⚖️ Research Ethics in Computer Science

“Science advances on trust. Every violation of research integrity corrodes the foundation of the entire enterprise.”

Research ethics in computer science spans a broad terrain: from plagiarism and data fabrication, to privacy violations in human-subjects research, to the societal impact of deployed AI systems. Understanding these norms is not optional — violations can end careers, trigger paper retractions, and cause real-world harm.

1. Academic Integrity Fundamentals

What Constitutes Misconduct

Type	Definition	Examples
Fabrication	Inventing data or results that were never collected	Reporting benchmark scores that were not actually run
Falsification	Manipulating data or experiments to misrepresent results	Selectively deleting unfavorable data points; cherry-picking random seeds
Plagiarism	Using another’s ideas, text, or code without attribution	Copying related work text from another paper; using unreferenced code
Duplicate Publication	Submitting substantially identical work to multiple venues simultaneously	Submitting the same paper to ICML and NeurIPS at the same time
Gift Authorship	Including authors who did not meaningfully contribute	Adding an advisor’s name to a paper they had no role in
Ghost Authorship	Excluding contributors who should be listed as authors	Omitting a collaborator who wrote the methodology section

[!CAUTION] These are not abstract rules. Fabrication and falsification are career-ending, and in some contexts (medical AI, safety-critical systems), they can cause direct physical harm. Retracted papers remain permanently marked in databases.

The Reproducibility Standard

A core ethical obligation in computational research is reproducibility: your results must be achievable by an independent researcher following your description. This requires:

Releasing code and model weights used to produce results.
Reporting all hyperparameters used in experiments.
Disclosing negative results and failed runs, not just the best run.
Specifying hardware, software, and dataset versions precisely.

See research/reproducibility.md for a full reproducibility framework.

2. Authorship and Credit

The CRediT Taxonomy

The Contributor Roles Taxonomy (CRediT) defines 14 distinct roles in research:

Conceptualization
Data Curation
Formal Analysis
Funding Acquisition
Investigation
Methodology
Project Administration
Resources
Software
Supervision
Validation
Visualization
Writing – Original Draft
Writing – Review & Editing

Most CS conferences and journals now support (and increasingly require) CRediT statements in submitted papers. Use them honestly.

Authorship Order Conventions in CS

Convention	Field	Meaning
First author	Universally	Primary contributor (usually the PhD student)
Last author	ML/AI/Systems	Senior researcher / PI / advisor
Alphabetical	Theory / Math	Equal contribution; no seniority implied
Co-first authorship	Common in biology	Explicitly denoted with † symbol; equal primary contributors

[!IMPORTANT] Discuss authorship and author order before the project begins, not after results are produced. Authorship disputes are among the most damaging conflicts in academic labs.

3. Human Subjects Research and IRB

Any research involving human participants — including user studies, surveys, interviews, dataset collection involving people, or analysis of personal data — is subject to ethical oversight.

When IRB Review Is Required

User studies with human participants
Analysis of data collected from people (e.g., scraped social media data, medical records)
Crowdsourced labeling tasks (e.g., Amazon Mechanical Turk)
Deployment of systems that interact with people in research contexts

IRB Process Overview

Submit a protocol describing your study: purpose, participants, data collection, privacy protections, risks, and benefits.
Receive approval (or exemption) before beginning data collection.
Obtain informed consent from participants.
Store data securely and anonymize where possible.
Report any adverse events to the IRB immediately.

[!WARNING] Publishing results from human-subjects research conducted without IRB approval is grounds for paper retraction and professional sanctions, regardless of the research quality.

4. Data Ethics

Responsible Dataset Collection

Concern	Guidance
Consent	Collect data only with explicit informed consent or clear public domain status
Privacy	Anonymize personally identifiable information (PII) before storage or release
Bias	Document demographic composition of datasets and known biases explicitly
Copyright	Respect intellectual property in training datasets (especially for generative models)
Provenance	Document data sources, collection methods, and preprocessing steps

The Datasheet for Datasets

Google Research introduced Datasheets for Datasets — a standardized documentation format that captures:

Motivation and intended use
Composition and collection process
Preprocessing and cleaning steps
Distribution, maintenance, and legal considerations

Publishing a datasheet alongside your dataset has become best practice.

5. AI Ethics and Responsible Research

Dual-Use Research Concerns

Many CS research contributions have dual-use potential — they can be used for beneficial or harmful purposes. Examples:

Face recognition → security applications OR surveillance abuse
Language models → writing assistance OR disinformation generation
Adversarial attacks → robustness research OR real-world attacks on deployed systems
Network traffic analysis → intrusion detection OR mass surveillance

[!NOTE] Most major venues (NeurIPS, ICML, ICLR, FAccT) now require a Broader Impacts or Ethics Statement in submitted papers. This is your opportunity to discuss potential harms and mitigation strategies honestly — not to minimize concerns.

The ACM Code of Ethics

The ACM Code of Ethics and Professional Conduct is the foundational ethical framework for computer scientists. Core principles:

Contribute to society and human well-being.
Avoid harm.
Be honest and trustworthy.
Be fair and take action not to discriminate.
Respect the work required to produce new ideas.
Respect privacy.
Honor confidentiality.

Fairness and Bias in Systems

Research involving machine learning must grapple with:

Algorithmic bias: Systems that discriminate against protected groups due to biased training data or model design.
Measurement bias: Metrics that do not capture the experience of underrepresented users.
Deployment context: A model accurate in aggregate may be harmful at the individual level.

Recommended frameworks and tools:

Fairlearn — Python toolkit for assessing and mitigating fairness issues.
AI Fairness 360 (AIF360) — IBM’s comprehensive fairness toolkit.
Model Cards — Standardized model documentation including fairness evaluation.

6. Open Science and Transparency

The Reproducibility Crisis

Meta-analyses across ML research have found that a significant fraction of published results cannot be reproduced by independent teams. Contributing causes:

Undisclosed hyperparameter tuning
Selective reporting of favorable random seeds
Missing implementation details
Dataset contamination (test set leakage)

Open Science Practices

Practice	Tool / Platform
Preprint publication	arXiv, OpenReview
Code release	GitHub, Papers with Code
Data release	Hugging Face Datasets, Zenodo, OSF
Model release	Hugging Face Hub
Pre-registration	OSF Preregistration — commit to methodology before seeing results

7. Citing and Attribution

What Must Be Cited

Direct quotations
Paraphrased ideas from other works
Datasets used in experiments
Code libraries and frameworks
Statistical methods not invented by you

Citation Tools

Zotero — Reference manager with browser plugin for one-click citation saving.
Better BibTeX — Generates consistent citation keys for LaTeX.
doi2bib — Convert DOI to BibTeX entry instantly.
Google Scholar — Cite button generates BibTeX for any paper.