AI tools are now woven into everyday research practice. They can draft text, summarise articles, generate interview questions, propose analytic categories, and produce outputs that look coherent enough to pass as competence.

That is precisely the risk.

Across many research and knowledge practices—especially in unequal, politically charged environments—the ethical failure is rarely “using AI”. The failure is treating fluent text as a substitute for accountable inquiry, and treating speed as a justification for weaker judgement.

Research is not a neutral exercise in knowledge accumulation. It is a situated, political, and ethical practice. AI can assist some tasks inside that practice, but it cannot hold the obligations that make research legitimate: historical awareness, power analysis, contextual fidelity, reflexive judgement, and responsibility to the people and realities implicated by what we produce.

Research is accountable inquiry, not fluent text

A simple boundary clarifies much of the current confusion:

AI can help with tasks. It cannot “do” research.

Research is a defensible chain: question → method → evidence → claim. When that chain is intact, readers can evaluate your reasoning, replicate key steps, and contest your conclusions. When the chain is broken, you do not have research—you have plausible text.

Ethical research requires more than technical correctness. It requires a stance on how knowledge is produced, by whom, for what purposes, and with what consequences. That stance becomes more—not less—important when tools can generate convincing outputs at scale.

A useful diagnostic question is:

•  Is this a claim, a description, or an interpretation?
  o    Description reports what is present.
  o    Interpretation argues what it means.
  o    Claim asserts what should be believed—and therefore must be anchored in method, evidence, and accountability.

AI can assist with organising descriptions. It cannot be accountable for interpretations or claims. That accountability sits with you.

Where AI misleads: coherence, confidence, and false consensus

AI’s most common failure mode is not “getting one fact wrong”. It is producing synthetic authority: a confident, coherent narrative that hides methodological gaps.

You have likely seen versions of this:

• a “perfect” literature synthesis that collapses disagreement into a single storyline;
• citations that look plausible but do not exist;
• an apparent consensus that no real scholarly community has actually reached;
• a tidy set of themes that erase context, contradiction, and power.

This is an ethical problem, not merely a technical one. When coherence replaces verification, the burden of error shifts outward—onto communities, institutions, or policy processes that may treat the output as credible. In politically charged contexts, that shift is rarely neutral: it tends to reproduce existing hierarchies of knowledge and legitimacy.

Build verification triggers into your workflow—moments where you must stop and check rather than continue producing. For example:

• Any specific statistic, legal claim, or institutional attribution.
• Any “everyone agrees” framing.
• Any summary that will be used to justify a decision affecting people’s lives.
• Any claim about marginalised communities that is not grounded in primary sources or lived expertise.

Integrity did not begin with AI—and AI amplifies old harms

It is tempting to treat AI ethics as a new domain. In practice, AI accelerates long-standing misconduct patterns:

• plagiarism and patchwriting,
• ghost authorship and blurred attribution,
• appropriation of community knowledge without reciprocity,
• “credit laundering” that erases who did the thinking and who bore the risk.

AI adds scale, speed, and plausible deniability. But “the tool did it” is ethically meaningless. Responsibility does not dissolve because the interface is convenient.

If you want one integrity anchor, use this:

If you cannot explain how a sentence was produced—and defend the evidence behind it—do not publish it.

Literature work with AI: mapping is not synthesis

Used carefully, AI can support scoping: identifying keywords, mapping debates, proposing reading sequences, or helping you structure what you intend to read. The line is crossed when AI output becomes a substitute for reading, evaluation, and responsible citation.

A disciplined workflow keeps three things distinct:

1. Mapping (what appears to be out there),
2. Reading (what the sources actually say),
3. Synthesis (your argued account of what the field supports, contests, and cannot conclude).

The danger is synthetic consensus: a flattened story that privileges dominant voices and erases disagreement—especially scholarship outside mainstream publishing systems, languages, and geographies. Citation is not neutral. It is part of how authority is built.

A minimal ethical practice is to ask:

• Who is missing from my reading list—and what does that absence do to my conclusions?
• Am I over-relying on the most visible sources because they are easier for AI to “recognise”?
• What would it take to broaden the evidence base beyond what is most legible to dominant institutions?
   

If your institution is integrating AI into research workflows, CTDC can support you to do so ethically—strengthening governance, verification, and accountability without undermining rigour.

CTDC Academy also has a forthcoming course on Research in the Age of AI, designed to translate these principles into practical methods and tools.