Claude Code on the ethics of LLM use

##AI

Compiled by Claude Code. Supporting doc for LLM use in work - experimenting with short guidelines. Prompts / process: see Claude Code chat about LLM ethics.

1. The Deeper Ethics of LLM Use: Structural Issues Behind the Research Debate

The question of whether and how to cite LLM use in research (covered below) sits inside a much larger set of ethical concerns. University disclosure policies tend to focus narrowly on attribution and academic integrity — but the decision to use an LLM at all implicates a researcher in a wider web of environmental, labour, intellectual property, and power-concentration issues. This section collates sources on those structural concerns and considers how they interact with the research-ethics debate.

Environmental Cost: Energy and Water

LLM use is not environmentally neutral. The computational demands of both training and inference are enormous and growing.

Implication for research ethics: If a university requires ethical review for research with significant environmental impact, the routine embedding of LLM tools into research workflows arguably deserves similar scrutiny — particularly at institutional scale.

Pirated and Copyrighted Training Data

The corpora on which major LLMs were trained include vast quantities of copyrighted material used without permission or compensation.

Implication for research ethics: Universities that insist on proper citation and intellectual property respect in their plagiarism policies face an uncomfortable tension: the tools they are integrating into research workflows were themselves built on unconsented use of others' intellectual property.

Labour Exploitation in the AI Supply Chain

The human labour behind LLMs extends well beyond engineering teams in San Francisco.

Implication for research ethics: Research ethics frameworks routinely ask whether a study's methods exploit vulnerable populations. The same question can be asked of the tools researchers use — if the instrument was produced through exploitative labour, does its use raise ethical concerns analogous to those in supply-chain ethics more broadly?

Centralised Control and the "Compute Divide"

LLM development is concentrated among a handful of firms, with consequences for who shapes knowledge production.

Implication for research ethics: When universities embed commercial LLMs into their research infrastructure, they deepen dependency on firms whose incentives are commercial, not educational — and whose market position gives them de facto influence over the conditions of knowledge production.

Epistemic Dependency: Who Controls What Counts as Knowledge?

Beyond practical dependency lies a deeper epistemological concern.

Implication for research ethics: The plagiarism debate (below) asks whether researchers properly cite their use of LLMs. The epistemic-dependency question asks something more fundamental: whether routine LLM use changes the nature of the knowledge researchers produce, and whether institutions have adequately reckoned with that possibility.

How These Structural Issues Connect to the Research-Use Debate

Most university LLM policies focus on a narrow set of questions: Did you disclose? Did you cite? Is the output reliable? These are important, but they leave the structural issues above largely unaddressed. A researcher can fully comply with disclosure requirements and still be:

None of this means LLMs should never be used in research. But it does suggest that the ethics of LLM use in academia cannot be reduced to citation practice — and that the current policy landscape has a significant blind spot. A fuller ethical framework would ask not just "did you disclose?" but "what are you participating in?"

Further Reading: Structural Ethics

2. LLM Use in Research: Is Uncited AI Use Plagiarism?

An initial claude-code-run survey of how universities, publishers, and researchers are grappling with whether uncited LLM use in research constitutes plagiarism — or something categorically different.

The Core Question

As LLMs become embedded in research workflows, institutions face a definitional puzzle: does submitting AI-assisted work without disclosure count as plagiarism? Traditional plagiarism assumes a human originator whose ideas or words are misattributed. LLM output complicates this — it has no single author, draws on vast training data, and can produce text that feels original to the user even when it closely echoes existing work. Researchers have noted an "illusion of originality" where users mistakenly believe AI-generated concepts are novel when they derive from training materials (Lemley & Ouellette, University of Chicago Law Review).

Where Universities Have Landed

The "Plagiarism" Camp

Several major institutions have explicitly expanded their definitions of plagiarism to encompass undisclosed AI use:

The "Instructor Discretion" Camp

Many elite US institutions — including Harvard, MIT, Stanford, Princeton, and Yale — operate "follow your instructor's policy" frameworks, allowing individual academics to set specific rules within broader institutional principles. This acknowledges that appropriate AI use varies by discipline and assessment type (Thesify policy update).

The Russell Group (UK) Principles

All 24 Russell Group universities adopted five shared principles in 2023, which emphasise AI literacy, ethical use, equal access, and upholding academic rigour — but leave detailed policy to individual institutions and departments. The principles recognise that "appropriate uses of generative AI tools are likely to differ between academic disciplines."

An Alternative Framing: Research Misconduct, Not Plagiarism

A strand of the scholarly literature argues that undisclosed AI use is better understood as research misconduct rather than plagiarism in the traditional sense. The argument runs:

Lemley and Ouellette's analysis in the Chicago Law Review draws a useful three-way distinction between copyright infringement (economic, legal), plagiarism (ethical, about attribution), and bad scholarly practice (rigour, familiarity with the literature). They argue these are routinely conflated in AI debates, and that each demands different remedies (Lemley & Ouellette).

What Publishers and Research Bodies Say

COPE (Committee on Publication Ethics)

COPE's position statement on authorship and AI tools is clear that AI cannot meet authorship criteria because it cannot take responsibility for submitted work, manage conflicts of interest, or handle copyright. Authors must disclose AI tool use in their Methods section and remain fully responsible for all content — including anything the AI produced. COPE's guidelines underpin policies at Springer Nature, Wiley, Elsevier, and other major publishers.

UK Research Integrity Office (UKRIO)

UKRIO's guidance on AI in research warns that using chatbots to obscure content origin creates serious plagiarism risks — the underlying ideas and wording still require proper attribution even when passed through an intermediary tool. They emphasise that "longstanding principles of good research practice can be applied to help navigate the challenges" and that researchers must fact-check every AI-generated statement.

ICLR 2026

The machine-learning conference ICLR has taken one of the strongest lines: papers making extensive undisclosed use of LLMs face desk rejection. Authors must state how LLMs were used both in the paper text and the submission form, and any falsehood or plagiarism produced by an AI is treated as a Code of Ethics violation.

APA (American Psychological Association)

The APA requires authors to explicitly document AI use in their methods sections and provide full AI outputs as supplemental materials.

The Emerging Consensus — and the Gap

A rough consensus is forming around several points:

  1. Disclosure is non-negotiable. Nearly all institutional and publisher policies now require transparent reporting of AI use.
  2. AI cannot be an author. Following COPE and ICMJE guidelines, AI is treated as a tool, not a contributor meeting authorship criteria.
  3. The author remains fully responsible for all content, including AI-generated material — a higher bar than many users appreciate.
  4. Whether undisclosed AI use is "plagiarism" remains genuinely contested. Some institutions have expanded plagiarism definitions; others treat it as a distinct category of misconduct; legal scholars argue it is a different kind of ethical failure altogether.

The gap is in the space between "you must disclose" and a shared understanding of why non-disclosure is wrong. Is it wrong because you're stealing someone's words (plagiarism)? Because you're misrepresenting the reliability of your claims (falsification)? Because you're breaking a social contract about what academic authorship means (misconduct)? Different answers to this question will shape how seriously institutions treat violations and what remedies they pursue.

Further Reading

3. Beyond Plagiarism: Industrial Language Production and the Denial-of-Service Problem

The debates above — about citation, disclosure, academic integrity — take place within a framework that still assumes language is produced at human scale. But what happens when it isn't? A different set of concerns emerges when we consider the sheer volume of AI-generated text now flooding institutions, and what that volume does to the systems built to process it.

The Flooding Problem

A useful framing comes from a 2025 Conversation article which describes what amounts to a denial-of-service attack through language. The core insight: institutions — journals, magazines, courts, legislatures, universities — historically relied on the difficulty of writing to limit the volume of submissions they needed to process. Generative AI has removed that bottleneck. The result is not a subtle shift but a structural overwhelm:

The metaphor of a denial-of-service attack is apt: the systems aren't being defeated by sophistication but by volume. The cost of producing text has collapsed; the cost of evaluating it has not.

"AI Slop" and Information Pollution

The broader internet reflects the same dynamic. "Slop" — selected as 2025 Word of the Year by both Merriam-Webster and the American Dialect Society — refers to the flood of low-quality, machine-generated content saturating platforms: generic articles, clickbait, synthetic images, deepfake video, and spam music.

The scale is striking:

The KR Institute frames this as pollution in our communication environment — a useful analogy because, like environmental pollution, the harm is cumulative, distributed, and borne disproportionately by those with the least power to filter it out (KR Institute — "AI Slop I: Pollution in Our Communication Environment"). The once-fringe "Dead Internet Theory" — the claim that most online content is no longer human-produced — is, by some measures, becoming an observable reality (Wikipedia — AI slop).

What Kind of Language Is This? The "Bullshit" Problem

A philosophical strand of this debate asks what kind of language LLMs actually produce. Hicks, Humphries, and Slater's widely discussed 2024 paper argues that LLM output is best understood as bullshit in the precise sense defined by the philosopher Harry Frankfurt: statements produced with no regard for whether they are true or false. The authors argue that terms like "hallucination" and "confabulation" are misleadingly anthropomorphic — they imply the system is trying to be truthful and sometimes failing, when in fact truth is simply not a dimension the system operates in (Hicks, Humphries & Slater, "ChatGPT is bullshit" — Ethics and Information Technology, 2024).

This has been contested — Gunkel and Coghlan argue that the bullshit framework may itself be misapplied to a non-agentive system, and that anthropomorphic metaphors like "hallucination" may still serve useful roles (Gunkel & Coghlan, "Cut the crap" — PhilPapers). But the core observation has force: the institutional flooding problem above isn't just about volume, it's about volume of text that has no epistemic relationship to truth. When a journal is swamped with AI-generated letters, the problem isn't just that there are too many — it's that the text was produced by a process indifferent to whether any of its claims are true. The volume and the epistemic emptiness compound each other.

The Arms Race — and Why Detectors Make It Worse

The natural institutional response has been to deploy AI detection tools. But this creates what the Conversation article calls a no-win arms race: as detectors improve, so do "humaniser" tools designed to evade them, in a cycle of "rapid, adversarial iteration" with no stable endpoint.

Worse, current detection tools are systematically biased:

The detection approach, in other words, risks punishing the people least likely to be cheating — non-native speakers, neurodivergent writers, and those with less conventional prose styles — while sophisticated users evade detection easily.

What Connects This to the Research Ethics Debate

The plagiarism-and-citation framework discussed in Sections 1 and 2 assumes a context where text is authored, attributable, and produced at human scale. The flooding/slop problem breaks all three assumptions simultaneously:

  1. Attribution becomes meaningless when a single person can generate thousands of submissions, or when propagandists generate synthetic "grassroots" campaigns. The question shifts from "who wrote this?" to "was a human involved at all?"
  2. Scale defeats gatekeeping. Peer review, editorial oversight, and judicial triage all depend on submission volumes that humans can process. Machine-speed production overwhelms these systems regardless of whether each individual piece is "properly cited."
  3. The epistemic contract erodes. Academic communication presupposes that writers have some relationship to the truth of what they claim. When the production tool is indifferent to truth (per the Hicks et al. argument), that presupposition fails — not at the margins but structurally.

These are not problems that better disclosure policies can solve. They require institutional responses at a different level.

Positive Approaches and Constructive Responses

The picture above is bleak, but there are emerging constructive responses — both technological and institutional — that move beyond the detect-and-punish paradigm.

1. Content Provenance and Labelling

Rather than trying to detect AI-generated content after the fact, the Coalition for Content Provenance and Authenticity (C2PA) — led by Adobe, Microsoft, the BBC, and now Google — is developing open standards for cryptographically signing content at the point of creation. Content credentials embed metadata about who created something, when, with what tool, and whether AI was involved (C2PA specification; Google's C2PA integration).

Google's SynthID embeds invisible watermarks into AI-generated text, audio, images, and video at the point of generation. The US Library of Congress has launched a Community of Practice for content provenance in cultural heritage institutions.

Limitations are real — no watermark is simultaneously robust, unforgeable, and publicly detectable, and provenance metadata can be stripped or forged (UK Parliament briefing on AI content labelling). But the approach is promising because it shifts the burden from detection (which creates an arms race) to attestation (which creates a positive signal of authenticity).

2. Redesigning Assessment Around Process, Not Product

In education, the most promising responses move away from policing final outputs and toward assessing the process of thinking:

The underlying principle is a shift from "can you produce a correct text?" to "can you demonstrate that you understand?" — a question that remains meaningful even when text production is trivially cheap.

3. AI Literacy as a Core Competency

Rather than treating AI as a threat to be defended against, several frameworks argue for embedding critical AI literacy into education:

4. Institutional Adaptation Over Technological Fix

Perhaps the most important lesson from the literature is that no technological fix will solve a structural problem. The Conversation article is explicit: "there won't ever be a way to totally stop" fraudulent AI use. What institutions can do:

Further Reading: Language Flooding, Slop, and Constructive Responses