AI in Practice: A Multilingual Survey of 2025 BioHackathon Participants

This paper presents a multilingual survey dataset from 105 participants of the 2025 DBCLS BioHackathon in Japan, detailing AI usage patterns, challenges, and demographics across genomics and bioinformatics, accompanied by cleaned data and documentation to support research on AI adoption in science.

The Gist

Imagine a massive, global "hackathon" (a coding and science marathon) where the world's best bioinformatics experts gather to solve tricky problems in genetics and biology. In 2025, this event happened in Japan. But instead of just looking at the code they wrote, a group of researchers decided to ask a different question: "How are you actually using Artificial Intelligence (AI) in your daily work?"

This paper is the report card and the treasure map of that investigation. Here is the story of what they found, explained simply.

1. The Big Picture: A Multilingual "Pulse Check"

Think of the BioHackathon community as a giant, diverse neighborhood. Some people speak English, some Japanese, and some Thai. To get a true feel for the whole neighborhood, the researchers didn't just ask one question in one language. They handed out a multilingual survey (like a menu in three languages) to 105 people.

They wanted to know:

• Are you using AI every day, or just occasionally?
• Is AI your helpful assistant, your draft writer, or does it do the whole job for you?
• What are you scared of? (Is it privacy? Is it making mistakes?)
• Does your boss or university support you, or are you flying solo?

2. The "Cleaned" vs. "Raw" Data: The Kitchen Analogy

The researchers collected a huge pile of answers. To make this data useful for other scientists, they did two things, which is like preparing a meal:

• The Raw Ingredients (Raw Data): This is the original survey responses. It's like a basket of fresh, unpeeled vegetables. It has everything in the original languages (English, Japanese, Thai), including long stories people wrote about their successes or failures with AI. It's rich and authentic, but a bit messy to cook with if you just want quick statistics.
• The Prepared Meal (Cleaned Data): This is the version scientists can use for quick math. The researchers "peeled and chopped" the data. They:
  • Translated everything into English.
  • Fixed typos (like changing "germini" to "Gemini").
  • Removed long stories to keep the spreadsheet tidy.
  • Crucially: They scrubbed out anything that could identify a person (like names or specific URLs), ensuring everyone's privacy was safe, like serving a meal without revealing who grew the tomatoes.

3. What Did They Find? (The Flavor Profile)

While the paper is mostly about sharing the data, the summary gives us a taste of the results:

• Adoption: AI is everywhere in this community. It's not just a fancy toy; it's a tool people use to write code, analyze data, and brainstorm ideas.
• The "How": People use AI in different ways. Some use it to assist (like a co-pilot), some to draft (write a first version to edit later), and some let it complete tasks entirely.
• The Worries: Even though people love the help, they have concerns. It's like driving a car with a very smart autopilot; you worry about the car making a mistake, your data getting stolen, or the autopilot being biased against certain groups.
• The Gaps: Some people feel their institutions (universities or companies) aren't giving them enough support or clear rules on how to use these powerful tools safely.

4. Why Does This Matter? (The Map for the Future)

Why publish a survey about a survey? Think of this dataset as a map for the future of science.

• For Tool Makers: If you are building the next AI for scientists, this map tells you where the potholes are. It shows you what tools people are already using and what features they are begging for.
• For Policymakers: It helps universities and governments understand what rules they need to write. Should they ban AI? Encourage it? How do they protect privacy?
• For Researchers: It allows scientists to compare how AI is used in Japan versus Thailand versus the US. Are the fears different? Are the tools different?

The Bottom Line

This paper isn't just a list of numbers; it's a community diary. It captures a specific moment in time (2025) where science and AI are colliding. By cleaning up the data and making it open to everyone, the authors are saying: "Here is the raw truth of how we are working with AI. Take this map, explore it, and help us build a better, safer, and smarter future for science."

They didn't just ask "Do you use AI?" They asked, "How does AI fit into your life, your fears, and your dreams?" and they gave the world the answers to dig into.

Technical Summary

1. Problem Statement

While Artificial Intelligence (AI) is rapidly transforming genomics, bioinformatics, and software development, there is a lack of empirical, community-level data regarding how these tools are actually being adopted in daily scientific workflows. Existing literature often focuses on theoretical potential or specific tool benchmarks, leaving a gap in understanding:

• Real-world usage patterns: How frequently and for what specific tasks (e.g., coding, drafting, analysis) are AI tools used?
• Barriers and concerns: What obstacles (institutional, technical, ethical) hinder adoption?
• Multilingual context: How do practices and perceptions differ across linguistic and geographic boundaries within the global bioinformatics community?

This study addresses the need for a structured, anonymized dataset to quantify AI adoption, sentiment, and challenges among practitioners in the BioHackathon community.

2. Methodology

The study employed a cross-sectional survey design targeting participants and networks of the 2025 DBCLS BioHackathon in Japan.

• Survey Instrument: A Google Forms questionnaire administered between September 16–18, 2025.
• Languages: The survey was offered in English, Japanese, and Thai to ensure broad accessibility.
• Scope & Content: The questionnaire covered six domains:
  1. Demographics: Role, institution type, country, age, gender.
  2. Usage Frequency: Daily vs. infrequent use, and reasons for non-use.
  3. Applications: Specific tools used and task-level usage (Coding, Research, Writing, Teaching, Translation).
     • Usage Intensity: Respondents categorized usage as "Assist" (help with parts), "Draft" (generate first draft), or "Complete" (finish with minimal editing).
  4. Challenges & Satisfaction: Pain points and confidence levels.
  5. Institutional Support & Harms: Perceived support levels and concerns regarding bias, privacy, IP, hallucinations, and environmental impact.
  6. Narratives: Open-ended questions on success/failure stories (included in raw data only).
• Data Collection: Distributed via event Slack channels and QR codes. Only two items were mandatory; all others were optional to reduce bias and fatigue.
• Sample Size: 105 finalized responses from a diverse range of countries and institution types.

3. Data Processing and Anonymization

A rigorous data curation pipeline was implemented to balance privacy with utility:

• De-identification: Adhered to HIPAA Safe Harbor guidelines. No names, institutional affiliations, or IP addresses were collected. Optional email addresses and links to external work were removed entirely.
• Redaction: All URLs in free-text responses were replaced with [URL redacted].
• Cleaning & Standardization (English Dataset):
  • Long narratives: Dropped from the cleaned dataset to minimize re-identification risk, retaining only structured variables.
  • Categorical "Other": Preserved custom text in companion columns (e.g., field_other_text) while extracting predefined categories.
  • Country Standardization: Mapped variations (e.g., "U.K." $\to$ "United Kingdom") and handled invalid entries as missing.
  • Tool Grouping: Normalized AI tool names via pattern matching (e.g., "chatGTP" $\to$ "ChatGPT"; "copilot" $\to$ "Copilot (unspecified)").
  • Missing Data: "Prefer not to answer" responses were converted to standard NAs.
• Deliverables: The release includes raw anonymized data (original languages), a translated raw English dataset, a cleaned analysis-ready English dataset, a codebook, the original questionnaire, and a translation lookup table.

4. Key Results and Findings

Note: As this is a data release paper, the primary "results" are the dataset itself and the high-level summary statistics presented in Figure 1.

• Adoption Levels: The dataset captures the distribution of self-reported AI usage levels (daily, infrequent, researcher/developer) among 105 participants.
• Geographic Reach: Responses were collected from multiple countries, with a focus on nations having $n > 1$ in the final dataset.
• Task-Specific Utilization: The data categorizes how AI is used across specific tasks:
  • Assist: Helping with specific parts of a task.
  • Draft: Generating initial content for editing.
  • Complete: Finishing tasks with little to no editing.
• Concerns: The dataset quantifies concern levels (1–5 scale) regarding algorithmic bias, data privacy, intellectual property, misinformation/hallucinations, and environmental impact.
• Data Quality Limitations: The authors explicitly note that missing data is high for optional items and that "missing" in task-specific questions conflates three states: non-use, non-applicability, or non-response. Similarly, missing concern scores may indicate a lack of opinion rather than low concern.

5. Key Contributions

1. Multilingual Dataset: A unique resource capturing AI practices in English, Japanese, and Thai, allowing for cross-cultural and cross-linguistic comparative analysis.
2. Granular Workflow Data: Unlike general surveys, this dataset breaks down AI usage by specific scientific tasks (coding, literature review, teaching) and intensity levels (assist vs. complete).
3. Dual-Format Release: Provides both raw anonymized data (for qualitative/multilingual research) and a cleaned, standardized English dataset (for rapid quantitative analysis).
4. Comprehensive Documentation: Includes a translation lookup table, a detailed codebook, and a full PDF of the survey logic, ensuring reproducibility and ease of reuse.
5. Community Benchmark: Establishes a baseline for "AI in practice" within the bioinformatics and genomics community, specifically around the BioHackathon network.

6. Significance and Re-use Potential

This dataset is significant for several research and development domains:

• Policy and Ethics: Helps institutions understand specific concerns (e.g., IP, hallucinations) to formulate better AI governance policies.
• Tool Development: Developers of scientific AI tools can identify pain points and desired improvements based on real-world feedback from bioinformaticians.
• Comparative Studies: Enables researchers to analyze how AI adoption varies by country, institution type, and language.
• Methodological Research: Serves as a case study for designing multilingual surveys and handling missing data in sensitive professional contexts.

The dataset is openly available under a CC-BY 4.0 International license via the GigaScience repository (GigaDB), supporting open science and further meta-research on AI adoption in the life sciences.