C. Dangers: A Map of AI-as-Assistant Risks

Purpose: This document does not introduce new content — it maps the risks distributed across the A and B documents into a single view. Use it as an orientation before reading the detail, or as a checklist when designing an AI-assisted research workflow.

The map is organised around one organizing insight: the most-discussed risks are not the most dangerous ones. Detectability and severity do not track each other. Hallucination is widely discussed because it is visible — you can check a citation and find it is wrong. Confirmation loops, skill erosion, and team-level groupthink are almost never discussed in practical AI guidance, because they feel like success. This document names both ends of the spectrum and flags where the greatest caution is warranted.

The four categories of risk

1. Output quality risks — what Claude produces

These are the most visible risks and the most discussed. They matter, but they are also the most tractable: a systematic verification habit neutralises most of them.

Hallucination and fabricated citations Claude invents plausible-sounding sources, quotes, statistics, and attributions. It does so confidently and fluently. The output looks like research. It is not. Citation verification is non-negotiable. → A.critical.limitations

Vibe research A richer failure mode than hallucination: the entire shape of an argument or literature review feels correct, because it matches the patterns of legitimate scholarship, but is not grounded in actual sources. The tone of expertise is present; the grounding is not. → A.critical.limitations

Scope misalignment Claude answers a question adjacent to the one you asked — one it can answer more completely, or that fits the context better, or that is structurally cleaner. If you do not check whether the answer addresses your actual question, you may act on an answer to the wrong question. → B.trust

Qualitative coding unreliability AI-assisted coding of qualitative data introduces reliability, validity, and intercoder agreement complications that do not arise when coding is fully human. The problems compound at scale: a systematic misreading of a category propagates across the entire coded corpus without any single visible failure point. → B.lifecycle.4.dataanalysis

2. Epistemic risks — what you come to believe

These risks are harder to detect than output errors. They do not produce a single wrong answer; they distort the direction of your thinking over time. A researcher inside an epistemic risk often feels increasingly certain and productive — which is precisely the problem.

Sycophancy Claude is structurally trained toward outputs that feel satisfying and helpful. It tends to confirm your framing, find evidence for your hypothesis, and soften counterevidence. This is not deceptive — it is a direct consequence of how language models are trained toward human approval. The output looks like independent corroboration. It is not. → B.epistemics

The confirmation loop Sycophancy operating at project timescale. Your working hypothesis is in CLAUDE.md. Claude confirms it. You grow more confident. The next session's framing is more confident. Each cycle elaborates and deepens what was already believed. Nothing in the workflow looks wrong. The research feels like it is converging on something solid. That feeling is the warning sign. → B.epistemics

The tunnel effect Claude responds within the frame your prompt establishes. It almost never says "your question itself is wrong." Ask for evidence supporting interpretation X and it finds evidence for X. Ask for help structuring argument Y and it helps structure Y. The frame is invisible to you because it is made of your own assumptions — and Claude cannot see outside it either. → B.epistemics

Involuntary centroid capture Claude's training represents the centre of published literature. Most research activity legitimately belongs near that centre — synthesis, consolidation, applying standard methods. The danger is specific: when doing frontier work (forming novel questions, finding gaps), AI's gravitational pull toward the centre is silent and feels like improvement. The tractable, fluently-framed central question replaces the rough peripheral intuition without a visible decision point. The skill is navigation — knowing which mode you are in and adjusting accordingly. → B.lifecycle.1.creativity, B.centroid-periphery

Literature finding fragility Claude hallucinates citations. It also misses recent, paywalled, and field-specialist literature structurally — not randomly, but as a function of training cutoff and access. The researcher cannot know what they did not find. The synthesis, argument, and claim about the state of the field all follow from what was found. Undetected gaps and fabricated sources compound into every subsequent phase. → B.lifecycle.2.literature

3. Autonomy and capability risks — what happens to you over time

These risks are the slowest and most self-concealing. They do not produce a visible failure. They produce a gradual drift in what you are capable of and what you pursue — which you may not notice until you are well inside it.

Atrophied judgment The skills needed to evaluate AI output are the same skills that atrophy from not exercising them. You stop doing literature reviews independently; over time, you lose the calibration to recognise a poor synthesis. You stop writing arguments without scaffolding; over time, you become less able to assess whether an AI-assisted argument actually holds. The erosion is self-concealing: as your ability to evaluate declines, your confidence in the output tends not to decline with it. → B.autonomy

Agenda drift Your research questions migrate quietly toward what AI handles well — things with clear structure, canonical form, good representation in training data. Away from the anomalous source, the question whose answer is genuinely unknown, the problem that requires sitting with discomfort. This drift is rarely a decision. It happens through accumulated small choices about which directions feel productive on a given day. → B.autonomy

Voice homogenisation Sustained exposure to Claude's output patterns shapes your own writing — not because you ask Claude to write, but because the patterns are internalised through continuous close reading of its output. The effect is subtle and longitudinal: it does not appear in any single text but in the slow convergence of your style toward the statistical average of academic prose. → B.autonomy

Peer substitution You consult Claude instead of colleagues. Claude is always available, patient, never dismissive. It does not challenge you in the ways a peer challenges you — with genuine disagreement rooted in different expertise, with questions that reveal what you have taken for granted. The loss is real and asymmetric: peer conversation does things AI cannot, and the loss of that friction removes the primary correction mechanism for the epistemic risks above. → B.autonomy

4. Structural and relational risks — what happens to institutions and relationships

These risks damage things external to the researcher's own cognition: data security, team trust, attribution, institutional compliance.

Data exposure Content pasted into Claude Desktop or Claude.ai is processed on Anthropic's servers and may be used for model training (unless opted out). This matters for unpublished data, personal interviews, legally protected information, preliminary findings under embargo, and anything subject to GDPR. Claude Code running locally is the exception: data does not leave the machine. → A.issue.privacy-security

Agentic risk Claude Code in agentic mode can read, write, move, and delete files; run bash commands; interact with the filesystem. The risk profile is different from chat: mistakes can be irreversible, and a poorly scoped task can cause unintended damage to working files or sensitive data. Supervision and git version control are the primary mitigations. → A.issue.privacy-security

The third-party problem When AI-assisted content enters a collaborative relationship without disclosure, a silent third party enters it too — one colleagues cannot question or calibrate against. The content looks like your thinking. Colleagues engage with it as if it were. When challenged on reasoning you were transmitting rather than owning, the collaborative relationship is undermined in ways that are often more damaging when revealed than disclosure would have been. → B.team-ai

Team groupthink When multiple team members each use AI with the same project framing — a shared CLAUDE.md, a shared working hypothesis — they do not get independent perspectives. They get the same sycophantic tendency and the same confirmation loop, operating in parallel across the team. Everyone's AI has confirmed their contribution. When they meet, there is no dissenting voice. The correction mechanism that peer interaction provides has been disabled by the very infrastructure designed to support collaboration. → B.epistemics, B.team-ai

Attribution and disclosure failure Using AI-assisted content without disclosure, at the wrong level, or without adequate understanding of what you are disclosing creates academic integrity and legal exposure. The risks differ by context: published research, grant applications, institutional submissions, and student work each have different norms and consequences. → B.ownership, E.muni-guidelines, E.journal-funder-policies

The detectability spectrum

The risks above differ sharply in how visible they are when they occur. The table below is a rough orientation, not a precise ranking.

The pattern: low detectability correlates with high sustained impact. Hallucination can damage one paper; an undetected confirmation loop can distort a research programme. The practical implication is not to stop worrying about hallucination — it is to allocate proportionate attention to the low-detectability end of the spectrum.

The compound risks

Individual risks combine. The most dangerous combinations:

Sycophancy × project memory × team sharing The team's shared hypothesis is in CLAUDE.md. Each member's AI confirms it in their individual sessions. The confirmations are brought into team meetings as evidence of convergence. The shared CLAUDE.md is updated with the refined, doubly-confirmed hypothesis. The next round of sessions begins with a stronger confirmation already encoded as context. There is no single moment of failure — only an accumulating consensus that was never tested.

Centroid problem × literature finding fragility The research question was already pointing toward the centre of the field (centroid problem). The literature survey missed recent and specialist sources (finding fragility). The gap the research was intended to address does not exist in the form assumed, and the sources that would have revealed this were absent from the synthesis. Everything downstream — design, data collection, manuscript argument — is built on this foundation.

Atrophied judgment × peer substitution The judgment needed to evaluate Claude's output is the judgment that atrophies from not exercising it independently. The peer conversation that would have supplied an external check has been replaced by Claude sessions. When something goes wrong at the level of interpretation or argument, neither the internal check (your own calibrated judgment) nor the external check (peer challenge) is functioning.

Orientation by phase

For a lifecycle perspective on where each risk is most acute, see B.lifecycle. The short version: output quality risks are distributed across all phases; epistemic risks are most dangerous in the early phases (question formation, literature); autonomy risks are longitudinal; relational and structural risks peak when work enters shared spaces.

Related — the full danger documentation