Demonstrate understanding of ethical standards governing scholarly research in computer science.
Lectures 4 and 5 addressed one specific violation and its prevention. This lecture supplies the principles from which those rules derive — rules are derivative, principles are generative.
The four principles are examined as a system, not as four separate definitions. Section 7 shows how a single act implicates all four, and supplies the five-test workflow that is the portable takeaway of this lecture.
Section 3.3 (honest error versus misconduct) is the most frequently misunderstood material in the unit: students assume any wrong result is misconduct, which is both false and corrosive to the culture of honest correction.
1. Learning Objectives
- Articulate the meaning and operational scope of each of the four foundational principles.
- Identify concrete ethical dilemmas that computer scientists face in applying these principles.
- Distinguish between honest error and research misconduct.
- Recognise forms of intellectual property — ideas, text, code, data, models — that require respect and attribution.
- Apply ethical reasoning to ambiguous, real-world scenarios.
2. Why Ethics Binds Computing Acutely
Research ethics provides the moral and professional compass that guides scholars in the creation, evaluation and dissemination of knowledge. In computer science — a discipline whose outputs increasingly shape commerce, governance, communication, healthcare and warfare — ethical conduct carries consequences that extend well beyond the laboratory or the publication page.
A single fabricated algorithm, a plagiarised model or a falsified benchmark can corrupt downstream engineering, mislead policymakers and erode public trust in the scientific enterprise. The propagation chain is concrete: an inflated accuracy claim becomes a cited baseline, then a procurement criterion, then a deployed system making decisions about people.
Four foundational principles anchor ethical practice: honesty, objectivity, integrity and respect for intellectual property. These are not abstract ideals: they are operationalised daily, from the design of an experiment to the writing of a paper, the reviewing of a peer's manuscript and the assignment of credit on a multi-author codebase.
- Belmont Report (1979).
- Singapore Statement on Research Integrity (2010).
- European Code of Conduct for Research Integrity (ALLEA, revised 2023).
- ACM Code of Ethics and Professional Conduct (2018) — the document you are professionally bound by as a computing practitioner.
3. Principle I — Honesty
3.1 Definition and Scope
Honesty in research is the truthful representation of methods, procedures, data, results and contributions. It is the bedrock upon which the credibility of scientific knowledge rests.
Honesty obliges the researcher to:
- Report results as observed, including negative or unexpected findings.
- Disclose methodologies faithfully, including their limitations.
- Avoid fabrication — inventing data, results or experiments.
- Avoid falsification — manipulating research materials, equipment or processes, or altering or omitting data.
- Report errors promptly and correct the record when mistakes are discovered.
3.2 Why Honesty Matters in Computer Science
Computer science research is largely driven by artefacts — software, hardware, datasets, benchmarks and trained models. The reproducibility of these artefacts depends on honest reporting.
- Benchmark performance. Selective reporting of metrics — reporting only the best of many runs — artificially inflates claimed performance gains. This constitutes selective honesty: technically true statements that omit context and mislead readers.
- Algorithm novelty. Claiming an algorithm is “novel” when it is essentially a renamed prior method is dishonest even though it is non-trivial to detect.
- Data integrity. Reporting on a cleaned subset of a dataset without disclosing exclusions can lead reviewers and readers to draw invalid conclusions.
Reporting a single best run without the variance across random seeds is selective honesty. Every sentence is true, and the report still misleads.
3.3 Honest Error versus Misconduct
Not every wrong result is misconduct. Science advances by correcting honest error. The distinction rests on intent and transparency.
| 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 |
The FFP framework — fabrication, falsification, plagiarism — is the basis of research-integrity policy worldwide. Honesty is the positive counterpart that prevents FFP in the first place. Note that reckless sits on the misconduct side: negligence severe enough to disregard obvious checks is not excused by absence of intent to deceive.
Response to discovery is the decisive dimension in practice: how a researcher behaves in the hour after finding their own error is the clearest available signal of character. Admitting an error is a professional act, not a confession — a culture that punishes correction manufactures concealment.
3.4 Operational Practices
- Maintain a tamper-evident research log — a version-controlled lab notebook. This is the cheapest available protection, because it produces exactly the documentation that distinguishes error from misconduct if a claim is challenged.
- Preserve raw data with metadata indefinitely where feasible.
- Adopt pre-registration for confirmatory studies.
- Uphold double-blind review norms, which deflate the incentive for selective reporting.
4. Principle II — Objectivity
4.1 Definition and Scope
Objectivity is the commitment to let evidence — not personal interest, ideology, prior commitment or external pressure — drive the design, analysis and interpretation of research. It is the discipline of allowing the data to speak, even when the message is unwelcome.
- Design experiments and evaluations that can in principle falsify a hypothesis.
- Avoid conflicts of interest, or disclose them when unavoidable.
- Distinguish empirical observation from interpretation.
- Submit claims to peer scrutiny rather than defend them rhetorically.
4.2 The CS-Specific Illusion, and Three Pressure Points
Computer science is often perceived as “objective” because it deals with formal systems. Yet every layer of an empirical CS study — from dataset choice, to hyperparameter selection, to evaluation metric — is shaped by human judgement. Objectivity is a regulative ideal, requiring continuous effort, not a property conferred by the subject matter.
- Confirmation bias in algorithm design. A researcher emotionally invested in a proposed algorithm may unconsciously design benchmarks that favour it. Robust objectivity requires running baselines under matched conditions, using standard suites, and reporting performance across the full distribution of inputs rather than cherry-picked cases.
- Conflicts of interest. Industry funding, equity stakes, consulting relationships and personal relationships can compromise the disinterested stance objectivity requires. Disclosure is necessary but never fully neutralising; readers should interpret funded work accordingly.
- Significance hunting. Although computing research rarely uses null-hypothesis significance testing in the biomedical sense, analogous risks exist. Tweaking architectures, hyperparameters, random seeds or preprocessing pipelines until a result “works” corrodes objectivity even when each individual technique is conventional.
4.3 Objectivity as Technical Competence
Objectivity is, in part, a competence: a researcher who cannot quantify uncertainty cannot report it honestly, however good their intentions. Consider a claim that a new compiler optimisation yields a mean speedup of 12 per cent over a baseline. Without confidence intervals, standard deviations, hardware variance and sample-size justification, the claim is technically a number but not an empirical finding.
The first expression gives effect size; the second gives uncertainty. Effect size without uncertainty overstates confidence; uncertainty without effect size obscures practical relevance. Together they allow readers to assess the claim on its merits rather than on the author's enthusiasm.
4.4 Operational Practices
- Use pre-registered experimental protocols where appropriate.
- Apply blinding in evaluation: whoever runs the benchmarks should not know which system is the proposed one.
- Disclose all conflicts of interest explicitly.
- Encourage adversarial collaboration — design the experiment a sceptic would accept, before running it.
5. Principle III — Integrity
5.1 Definition and Scope
Integrity is the consistent and coherent application of ethical principles across the entire research lifecycle. Where honesty pertains to specific statements and data, integrity pertains to the character and consistency of the researcher.
A researcher of integrity acts ethically even when no one is watching, when peer pressure pushes the other way, and when expedience tempts compromise. Integrity subsumes:
- Adherence to institutional and professional standards.
- Consistency between what one says publicly and what one does privately.
- Willingness to take ethical stands that may be costly — retracting one's own work, or reporting a colleague's misconduct.
- The disposition to ask “Is this the right thing to do?” before every significant research decision.
Honesty can be satisfied statement by statement; integrity cannot be satisfied piecemeal, because it is the property of a pattern of conduct — which is why the frequency row in Section 3.3 matters. Integrity is also only observable under pressure: conduct that costs nothing tests nothing.
5.2 Integrity in Multi-Author Work
Modern CS research is collaborative, so integrity must scale across teams, institutions and nations.
- Authorship attribution. The CRediT (Contributor Roles Taxonomy) provides fourteen roles: conceptualisation, methodology, software, validation, formal analysis, investigation, data curation, writing – original draft, writing – review and editing, visualisation, supervision, project administration, funding acquisition, resources. Misattribution of these roles is an integrity violation.
- Gift and ghost authorship. Including someone who did not meaningfully contribute, or omitting a contributor whose work materially shaped the paper, are both breaches. The second is the one that harms students.
- Order of authorship. The CS convention of alphabetical ordering sometimes masks the actual contribution order. Integrity requires transparent disclosure when the alphabet misrepresents the work distribution.
Agreeing roles in writing before submission prevents nearly every authorship dispute that reaches a supervisor.
5.3 Integrity in Mentoring and Supervision
Computer science research is heavily mentored, and an advisor's integrity directly shapes a student's professional formation. Practices that compromise integrity include:
- Pressuring students to keep the lab's “story” consistent in review responses.
- Accepting authorship in exchange for funding or lab resources.
- Failing to credit students whose code, datasets or experiments populate papers.
The power asymmetry here is real rather than symmetric: students are the party structurally exposed to both ghost authorship and mentorship failure.
5.4 Whistleblowing and the Costs of Integrity
Whistleblowers often pay a price — retaliation, social ostracism, stalled careers — yet whistleblowing remains essential to the integrity of the scientific system. Institutional ombudspersons, confidential reporting channels and protective policies such as the U.S. Whistleblower Protection Act and EU Directive 2019/1937 exist to mitigate these costs. B.Tech graduates entering industry research may one day face such choices.
6. Principle IV — Respect for Intellectual Property
6.1 Definition and Scope
Respect for intellectual property acknowledges that ideas, expressions, methods, data and code have authors, and that these authors have legitimate claims over attribution and, in some cases, over use and commercialisation. It bridges ethics and law — and it extends beyond what the law strictly requires.
6.2 Five Categories in CS Research
- Textual IP. Verbatim plagiarism (copying without quotation marks or citation) and idea plagiarism (paraphrasing an argument so closely that structure and substance remain the source's). Computer science papers are densely argumentative, so the threshold for attribution is lower than in many other disciplines: even a single lifted sentence is misconduct.
- Software and code IP. Do not copy open-source code without honouring licence terms — GPL, MIT, Apache and BSD impose materially different obligations. Do not re-implement a protected expressive structure after reading another team's code. Cite the software libraries used in experiments, and deposit your own code where possible under a permissive licence.
- Data IP. Cite dataset creators; respect licences (Creative Commons variants, dataset-specific terms); do not redistribute restricted-access datasets through side channels; respect the consent terms under which human-subjects data were collected.
- Algorithmic and methodological IP. Even unpatented algorithms have authors. The community convention is eponymous citation: RSA (Rivest–Shamir–Adleman), Dijkstra's algorithm, the Adam optimizer (Kingma and Ba), the Transformer (Vaswani et al.), ResNet (He et al.).
- Trademark and brand IP. Naming a system after a competitor's trademarked product, or using screenshots of proprietary software beyond fair-use bounds, can infringe trademark and copyright — and even where legally defensible may be unprofessional.
6.3 The Legal–Ethical Boundary
IP law varies by jurisdiction and is broader than research ethics in some respects (patent law) and narrower in others: many forms of idea plagiarism are legal but unethical.
| Act | Legal status | Ethical status |
|---|---|---|
| Verbatim copying of text without citation | Generally not illegal in small amounts (fair use) | Unethical |
| Re-implementing an algorithm from 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 |
The ethical principle of respect for IP operates independently of, and often more strictly than, the legal minimum. “It is legal” is never a sufficient answer to an ethical question. Legal compliance is assessed by counsel after the fact; ethical compliance is assessed by the community continuously, and reputational sanction does not require a court.
Operational practices: track every idea, method and quotation in a citation manager (Zotero, Mendeley, EndNote); run submitted papers through plagiarism-detection software before submission; maintain a licence-awareness checklist for every dependency used in research code; and when paraphrasing, ensure the structure as well as the wording is your own.
7. How the Four Principles Interact
The four principles are not independent. A single act of conduct can implicate all four:
- Fabricating data is dishonest; it lacks objectivity in the sense that evidence is decoupled from inference; it reflects a failure of integrity; and it disrespects the community whose trust is being betrayed.
- Plagiarising disrespects IP, is dishonest about authorship contribution, compromises the integrity of the scholarly record, and skews the objectivity of the literature by inflating apparent support for claims.
- Honesty test — can I state this plainly, including what is less flattering?
- Objectivity test — has the evidence, not my preference, driven my conclusion?
- Integrity test — would I be comfortable if this decision were reported in the campus newspaper?
- IP-respect test — have I given proper credit to sources, code, data and methods?
- Public-reason test — can I justify this decision in a forum of respected peers?
If all five pass, the decision is almost certainly ethical. If any fails, further reflection — or consultation with an advisor or ethics officer — is warranted. A failed test is a prompt to consult, not a verdict of guilt.
8. Case Studies
| Scenario | Principle implicated | Verdict and best practice |
|---|---|---|
| Trimmed benchmarks. A proposed architecture underperforms three baselines; after re-runs and tuning it wins; only the final results are reported | Honesty, partly Integrity | Misconduct. Report the full set of runs, document the hyperparameter 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 verbatim with marks |
| The ghost in the code. A student implements and evaluates; the advisor adds a paragraph, places both names on the byline, and says authorship will be sorted out later | Integrity | Agree authorship before submission under a documented policy consistent with the venue's standards |
| The open-source re-use trap. A GPL-licensed library is forked, modified and used inside a closed-source commercial product | Respect for IP, Integrity | Derivative work. High legal exposure and a clear ethical breach. Maintain a dependency licence inventory |
| The convenient control experiment. A known, recently published open-source baseline that outperforms the system is silently omitted | Objectivity | High severity. Disclose all known relevant comparators and explain any exclusion |
In four of these five cases the offending act is an omission, not a falsehood. This is why the honesty test asks specifically about what is less flattering. Students expect misconduct to look like lying; most of it looks like things left unsaid.
9. Summary
- Honesty demands truthful reporting of methods, data and findings; it is the antidote to fabrication, falsification and selective reporting.
- Objectivity demands that evidence rather than interest drives inference; it requires management of conflicts of interest, avoidance of confirmation bias, and methodological rigour.
- Integrity demands consistency of ethical conduct across the entire research lifecycle, including authorship, mentorship and the willingness to defend the truth when it is costly.
- Respect for intellectual property demands attribution of ideas, text, code, data and methods, together with compliance with the licences under which they are shared.
- These principles operate together: weakness in one weakens the system as a whole. In computer science, adherence is constitutive of the discipline's claim to scientific status.
10. Exam Preparation
- Honest error versus misconduct — intent, response, documentation, frequency.
- Fabrication — Falsification — Plagiarism (FFP).
- Selective reporting versus negative-results reporting.
- Objectivity versus confirmation bias versus conflict of interest.
- Integrity versus honesty — character against specific statement.
- Authorship: gift, ghost, alphabetical, CRediT roles.
- IP categories: text, code, data, algorithms, trademarks.
- Licence compliance: GPL, MIT, Apache, BSD, CC variants.
- The five-test diagnostic workflow.
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. Answers that merely narrate the scenario lose most of the available marks.
10.1 Short-Answer Questions
- Define research integrity and distinguish it from research honesty, with one example from computer science. 3 marks
- Explain why selective reporting of benchmark results violates honesty even when no data are fabricated. 3 marks
- Distinguish idea plagiarism from verbatim plagiarism and provide one example relevant to a literature review. 3 marks
- Describe the role of the CRediT taxonomy in supporting integrity in multi-author work. 2 marks
- Why is the legal–ethical boundary important when discussing respect for intellectual property? Provide one example. 3 marks
- State the five-test diagnostic workflow and apply it to a scenario of your choosing. 5 marks
10.2 Long-Answer Questions
- “Objectivity is a regulative ideal, not a state.” Discuss with reference to conflicts of interest, confirmation bias and benchmark design in contemporary machine learning research. 10 marks
- Analyse a case in which each of the four foundational principles is implicated. Discuss the institutional mechanisms that can catch or correct such a violation. 10 marks
- Compare the protection of textual IP, software IP and data IP in computer science research. What licence-attribution practices should a B.Tech project adopt? 10 marks
10.3 Applied Scenarios
- You are a third-year B.Tech student co-authoring a paper with a peer who has removed your name from the byline because they disagree with your recent methodological criticism. How do you proceed, and which principles are at stake?
- Your advisor pressures 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.
- You find an elegant open-source implementation that does almost what you need. Compare three options: (a) fork and modify, citing the source; (b) reimplement from scratch using only the paper's description; (c) use as-is and relicense. Discuss the IP and integrity implications of each.
11. Further Reading
- Singapore Statement on Research Integrity (2010).
- European Code of Conduct for Research Integrity, revised edition (ALLEA, 2023).
- ACM Code of Ethics and Professional Conduct (2018).
- Resnik, D. B. What Is Ethics in Research & Why Is It Important? National Institute of Environmental Health Sciences.
- Steneck, N. H. Introduction to the Responsible Conduct of Research. Office of Research Integrity, U.S. Department of Health and Human Services.
- Pineau, J., et al. (2021). Improving Reproducibility in Machine Learning Research: A Report from the NeurIPS 2019 Reproducibility Program. Journal of Machine Learning Research.