What Is AI Research Ethics? Principles and a Checklist for Researchers, Covering Data, Generative AI, Copyright, and Bias

Before you read: This article is for general information only and is not legal advice.
Laws and policies change often, so always check the latest official sources. The examples are hypothetical and meant to aid understanding.

Introduction: Is High Performance Enough to Make Good Research?

Suppose a model reaches 95% accuracy. But it was built on data collected without anyone's consent, and its errors are concentrated in one particular group. The number looks impressive, yet it is hard to call this good research.

AI research ethics starts from questions like this. It asks not only "can we build it?" but also "how did we build it, and who is affected, and how?"

1. What Is AI Research Ethics?

Concept and scope
AI research ethics refers to the values and procedures to uphold throughout AI research and development. Its scope runs from data collection and model design through experiments, paper writing, deployment, and post-deployment management.

Relationship to general research ethics
The basic framework of existing research ethics applies to AI as it is: no fabrication, falsification, or plagiarism; protection of research participants; honest reporting. AI adds its own characteristics:

Data is large-scale, and its provenance is complicated.
It can be hard to explain why a result came out the way it did.
A model, once built, may be used in unexpected ways across several services.

Why review is needed
Experiments that once stayed in the lab now reach society quickly. The earlier a problem is examined, the cheaper it is to address; fixing things after the fact costs more. This is not an opinion so much as a widely accepted rule of thumb in software development.

2. Core Ethical Principles in AI Research

AI principles put forward by international bodies and national institutions differ in wording but share a lot in common. Representative examples are UNESCO's Recommendation on the Ethics of Artificial Intelligence (2021) and the OECD AI Principles. Check each organization's latest documents for details.

Principle	Key question	Example
Human dignity and autonomy	Do people keep final decision-making authority?	A hiring AI's score is used only as a reference; a person decides
Fairness	Does it avoid disadvantaging particular groups?	Comparing error rates by group
Transparency and explainability	Can we explain how it works?	Writing a model card
Privacy and data protection	Does it avoid collecting or retaining more than necessary?	Anonymization, access control
Safety and reliability	Can harm from malfunction or misuse be reduced?	Stress testing
Responsibility and accountability	Who is responsible when something goes wrong?	Named owners and logs
Reproducibility and research integrity	Can others verify it?	Publishing code and settings

A caution: principles can conflict. Publishing data to increase transparency can collide with privacy. Increasing explainability can sometimes lower performance. So ethics is less about finding the one right answer and more about striking a balance and recording why.

3. AI Training Data and Ethics

Problems in the collection process

The purpose of collection differs from the actual use
Data subjects don't know their data is used for training
The source of data scraped in bulk from the web is hard to trace

Personal and sensitive information
Health, beliefs, biometrics, and similar information need especially strict protection. Anonymization does not guarantee safety, because combining several pieces of information can make an individual identifiable again. The applicable law (Korea's Personal Information Protection Act, or GDPR abroad, for example) varies by research subject and region, so it is safest to check with your institution's privacy office or review board.

Bias and representativeness
If data from a certain region, age group, or language is overrepresented, the model optimizes for that group. But data bias is only one cause among several. Bias can also arise from how the problem is defined, labeling criteria, and evaluation methods.

Source and permission to use
"Publicly available" is not the same as "free to use." Check licenses, terms of service, and any restrictions on collection methods (for example, conditions related to crawling). Copyright exceptions and limitations (such as for research or information analysis) differ by country, so they cannot be generalized.

Synthetic and public data

Synthetic data can reduce privacy risk. However, it can inherit the biases of the original data or partially reproduce original information, so it cannot be called "completely safe."
For public datasets, check the collectors' conditions and limitations in the documentation (such as datasheets).

4. Ethical Issues in Generative AI Research

Generative AI is AI that produces text, images, and code.

False or distorted information: It can produce content that sounds plausible but is untrue (commonly called "hallucination"). Cases of made-up references that don't exist have also been reported.
Biased content: Stereotypes in training data can show up in outputs.
Copyright and usage rights: The use of data for training and the rights in outputs are still being debated and decided country by country, and no single conclusion has settled (needs verification).
Personal information exposure: Research is examining the possibility that a model reproduces parts of its training data verbatim. The degree of risk varies by model and settings.

Why researchers must verify
Generated output is a candidate, not evidence. Citations, figures, and factual claims must be checked against the originals.

Transparency when using results as research output
It is advisable to state where and how AI was used (organizing ideas, editing prose, generating code, analysis, and so on). Policies differ by conference and journal, so check the relevant rules before submitting. AI cannot bear responsibility, so responsibility as author rests with the person. Many academic institutions share this position, but exceptions may exist.

5. AI Research and Copyright

When using others' works

Quotation should be limited to what is necessary, with the source stated, and your own writing should be the main body.
Figures, tables, and data can also be copyrighted works.

Checking sources and licenses
Even among things labeled "open," conditions differ. Creative Commons licenses, for example, come with varied conditions such as attribution, limits on commercial use, and no derivatives. Using material without reading the conditions can be a violation.

Why write original work
Copying someone else's sentences, or merely reordering their words, raises plagiarism and can raise copyright infringement issues. Above all, it offers readers no new value. The principle is to understand the material, restate it in your own words, and cite the source.

Practical steps when producing content with AI

Check the tool's terms of service for permitted use of outputs (commercial use, attribution).
Check whether the output is excessively similar to existing works, using plagiarism checkers or searches.
Record the materials and tools used, and the dates.
If in doubt, consult your institution or a legal professional.

6. Bias and Fairness in AI Research

Effect of data bias
When a group has little data, errors for that group can increase. But it can't be simplified to "biased data → discrimination necessarily follows." The context of use also plays a part.

Hypothetical cases where results may disadvantage a group

A voice assistant that recognizes speech with certain accents poorly
A résumé-screening tool trained on past hiring records that repeats past practices
An image analysis tool whose accuracy varies with skin tone

Detecting and mitigating

Measure performance (accuracy, error rate) separately for each group.
Review the composition of the data and supplement under-represented groups.
Apply mitigation techniques at the training or post-processing stage.
Have people from diverse backgrounds review the results.

Cautions in interpretation
There is no single definition of fairness. It is known that different fairness metrics can be hard to satisfy fully at the same time in some cases, so you should state which standard you chose and why. Also, an improved bias metric does not justify concluding "it is fair."

7. Transparency and Reproducibility in AI Research

Why record
Results are trustworthy only if the same experiment, run again, yields the same conclusion. Records also let you trace the cause when there's an error. That said, personal data, security, or trade secrets sometimes make full disclosure impossible, so record "to the extent possible" and state why anything is withheld.

How to document

Data: source, collection period, preprocessing, limitations (datasheet)
Model: purpose, performance, limitations, situations where it should not be used (model card)
Experiments: hyperparameters (training settings), random seeds, code version
Evaluation: metrics, data splits, baselines

Elements of a reproducible environment
Fix or record software versions, hardware, and random-number handling. Note also that results may differ slightly even with the same settings because of hardware differences.

8. The Responsibility of AI Researchers

Reviewing social impact: Better performance is not automatically a social benefit. Look at who gains and who bears the burden.
Finding and recording risks early: In the early stage of research, write down "what happens if this is misused?" as a risk list, and update it along with mitigation plans. Dual-use problems, where a technology built with good intent can be abused, are handled here too.
Disclosing errors and limitations: Stating failed experiments, conditions of low performance, and the range of applicability raises trust in the research.
Considering societal use: Anticipate uses beyond the research purpose and design usage conditions, licenses, or access limits accordingly. This kind of control has limits, however, and there is a counterview that excessive restriction can dampen the sharing of research.

9. An Ethics Checklist for Real AI Research

Is the research purpose clear, and do you use only the data that purpose requires?
Have you confirmed the source and usage rights of the data (license, terms, consent)?
Have you applied privacy protections (minimal collection, anonymization, access control)?
Have you checked data and model bias by group?
Have you verified performance and limitations together?
Have you avoided overstating the results?
Have you properly cited sources and references and disclosed whether AI was used?
Have you secured reproducibility and verifiability?
Have you examined and recorded expected social risks and side effects?

10. The Future of AI Research Ethics

Why it matters more
As AI enters areas close to daily life, such as healthcare, hiring, and education, the cost of errors grows. There is broad consensus on this point.

Division of roles

Researchers and developers: record risks and disclose limitations.
Companies: set up internal review systems and clear lines of responsibility.
Governments: set minimum standards and provide support.
Users: read results critically and raise problems.

Policy trends (needs verification)

In Korea, the Framework Act on the Development of Artificial Intelligence and Establishment of Trust (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법) is confirmed as having taken effect on January 22, 2026. The government is reported to have said it will provide a grace period, but detailed obligations and timing may change, so check the responsible ministry's latest notices.
The EU AI Act applies in stages. There are reports and analyses that the schedule for high-risk AI has been adjusted (postponed), but the exact dates and scope are safest confirmed from official EU documents.

Can innovation and responsibility go together?
One view sees ethics as an obstacle that slows things down; another sees it as a foundation for long-term trust. This article leans toward the latter, but concern that excessive procedure can dampen research is real too. What matters is less adding rules than making a level of review proportionate to the research purpose into a habit.

Conclusion

AI research ethics is not a declaration meant only for special people. It begins with small habits: confirming where data came from, recording limitations, and verifying results. Before your next project, run through the checklist above once.