Notes: Please share code/splits to enable replication.

Evaluation Process(es)¶

No official baselines yet; suggested protocol below.

Method used (suggested):

Split images into train/validation/test with stratification by emotion (and, if possible, balanced culture marginals).
Input: the image (pixels); Targets: per-image human means (18 scales) and/or target emotion.
Early stop on validation MAE (macro-averaged across 18 scales).

Process (minimum to report):

Preprocessing (resize, normalization), model architecture, optimizer, epochs.
Exact split manifest (IDs) and OSF version.
Metrics overall.

Results:

Provide a table: overall MAE/RMSE/ρ; confusion matrix for 12-way classification (optional).

Expected Performance and Known Caveats¶

No expectations set yet. Recommended baseline to establish:

Image-only baseline: regression to 18 targets.

Known Caveats (general):

Human ratings are subjective; ceiling is < 1.0 correlation.
Some emotions (e.g., anger) are harder to model from still images.
Synthetic photorealism may bias attractiveness → check slice performance.
Cultural cues can proxy other attributes → report fairness slices.

Additional Notes: When you publish results, please include: OSF DOI + version, exact split manifests, and the metric script so others can reproduce your scores.

Access, Retention, & Wipeout¶

Access¶

Access Type

External – Open Access

Documentation Link(s)

Main Repo (DOI): https://doi.org/10.17605/OSF.IO/V8DKM
Images component (DOI): https://doi.org/10.17605/OSF.IO/K8XVH
Analysis Code & Outputs (DOI): https://doi.org/10.17605/OSF.IO/8P572
Preregistrations (Studies 1–3): https://doi.org/10.17605/OSF.IO/72RJ3
Data Card (this file) and repository README: included in OSF project

Dataset Website URL

OSF project landing page (see DOIs above)

Prerequisite(s)

None. Public, citable DOIs.
Users must follow the lincences.
Recommended: read the paper (Behnke et al., 2025), and Data Card.

Policy Link(s)

Images: Creative Commons Attribution 4.0 (CC BY 4.0)
Metadata: Creative Commons Attribution 4.0 (CC BY 4.0)
Code: MIT License

Direct download URL / Other repository URL

Use the OSF DOIs above.

Access Control List(s)¶

Not applicable — open access.

Notes: If any file is restricted in the future (e.g., raw participant-level data), the OSF component will state access conditions and contact details. File integrity is provided via OSF versioning.

Retention¶

Duration¶

Indefinite (persistent DOIs on OSF).

Policy Summary¶

Summary: Releases are versioned on OSF. Older versions remain accessible for reproducibility; newer versions supersede prior ones. Critical issues are handled via hotfix versions with notes in the OSF version description.

Process Guide¶

Complies with OSF hosting and versioning guidelines; follows institutional ethics approvals and publication policies.

Wipeout and Deletion¶

Duration¶

No scheduled deletion. Items remain available via OSF DOIs.
Takedown is on demand (see Deletion Event Summary).

Deletion Event Summary¶

If a takedown/correction is needed:

Deprecate the item in the current release and remove it from active bundles.
Record the reason and replacement in a plain-text file changes.txt (e.g., [2025-11-13] v1.0.2 — Removed LAI-GAI_fake_image.jpg - cultural concern).
Publish a new version; keep prior versions for provenance unless OSF/legal requires full removal.

Acceptable Means of Deletion¶

Removal from active release; mark as Deprecated.
Replace with corrected asset.
(If legally required) request OSF support to suppress prior file versions.

Post-Deletion Obligations¶

Add an entry to changes.txt (date, OSF version, item removed/replaced, brief reason).
Update this Data Card if the change affects documentation or guidance.
Notify users in the project README (or OSF version description/release notes).
(If applicable) Re-run basic integrity checks (e.g., verify file counts/paths) and update any file lists.

Additional Notes¶

Because content is synthetic and non-identifying, full “right-to-be-forgotten” workflows generally do not apply; cultural or ethical concerns are handled via deprecation/replace.

Operational Requirement(s)¶

Wipeout Integration Operational Requirements¶

Maintain a simple changes.txt at the project root (date, OSF version, added/removed/replaced items, brief reason).
Validate basic file integrity before each release

Exceptions and Exemptions¶

Policy Exception bug: N/A (no internal-only retention constraints at present).

Summary¶

Open, versioned, DOI-backed dataset on OSF; no PII; deprecation-first approach for corrections; transparent changes for reproducibility.

Additional Notes¶

If you later host mirrors (e.g., Zenodo/GitHub), state mirroring cadence and point-of-truth as the OSF DOIs.

Versioning and Maintenance¶

First Version

Release date: 11/2025
Link to dataset: Main Repo (DOI): https://doi.org/10.17605/OSF.IO/V8DKM
Status: Actively Maintained (static library with occasional hotfixes)
Size of Dataset: (to be filled after upload)
Number of Instances: 847 images + annotations

Note(s) and Caveat(s)

Some stimuli are intentionally aversive for scientific purposes; use content warnings and follow IRB/ethics guidance.

Cadence

Static (with hotfix updates only)

Last and Next Update(s)

Date of last update: 11/2025
Total data points affected: (N/A for initial release)
Data points updated/added/removed: (fill on future updates)
Date of next update: (as needed; hotfix basis)

Changes on Update(s)

Future updates will list added/removed/replaced items and reasons in changes.txt.

Additional Notes

OSF version descriptions should include a short summary of changes for each update.

Citation Guidelines¶

Guidelines & Steps: Cite both the dataset (OSF components) and the associated manuscript.

BiBTeX (paper: Using AI to Generate Affective Images: Methodology and Initial Library):

@article{Behnke_LAIGAI_2025,
title = {Using AI to Generate Affective Images: Methodology and Initial Library (LAI-GAI)},
author = {Behnke, Maciej and Kłoskowski, Maciej and Klichowski, Michał and {others}},
year = {2025},
note = {Manuscript / preprint},
url = {<add preprint/DOI when available>}
}

BibTeX (dataset — Images component):

@misc{LAIGAI_images_2025,
title = {Library of AI-Generated Affective Images (LAI-GAI) — Images Component},
author = {Behnke, Maciej and collaborators},
year = {2025},
doi = {10.17605/OSF.IO/K8XVH},
url = {https://doi.org/10.17605/OSF.IO/K8XVH}
,
note = {Version X.Y}
}

Library of AI‑Generated Affective Images (LAI‑GAI)¶

Article Link¶

Dataset Link¶

Data Card Author(s)¶

Authorship¶

Publishers¶

Publishing Organization: -Adam Mickiewicz University (AMU)¶

Industry Type(s)¶

Contact Detail(s)¶

Dataset Owners¶

Team¶

Contact Detail(s)¶

Author(s)¶

Authors' contributions: Roles and Responsibilities in the Project (CRediT taxonomy)¶

Funding Sources¶

Dataset Overview¶

Data Subject(s)¶

Dataset Snapshot¶

Content Description¶

Motivations & Intentions¶

Motivations¶

Purpose(s)¶

Domain(s) of Application¶

Motivating Factor(s)¶

Intended Use¶

Dataset Use(s)¶

Suitable Use Case(s)¶

Unsuitable Use Case(s)¶

Initial Research and Problem Space(s)¶

Detailed Methodology Description: [https://osf.io/v8dkm/files/8t34y]¶

Example of Data Points¶

Primary Data Modality¶

LAI-GAI Repository Map¶

Participant level data¶

Example of Data Points¶

Participant-level data (field explanations)¶

Aggregated per image¶

Example of Data Points — Aggregated per image (means/SDs)¶

Aggregated per image(field explanations)¶

Sensitivity of Data¶

Sensitivity Type(s)¶

Field(s) with Sensitive Data¶

Security and Privacy Handling¶

Risk Type(s)¶

Risk(s) and Mitigation(s)¶

Sensitive Human Attribute(s)¶

Intentionality¶

Intentionally Collected Attributes¶

Unintentionally Collected Attributes (potentially inferable)¶

Rationale¶

Methodology Detail(s) - Sensitivity Related¶

Risk(s) and Mitigation(s)¶

Provenance¶

Collection¶

Data Collection¶

Collection Method or Source:¶

Data Selection ((Images: high-level)¶

Data Inclusion (Images)¶

Data Exclusion (Images)¶

Data Exclusion (Ratings)¶

Relationship to Source: Use & Utility(ies)¶

Benefit and Value(s)¶

Limitation(s) and Trade-Off(s)¶

Transformations¶

Synopsis¶

Field(s) Transformed — see notebooks (single source of truth)¶

Libraries and Methods Used¶

Breakdown of Transformations¶

Cleaning Missing Value(s)¶

Cleaning Mismatched Value(s)¶

Anomalies¶

Dimensionality Reduction (not applied to release)¶

Joining Input Sources¶

Human Oversight Measure(s)¶

Additional Considerations¶

Annotations & Labeling¶

Annotation Workforce Type¶

Annotation Description for Human Annotations (Crowdsourcing)¶

Validation Types¶

Method(s)¶