CSEG3060 · Research Methodology in Computer Science
Honesty, objectivity, integrity and respect for intellectual property
A fabricated benchmark does not stay in the paper. It ships.
Unit III · Lecture 6 of 8 · 60 minutes · CO3
Dr. Mohsin Furkh Dar · School of Computer Science, UPES
→ / Space next · ← back · O overview · N speaker notes · F full screen
Learning objectives · CO3
Articulate the meaning and operational scope of each of the four principles
Identify concrete ethical dilemmas computer scientists face
Distinguish honest error from research misconduct
Recognise the forms of IP — ideas, text, code, data, models — requiring attribution
Apply ethical reasoning to ambiguous, real-world scenarios
Context
Computing outputs shape commerce, governance, communication, healthcare and warfare — so failures propagate beyond the laboratory and the publication page.
A fabricated algorithm, a plagiarised model or a falsified benchmark can corrupt downstream engineering, mislead policymakers and erode public trust in the scientific enterprise.
Truthful representation of methods, data and contributions
Evidence, not interest, drives inference
Consistency of conduct across the whole lifecycle
Attribution of ideas, text, code, data and methods
Codified in the Belmont Report (1979) · Singapore Statement (2010) · European Code of Conduct for Research Integrity (ALLEA, 2023) · ACM Code of Ethics (2018).
Principle I
Reporting only the best of many runs artificially inflates claimed gains. This is selective honesty — technically true statements that omit context and mislead.
Claiming an algorithm is “novel” when it is a renamed prior method · reporting on a cleaned subset without disclosing exclusions.
CS research is driven by software, hardware, datasets, benchmarks and trained models — and the reproducibility of every one of them depends on honest reporting.
Principle I · continued
| Dimension | Honest error | Misconduct |
|---|---|---|
| Intent | Unintentional, accidental | Willful or reckless |
| Response to discovery | Open acknowledgment, correction | Concealment, denial |
| Documentation | Reproducible records of process | Inconsistent, missing |
| Frequency | Sparse and explainable | A pattern of irregularities |
Research-integrity policy worldwide centres on Fabrication, Falsification, Plagiarism. Honesty is the positive disposition that prevents all three.
Principle II
Computing feels objective because it handles formal systems. Yet every layer of an empirical CS study — dataset choice, hyperparameter selection, evaluation metric — is shaped by human judgement. Objectivity requires continuous effort.
Emotional investment in a proposed method leads to benchmarks that unconsciously favour it.
Remedy: matched baseline conditions, standard suites, report across the full input distribution.
Industry funding, equity, consulting and personal relationships compromise disinterest.
Remedy: disclosure — necessary but never fully neutralising.
Tweaking architectures, hyperparameters, seeds or preprocessing until a result “works”.
Remedy: pre-registration, blinding, adversarial collaboration.
Every individual action in the significance-hunting loop is standard practice — which is exactly why the aggregate is so hard to see and so easy to rationalise.
Principle II · continued
“12% faster.”
Without confidence intervals, standard deviations, hardware variance and sample-size justification, this is a number but not an empirical finding.
Effect size — how large is the difference?
Uncertainty — how well is it pinned down?
Objectivity is partly a technical competence, not only a disposition — a researcher who cannot quantify uncertainty cannot report it honestly, however good their intentions.
Principle III
Pertains to specific statements and data. Can be satisfied statement by statement.
Pertains to the character and consistency of the researcher across the whole research lifecycle. Cannot be satisfied piecemeal — it is the property of a pattern.
Principle III · continued
CRediT — 14 contributor roles: conceptualisation · methodology · software · validation · formal analysis · investigation · data curation · writing – original draft · writing – review and editing · visualisation · supervision · project administration · funding acquisition · resources
Gift authorship — including someone who did not meaningfully contribute.
Ghost authorship — omitting a contributor whose work materially shaped the paper. The second is the one that harms students.
The CS convention of alphabetical ordering sometimes masks the actual contribution distribution. Integrity requires transparent disclosure when the alphabet misrepresents the work.
Whistleblowers face retaliation, ostracism and stalled careers, yet the practice is essential. Ombudspersons, confidential channels and protective law — U.S. Whistleblower Protection Act, EU Directive 2019/1937 — mitigate the cost.
Principle IV
Verbatim and idea plagiarism. CS papers are densely argumentative, so the threshold for attribution is lower than in many disciplines: even one lifted sentence is misconduct.
Honour licence terms — GPL, MIT, Apache, BSD differ materially · do not re-implement a protected expressive structure after reading another team's code · cite the libraries you use · deposit your own code where possible.
Cite dataset creators · respect licences (Creative Commons, dataset-specific terms) · never redistribute restricted data through side channels · respect the consent terms under which human-subjects data were collected.
Unpatented algorithms still have authors. The community convention is eponymous citation: RSA, Dijkstra's algorithm, Adam (Kingma & Ba), Transformer (Vaswani et al.), ResNet (He et al.).
Naming a system after a competitor's trademark, or using proprietary screenshots beyond fair use, risks infringement — and is in any case unprofessional.
Principle IV · continued
| Act | Legal status | Ethical status |
|---|---|---|
| Verbatim copying of text without citation | Generally not illegal in small amounts (fair use) | Unethical |
| Re-implementing an algorithm after reading prior code | Generally legal | Often unethical |
| Using a copyrighted figure in a review paper | Often permitted under fair use | Still requires citation |
| Falsifying results | Not typically illegal | Unethical; sometimes illegal (e.g. federal grants) |
| Stockpiling research data for years | Generally legal | May be unethical if it blocks community progress |
IP law is broader than research ethics in some respects (patent law) and narrower in others — many forms of idea plagiarism are perfectly legal and clearly unethical.
Synthesis
Fabricating data is dishonest, decouples inference from evidence (objectivity), reflects a failure of character (integrity), and betrays the community's trust.
Plagiarism disrespects IP, is dishonest about authorship, corrupts the integrity of the record, and skews objectivity by inflating apparent support for a claim.
Case studies
| Scenario | Principle | Verdict and best practice |
|---|---|---|
| Trimmed benchmarks — architecture loses to three baselines; after re-runs and tuning it wins; only final results reported | Honesty, partly Integrity | Misconduct. Report all runs, document the search protocol, use statistical tests |
| The borrowed paragraph — a survey paragraph, synonyms swapped, no citation | Respect for IP | High severity. Reread, set aside, rewrite with citation — or quote with marks |
| The ghost in the code — student implements and evaluates; advisor adds a paragraph, both names on the byline, “sorted out later” | Integrity | Agree authorship before submission under a documented policy |
| The open-source trap — GPL library forked, modified, shipped in a closed-source product | Respect for IP, Integrity | Derivative work. High legal exposure. Maintain a dependency licence inventory |
| The convenient control — a known, recent, better-performing open-source baseline silently omitted | Objectivity | High severity. Disclose all known comparators and explain any exclusion |
Consolidation
Truthful reporting of methods, data and findings — the antidote to fabrication, falsification and selective reporting.
Evidence rather than interest drives inference — managing conflicts, resisting confirmation bias, methodological rigour.
Consistency across the whole lifecycle — authorship, mentorship, and defending truth at a cost.
Attribution of ideas, text, code, data and methods, with licence compliance.
A co-author removes your name from the byline because they disagree with your methodological criticism. How do you proceed, and which principles are at stake?
Your advisor presses you to report benchmark results from a partial dataset because “the full dataset will take too long”. Identify the principles and propose a constructive course of action.
An elegant open-source implementation does almost what you need. Compare (a) fork and modify with citation, (b) reimplement from the paper alone, (c) use as-is and relicense.
Exam preparation
Name the principle, then the specific obligation breached, then the remedy. Prefer the omission-detecting reading of a scenario — most examination cases turn on what was left undisclosed.
Singapore Statement (2010) · European Code of Conduct (ALLEA, 2023) · ACM Code of Ethics (2018) · Resnik, What Is Ethics in Research? · Steneck, Introduction to the Responsible Conduct of Research · Pineau et al., JMLR (2021).
Next — Lecture 7: Ethical Considerations in Research: informed consent, confidentiality, fabrication and falsification, authorship norms and conflict of interest.