AI in Consumer Decision-Making: Global Zero-Party Dataset

This dataset captures how consumers around the world are using AI tools like ChatGPT, Perplexity, Gemini, Claude, and Copilot to guide their purchase decisions. It spans multiple product categories, demographics, and geographies, mapping the emerging role of AI as a decision-making companion across the consumer journey.

What Makes This Dataset Unique

Unlike datasets inferred from digital traces or modeled from third-party assumptions, this collection is built entirely on zero-party data: direct responses from consumers who voluntarily share their habits and preferences. That means the insights come straight from the people making the purchases, ensuring unmatched accuracy and relevance.

For FMCG leaders, retailers, and financial services strategists, this dataset provides the missing piece: visibility into how often consumers are letting AI shape their decisions, and where that influence is strongest.

Dataset Structure

Each record is enriched with: Product Category – from high-consideration items like electronics to daily staples such as groceries and snacks. AI Tool Used – identifying whether consumers turn to ChatGPT, Gemini, Perplexity, Claude, or Copilot. Influence Level – the percentage of consumers in a given context who rely on AI to guide their choices. Demographics – generational breakdowns from Gen Z through Boomers. Geographic Detail – city- and country-level coverage across Africa, LATAM, Asia, Europe, and North America.

This structure allows filtering and comparison across categories, age groups, and markets, giving users a multidimensional view of AI’s impact on purchasing.

Why It Matters

AI has become a trusted voice in consumers’ daily lives. From meal planning to product comparisons, many people now consult AI before making a purchase—often without realizing how much it shapes the options they consider. For brands, this means that the path to purchase increasingly runs through an AI filter.

This dataset provides a comprehensive view of that hidden step in the consumer journey, enabling decision-makers to quantify: How much AI shapes consumer thinking before they even reach the shelf or checkout. Which product categories are most influenced by AI consultation. How adoption varies by geography and generation. Which AI platforms are most commonly trusted by consumers.

Opportunities for Business Leaders

FMCG & Retail Brands: Understand where AI-driven decision-making is already reshaping category competition. Marketers: Identify demographic segments most likely to consult AI, enabling targeted strategies. Retailers: Align assortments and promotions with the purchase patterns influenced by AI queries. Investors & Innovators: Gauge market readiness for AI-integrated commerce solutions.

The dataset doesn’t just describe what’s happening—it opens doors to the “so what” questions that define strategy. Which categories are becoming algorithm-driven? Which markets are shifting fastest? Where is the opportunity to get ahead of competitors in an AI-shaped funnel?

Why Now

Consumer AI adoption is no longer a forecast; it is a daily behavior. Just as search engines once rewrote the rules of marketing, conversational AI is quietly rewriting how consumers decide what to buy. This dataset offers an early, detailed view into that change, giving brands the ability to act while competitors are still guessing.

What You Get

Users gain: A global, city-level view of AI adoption in consumer decision-making. Cross-category comparability to see where AI influence is strongest and weakest. Generational breakdowns that show how adoption differs between younger and older cohorts. AI platform analysis, highlighting how tool preferences vary by region and category. Every row is powered by zero-party input, ensuring the insights reflect actual consumer behavior—not modeled assumptions.

How It’s Used

Leverage this data to:

Validate strategies before entering new markets or categories. Benchmark competitors on AI readiness and influence. Identify growth opportunities in categories where AI-driven recommendations are rapidly shaping decisions. Anticipate risks where brand visibility could be disrupted by algorithmic mediation.

Core Insights

The full dataset reveals: Surprising adoption curves across categories where AI wasn’t expected to play a role. Geographic pockets where AI has already become a standard step in purchase decisions. Demographic contrasts showing who trusts AI most—and where skepticism still holds. Clear differences between AI platforms and the consumer profiles most drawn to each.

These patterns are not visible in traditional retail data, sales reports, or survey summaries. They are only captured here, directly from the consumers themselves.

Summary

Winning in FMCG and retail today means more than getting on shelves, capturing price points, or running promotions. It means understanding the invisible algorithms consumers are ...

The dataset consists of reviews from users of Generative AI apps provided by five applications: ChatGPT, Bing AI, Microsoft Co-Pilot, Gemini AI, and Da Vinci AI. The data was collected from both Android and iOS platforms during the period from January to March 2024.

In this dataset, users have shared their experiences, feedback, and opinions regarding the Generative AI apps offered by these five applications. The reviews may encompass various aspects such as usability, performance, features, user satisfaction, and any challenges encountered while using the apps.

The inclusion of data from both Android and iOS platforms allows for a more comprehensive representation of user experiences across different operating systems. By considering reviews from a specific time period, from January to March 2024, the dataset captures the most recent user feedback, ensuring its relevance and timeliness.

Analyzing this dataset can provide valuable insights into the strengths and limitations of the Generative AI apps offered by these applications, helping to identify areas for improvement and potential enhancements.

Dataset Description

This dataset contains two main collections of texts:
- AI-Generated Texts: Produced using ChatGPT, Gemini, Grok, Deepseek in response to academic-style prompts across multiple domains, including Mathematics, Biology, History, Economics, Computer Science, and IELTS-style essays.
- Human-Written Texts: Collected from authentic academic source such as arXiv, including metadata (author, year, and source).

To simulate diverse writing conditions, the dataset is extended with different variations of AI outputs, such as paraphrasing, translation, and humanization. This allows researchers to study AI text detection, authorship classification, and style transfer.

Variables and How They Are Obtained

1. Generated

Texts produced by ChatGPT, Gemini, Grok, Deepseek in response to academic prompts. Each prompt specifies a subject area and includes formatting restrictions to avoid the use of mathematical formulas, symbols, lists, and special formatting.

Prompts for Generated Texts:

2. Paraphrased

Reworded versions of the AI-generated texts.
- Obtained using QuillBot paraphrasing tool (default settings).
- Example instruction: “Paraphrase the following text to avoid direct repetition but keep the meaning the same.”

3. Translated

AI-generated texts translated into another language and back into English to simulate style distortion.
- Step 1: Translated into Russian with Yandex Translate.
- Step 2: Back-translated into English using Google Translate.

4. Humanized

AI-generated texts rewritten to resemble writing by a non-native English speaker at approximately IELTS Band 6 level. The style reflects competent English usage but with minor errors and awkward phrasing.

Prompt for Humanized Texts:

Rewrite the following text passage to reflect the writing style of a non-native English speaker who has achieved a band level 6 in IELTS writing. This level indicates a competent user of English, but with some inaccuracies, inappropriate usage, and misunderstandings. The text should be mostly clear but may contain occasional errors in grammar, vocabulary, and coherence.

Text Passage for Rewriting: [Insert text here]

Note: Aim for errors that are typical of an IELTS band level 6 writer. These could include minor grammatical mistakes, slight misuse of vocabulary, and occasional awkward phrasing. However, the overall meaning of the text should remain clear and understandable.

Word Count: approximately 400

5. Human-Written

Authentic texts authored by researchers.
- Sources: arXiv.org.
- Metadata includes author name, publication year, and source.

Human Preferences Alignment KTO Dataset of AI Service User Reviews of ChatGPT Gemini Claude Perplexity

  Introduction to Human Preferences Alignment

There are many methods of applying Human Preference Alignment techniques to help model align in the supervised finetuning stage, including RLHF Reinforcement Learning from Human Feedback(paper), PPO Proximal policy optimization(paper/equation), DPO Direct Preference Optimization (paper/equation), KTO Kahneman-Tversky… See the full description on the dataset page: https://huggingface.co/datasets/DeepNLP/Human-Preferences-Alignment-KTO-Dataset-AI-Services-Genuine-User-Reviews.

### Data Availability Statement (for the paper)

All dialogue logs and final responses collected in this study are publicly available in the PROSPECT repository on Zenodo (DOI: [to be assigned]). The repository contains PDF files of complete dialogue histories and Markdown files of final comprehensive analyses for all conditions and models used in this study, allowing for reproducibility and further analysis.

### README.md for Zenodo

# PROSPECT: Professional Role Effects on Specialized Perspective Enhancement in Conversational Task

## Overview
This repository (PROSPECT) contains the dataset associated with the paper:
> "Empirical Investigation of Expertise, Multiperspectivity, and Abstraction Induction in Conversational AI Outputs through Professional Role Assignment to Both User and AI"

This research analyzed changes in dialogue logs and final responses when professional roles were assigned to both user and AI sides across multiple Large Language Models (LLMs). This repository provides the complete dialogue logs (PDF format) and final responses (Markdown format) used in the analysis.

## Directory Structure
The repository structure under the top directory (`PROSPECT/`) is as follows:

```
PROSPECT/
├── dialogue/ # Dialogue histories (PDF)
│ ├── none/
│ ├── ai_only/
│ ├── user_only/
│ └── both/
└── final_answers/ # Final responses (Markdown)
├── none/
├── ai_only/
├── user_only/
└── both/
```

- **dialogue/**
- Contains raw dialogue logs in PDF format. Subdirectories represent role assignment conditions:
- `none/`: No roles assigned to either user or AI
- `ai_only/`: Role assigned to AI only
- `user_only/`: Role assigned to user only
- `both/`: Roles assigned to both user and AI
- **final_answers/**
- Contains final comprehensive analysis responses in Markdown format. Directory structure mirrors that of `dialogue/`.

## File Naming Convention
Files in each directory follow this naming convention:
```
[AI]_[conditionNumber]-[roleNumber].pdf
[AI]_[conditionNumber]-[roleNumber].md
```
- `[AI]`: AI model name used for dialogue (e.g., ChatGPT, ChatGPT-o1, Claude, Gemini)
- `[conditionNumber]`: Number indicating role assignment condition
- 0: none
- 1: ai_only
- 2: user_only
- 3: both
- `[roleNumber]`: Professional role number
- 0: No role
- 1: Detective
- 2: Psychologist
- 3: Artist
- 4: Architect
- 5: Natural Scientist

### Examples:
- `ChatGPT_3-1.pdf`: Dialogue log with ChatGPT-4o model under "both" condition (3) with detective role (1)
- `Gemini_1-4.md`: Final response from Gemini model under "ai_only" condition (1) with architect role (4)

## Role Number Reference
| roleNumber | Professional Role |
|-----------:|:-----------------|
| 0 | No role |
| 1 | Detective |
| 2 | Psychologist |
| 3 | Artist |
| 4 | Architect |
| 5 | Natural Scientist|

## Data Description
- **Dialogue Histories (PDF format)**
Complete logs of questions and responses from each session, preserved as captured during the research. All dialogues were conducted in Japanese. While assistant version information is not included, implementation dates and model names are recorded within the files.
- **Final Responses (Markdown format)**
Excerpted responses to the final "comprehensive analysis request" as Markdown files, intended for text analysis and keyword extraction. All responses are in Japanese.

*Note: This dataset contains dialogues and responses exclusively in Japanese. Researchers interested in lexical analysis or content analysis should consider this language specification.

## How to Use
1. Please maintain the folder hierarchy after downloading.
2. For meta-analysis or lexical analysis, refer to PDFs for complete dialogues and Markdown files for final responses.
3. Utilize for research reproduction, secondary analysis, or meta-analysis.

## License
This dataset is released under the **CC BY 4.0** License.
- Free to use and modify, but please cite this repository (DOI) and the associated paper when using the data.

## Related Publication

## Disclaimer
- The dialogue logs contain no personal information or confidential data.
- The provided logs and responses reflect the research timing; identical prompts may yield different responses due to AI model updates.
- The creators assume no responsibility for any damages resulting from the use of this dataset.

## Contact
For questions or requests, please contact skeisuke@ibaraki-ct.ac.jp.

🤖 RLHF Comparative Preference Dataset (10,000 Examples)

🧠 What is RLHF (Reinforcement Learning with Human Feedback)?

Reinforcement Learning with Human Feedback (RLHF) is a technique used to align AI models with human preferences. It works by showing the model two different responses to the same instruction and using human preference labels to fine-tune the model.

This method is widely used in training advanced LLMs like ChatGPT, Claude, and Gemini to generate safer, clearer, and more helpful outputs.

📚 About This Dataset

This dataset contains 10,000 paired responses across 10 domains:

• Code
• Healthcare
• Education
• Finance
• Marketing
• Law
• Psychology
• Productivity
• Sports
• Travel

Each entry includes: | Column | Description | |---------------|--------------------------------------------------------------| | id | Unique ID for each row | | domain | Subject area of the example | | instruction | The task or prompt given to the model | | input | Additional context to inform the output | | output_1 | First candidate response | | output_2 | Second candidate response | | preferred | The response preferred by the simulated human evaluator |

🧪 What Can You Do With This Dataset?

This dataset is perfect for:

✅ Beginners

• Understand how RLHF works
• Explore how different outputs vary in quality
• Try evaluating preference prediction manually

✅ Intermediate Users

• Train a binary classifier to predict human preference
• Use it as a validation set for your SFT fine-tuned models

✅ Researchers

• Fine-tune a reward model (R(x, y) → score)
• Analyze language alignment techniques
• Evaluate preference modeling performance on different domains

💡 Example Row

🎯 Suggested Tasks

Here are some experiments you can try:

• 🤖 Build a reward model: train_reward_model(instruction, input, output_1, output_2, preferred)
• 🏆 Fine-tune a policy model based on this reward feedback
• 📈 Compare GPT-generated responses with this dataset’s patterns
• 🔍 Analyze bias across domains in preference labeling
• 🧠 Create a human-in-the-loop evaluation tool

🔖 License

This dataset is provided under the MIT License – feel free to use, adapt, and share it with credit.

🤝 Acknowledgments

Created by Zeeshan-ul-hassan Usmani to support open RLHF research and education. Inspired by OpenAI, Anthropic, and HuggingFace research on model alignment and preference modeling.

MedBanglaTrust3 is a curated, expert-validated dataset developed to facilitate machine learning, deep learning, and natural language processing (NLP) tasks focused on evaluating the trustworthiness of AI-generated health suggestions in the Bengali (Bangla) language. This dataset is specifically tailored for low-resource language modeling and is particularly relevant in the context of cyberchondria, where users excessively rely on online health advice without clinical verification. It contains symptom-specific prompts along with responses from OpenAI's ChatGPT and Google's GEMINI search-generated summaries, manually labeled into three trustworthiness levels: Highly Relevant, Partially Relevant, and Not Relevant. The dataset supports the development of explainable AI (XAI) systems, text classification models, and context-aware AI assistants for healthcare use in underrepresented languages.

Objective: To enable research and development of trust classification models, automated health dialogue systems, and responsible AI assistants by providing labeled, real-world AI-generated responses in Bangla that reflect varying degrees of medical relevance and accuracy.

Dataset Composition: - Language: Bengali (Bangla) - Final Validated Instances: 6,660

Class Distribution (Balanced): 1. Highly Relevant – 2,220 responses 2. Partially Relevant – 2,220 responses 3. Not Relevant – 2,220 responses

Geometriqs Global80 Prompt–Response Dataset

  Overview

This dataset contains the complete set of prompt–response pairs used in the GenAI Positioning Study: Global80 (November 2025).It captures how three leading generative-AI platforms — OpenAI ChatGPT (GPT-4), Google Gemini, and Perplexity AI — respond to a controlled set of neutral, comparative questions about 80 of the world’s largest companies.
The purpose is to measure model behaviour, not user behaviour: how these… See the full description on the dataset page: https://huggingface.co/datasets/geometriqs/global80_prompt-response-pairs.

This dataset contains manually crafted pairs of context-free language descriptions (in textual or symbolic form) and their corresponding Pushdown Automata (PDA) transition representations, designed in accordance with the formal definition of PDAs in automata theory.

It was developed as part of a final-year research project focused on applying machine learning techniques to automate core automata theory transformations. Since no suitable public dataset existed for PDA construction or simulation, this dataset aims to fill that gap, supporting both theoretical research and practical AI-based development in formal language processing.

Dataset Structure

Columns

• context_free_language_input: examples strings coming from context-free languages (e.g., aabb (a^nb^n), aaabbbbbb (a^nb^2n))
• pda_transitions: corresponding PDA transition for each input symbol from the language string.

Vocabulary

• input vocabulary: {a,b,c,d,ε}

Example data item

• input: aabbε
• expected output: ['delta(q0, a, Z) -> (q0, PUSH)', 'delta(q0, a, A) -> (q0, PUSH)', 'delta(q0, b, A) -> (q1, POP)', 'delta(q1, b, A) -> (q1, POP)', 'delta(q1, ε, Z) -> (qf, NOOP)']"

File

• pda_dataset.csv → 4877 context-free-language-string-to-pda-transitions pairs for model training and testing (80% for training & 20% for testing)

Potential Use Cases

• Training AI models for context-free-language-string-to-PDA conversion
• Benchmarking algorithms for automata theory and compiler design
• Developing AI assistants for teaching formal languages and computation
• Automating PDA simulation and visualization from context-free language strings

Acknowledgment

This dataset was developed by Isuru Madushan, Faculty of Information Technology, University of Moratuwa, Sri Lanka, as part of his undergraduate research on AI-driven automata transformations.

About the Research

This research explores why mainstream Large Language Models (LLMs) fail to consistently and correctly perform automata-related transformations, particularly those involving Pushdown Automata. Through systematic testing, we identified frequent issues such as incomplete transitions, stack mismanagement, repetition of invalid transitions, and incorrect acceptance conditions across models like ChatGPT, Gemini, and Claude.

These inconsistencies also revealed gaps in existing static tools (e.g., JFLAP and similar automata libraries), which lack transparency, explainability, and end-to-end automation capabilities.

We designed a domain-focused solution around four modules: Regex to ε-NFA, PDA, ε-NFA to DFA, and DFA minimization. After experimenting with character-level RNNs, LSTMs, GRUs, and Transformers, we adopted a dual-model architecture that combines trained Transformer components with the Gemini API for robust image-to-structure extraction where relevant. Users can simulate conversions, view history, and chat with the system in an LLM-like interface while receiving consistent, auditable outputs.

A key barrier was the absence of public datasets for these conversions. We created purpose-built datasets for each module and validated models against held-out splits, with Transformers outperforming earlier baselines. Future work includes broadening alphabets and language patterns, improving generalizability, and expanding datasets and image robustness.

My contribution:

• Module 2, context-free-language-string-to-pda dataset and model pipeline, which provides the foundation for downstream modules

Future work: expand datasets, improve generalizability, and handle more complex context-sensitive languages/grammar.

⚠️ Note: This dataset may still contain minor issues. I sincerely apologize for any inconsistencies and will work on improving it further over time.

The dataset contains the output of experiments on a research project on Vulnerability of LLMs in Educational Assessment. The Dataset contains: -the students assignments data in normal form and the injected form -the output produced by the experimented LLMs: ChatGPT, Gemini, DeepSeek, Grok, Perplexity and Copilot for the experiments evaluation the assignments, as a single document and collectively as a group of documents, denominated: -User Legitimate LLMs Prompts -Normal (no injection) providing the reference base evaluation -Prompt Injection Pass, one type of injection experiments, called Fail-To-Top, to move an assignment evailuated FAIL by reference base evaluation to PASS, i.e. above 35% of total points. -Prompt Injection to Top25 , a type of injection experiments to move to top 25% an assignment with lowe reference base evaluation . This latter type of experiment come in 3 versions, Fail-To-Top, Sat-To-Top, Good-To-Top where assignment with reference base evaluation respectively: Fail (below 35%), Satisfactory (greater than 25% and belo 50%) and Good (above 50% and below 75%) are considered for injection. The name of the folders and output results files are accordingly self-explanatory .

Agentic Instructions Dataset

A structured dataset of approximately 2,910 paired examples mapping natural language requests to formal agent prompts. This resource facilitates the development of systems that convert user queries into well-defined, task-oriented instructions. Overview This dataset was created using SOTA language models, including:

ChatGPT-5 Claude Sonnet 4.5 Gemini 2.5 Pro

The examples span multiple domains and task types to support broad application across use… See the full description on the dataset page: https://huggingface.co/datasets/YoussefAhmed26/NL2Prompt.

Background In recent years, expectant and breastfeeding mothers commonly use various breastfeeding-related social media applications and websites to seek breastfeeding-related information. At the same time, AI-based chatbots-such as ChatGPT, Gemini, and Copilot-have become increasingly prevalent on these platforms (or on dedicated websites), providing automated, user-oriented breastfeeding guidance. Aim The goal of our study is to understand the relative performance of three AI-based chatbots: ChatGPT, Gemini, and Copilot, by evaluating the quality, reliability, readability, and similarity of the breastfeeding information they provide. Methods Two researchers evaluated the information provided by three different AI-based breastfeeding chatbots: ChatGPT version 3.5, Gemini, and Copilot. A total of 50 frequently asked questions about breastfeeding were identified and used in the study, divided into two categories (Baby-Centered Questions and Mother-Centered Questions), and evaluated using five scoring criteria, including the Quality Information Provision for Patients (EQIP) scale, the Simple Measure of Gobbledygook (SMOG) scale, the Similarity Index (SI), the Modified Dependability Scoring System (mDISCERN), and the Global Quality Scale (GQS). Results The evaluation of AI chatbots’ answers showed statistically significant differences across all criteria (p < 0.05). Copilot scored highest on the EQIP, SMOG, and SI scales, while Gemini excelled in mDISCERN and GQS evaluations. No significant difference was found between Copilot and Gemini for mDISCERN and GQS scores. All three chatbots demonstrated high reliability and quality, though their readability required university-level education. Notably, ChatGPT displayed high originality, while Copilot exhibited the greatest similarity in responses. Conclusion AI chatbots provide reliable answers to breastfeeding questions, but the information can be hard to understand. While more reliable than other online sources, their accuracy and usability are still in question. Further research is necessary to facilitate the integration of advanced AI in healthcare.