B. Autonomy: Staying in Charge of Your Own Thinking

Purpose: Sustained AI use changes you — not just your outputs. This document is about what happens to your intellectual capacity over time: how AI can expand it, how it can erode it, and what practices keep you in the driver's seat. The question is not whether to use AI, but how to use it in a way that leaves you more capable, not less.

The paradox: AI can increase autonomy

The anxiety-first framing — "AI threatens your autonomy" — misses something important. For many researchers, carefully used AI genuinely expands what they can do independently:

These are genuine autonomy gains. The point of this document is not to add to a general anxiety about AI but to distinguish the kinds of use that expand capacity from the kinds that erode it.

The distinction that matters: Autonomy lives in the judgment calls, not the execution. Letting Claude format your bibliography does not erode anything. Letting Claude define what your research question is — does.

How autonomy erodes: the mechanisms

None of these happen dramatically. They accumulate slowly, often feeling like productivity gains at the time.

Epistemic capture

Claude reframes your question, and you follow the new frame without noticing the shift. You came with question A — a hard, specific question shaped by years of familiarity with your field. Claude restructures it as question B: more tractable, better representable in text, closer to what its training handles well. You end up answering B.

The reframing often feels like an improvement — clearer, more structured, more operationalisable. Sometimes it is. Sometimes you have traded a genuine question for an answerable one.

Atrophied judgment

You stop exercising the skills needed to evaluate Claude's output, because you have outsourced the exercise. Over time, your calibration drifts. You can no longer tell whether a literature review is good, because you have not done literature reviews independently long enough to maintain the intuitions that judgment requires. The output looks plausible. You cannot see what is missing.

This is the slowest and most serious form of erosion, because it is self-concealing: as your ability to evaluate declines, your confidence in the output tends not to decline with it.

The risk is not speculative. Google DeepMind's paper introducing their own AI co-scientist system explicitly flags it as a concern about advanced AI in research: "Over-reliance on AI-generated suggestions in collaborative AI systems could diminish critical thinking." This is a warning from the team that built one of the most capable AI research tools yet developed, about their own system. The logic applies at every scale of AI use, from a researcher drafting a paragraph in Claude Desktop to a team running a multi-agent pipeline.

Gottweis, J., Weng, W.-H., Daryin, A., Tu, T. et al. (2025). Towards an AI co-scientist. arXiv:2502.18864.

Agenda drift

Your research questions migrate quietly toward what AI handles well — things with clear structure, canonical form, good representation in training data, recoverable from text. Away from what matters to your field but resists that form: 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 questions feel productive to pursue on a given day.

Confirmation accumulation

You show Claude your conclusions, ask it to develop or support them, it does. You feel confirmed. You have not tested anything — you have generated a more elaborate version of what you already believed, and the elaboration creates an illusion of evidence.

Voice homogenisation

Your own writing starts absorbing Claude's patterns — not because you asked it to write, but from sustained exposure to its output. The effect is subtle and insidious: it does not show up in any single text but in the drift of your style over months.

Peer substitution

You consult Claude instead of colleagues. Claude is available, patient, never busy, never dismissive. It does not challenge you in the ways a peer challenges you — with genuine disagreement rooted in different expertise, with the friction of a different intellectual agenda, with questions that reveal what you have taken for granted. The convenience is real. The loss is also real, and not symmetric: peer conversation does things AI cannot.

The signals: how to recognise erosion

These are not failure states — they are early warnings.

In the moment:

Over time:

The last signal is worth dwelling on. It is often partly true — Claude has broad coverage of published knowledge, available at any moment. But colleagues have something Claude does not: stakes, a shared project, genuine intellectual disagreement, the capacity to be actually wrong about something in ways that matter. If you find yourself preferring Claude's responses to peer conversation, that preference is worth examining.

The peer group as the final calibrator

The best check on autonomy erosion is not introspective — it is social. Your field's community of peers, collectively, maintains the standards that determine whether your work is good. Claude has no access to those standards except through its training, which is always behind, always partial, always oriented toward the average of what has been published rather than toward what your specific community currently values.

This means: the peer group is the ultimate evaluator that AI cannot replace.

Peer review, seminar discussion, conference feedback, the reaction of your supervisor or collaborators — these are the moments where you find out whether your work holds up in the community of people who share your standards and your stakes. They are also the moments where autonomy erosion becomes visible, because it is harder to hide in front of people who know your field and know you.

Use this as a design principle: maintain the conditions that keep peer conversation alive and productive. Not as a check on Claude, but as the relationship that gives your work its ultimate meaning. Claude can help you prepare for that conversation; it cannot replace it.

Good practice: maintaining and building autonomy

These are not rules — they are habits that keep your intellectual capacity exercised.

Write before you ask

For any intellectual product you care about, write a rough draft — even a rough, fragmentary one — before showing Claude. Force yourself to externalize your own thinking first. This does two things: it ensures Claude responds to your thinking rather than substituting for it, and it gives you a baseline to compare against, so you can see what Claude changed and decide whether the change is an improvement.

Interrogate reframings

When Claude restructures your question, pause before following. Ask explicitly: is this a better question, or a more tractable one? The two are not the same. Tractability — ease of processing, clarity of output — is not a measure of intellectual importance. Your field's genuinely hard questions are often hard precisely because they resist clean framing.

Preserve your disagreement reflex

Actively look for where Claude is wrong, incomplete, or subtly misaligned with your actual question. If you find that you rarely or never disagree with Claude's outputs, that is not a sign of quality — it is a sign that your critical engagement has softened. Disagreement is the exercise that keeps judgment sharp.

Maintain peer conversation

Use Claude for things peers cannot provide — availability, volume, patience, breadth. Use peers for what Claude cannot provide — genuine intellectual challenge, shared stakes, field-specific standards, the friction of a different agenda. These are not competing relationships; they are complementary ones. But they require deliberate maintenance, because Claude is always available and peers are not.

Keep some slow practices

Periodically do things the long way: read a paper without asking for a summary. Write a paragraph without drafting prompts. Work through a problem before opening Claude. These are not performances of virtue — they are maintenance of calibration. The point is not to avoid AI assistance but to ensure your judgment remains exercised enough to evaluate it.

Switch modes deliberately

Distinguish sessions where Claude is a tool (processing, formatting, searching, transforming) from sessions that are your own thinking (no Claude, or Claude only at the end). Do not let the modes blur. The blurring is where agenda drift begins.

Audit the relationship periodically

Every few months, ask yourself: Has my research question changed since I began using AI heavily? Is that change an improvement I can defend, or a drift I did not decide? Am I still talking to colleagues about my work? Can I reconstruct how my recent AI-assisted outputs were produced, and defend the key decisions in each?

The DISSINET context

For historians and social scientists working with historical sources, some autonomy risks take specific forms.

Interpretive authority. The interpretive work of reading primary sources — establishing what a source means, in its historical context, given what you know about the archive — is the core disciplinary skill. Claude can assist with translation, with pattern-finding across a corpus, with structuring what you have found. It cannot perform the interpretive act itself: that requires domain expertise, archival knowledge, and accountability to a scholarly community. The risk is not that Claude will interpret your sources badly. The risk is that you will mistake Claude's processing of sources for interpretation, and stop distinguishing between the two.

The anomalous source. Historical research often turns on anomalies — the source that does not fit, the gap in the record, the evidence that complicates the prevailing account. These are the hardest things to hand to AI, because AI tends toward the central tendency of what it has been trained on. It handles the typical well and the anomalous poorly. If your research depends on taking anomalies seriously, be alert to the possibility that AI-assisted workflows will systematically underweight them.

Scale and corpus work. DISSINET works with large corpora — archive scans, network data, batches of sources that exceed what any individual researcher could process manually. AI assistance at this scale is not a luxury; it may be the only way the work is possible at all. This is the autonomy gain in clearest form. The caution here is different: at scale, errors and biases in AI processing propagate across the whole corpus. A misreading of one source is one error; a systematic misreading embedded in a batch-processing workflow is a structural problem. The supervision that matters at scale is not source-by-source reading but audit of the workflow itself — understanding what the AI is doing and checking it at representative points.

Latin and other historical languages. Claude has meaningful capability with Latin and with many historical languages, but its training is uneven across periods, genres, and text types. It handles classical Latin better than medieval administrative Latin; it handles well-edited printed texts better than manuscript hands. Your field knowledge of what a text is and what kind of reading it requires is irreplaceable. Use Claude's language capabilities as a first pass, not as an authority.

The inter- and transdisciplinary dimension. Working across disciplinary boundaries creates both a specific vulnerability and a specific protective resource when AI enters the picture.

The vulnerability: in an inter/transdisciplinary team, you are frequently working in a domain that is not entirely your own. You already practice appropriate deference — you trust the network analyst's methodology, the historian's reading of the source, the programmer's code. AI can produce outputs that look authoritative precisely where that deference is already habitual. A historian may not question Claude's handling of network data because "that's the technical side." A programmer may not question Claude's interpretation of a medieval source because "that's the history." The existing respect for expertise boundaries can become a channel through which AI output passes unchecked.

The protective resource is the mirror of the same thing. Cross-disciplinary questioning is a form of verification that disciplinarily homogeneous teams cannot perform. The historian who asks naive questions about the network model may expose an assumption the programmer has stopped seeing. The programmer who reads the source interpretation with fresh eyes may notice what domain familiarity has made invisible. A team with a strong culture of mutual interrogation across expertise boundaries is structurally better positioned to catch AI errors — because it already practises the habit of not treating any single person's output as unchallengeable, and that same habit applies to Claude's.

The deeper principle

Autonomy is not about avoiding AI assistance. It is about remaining the person who decides what the question is, evaluates the answer, and takes responsibility for the result — across individual outputs and across your intellectual development over time.

The researchers who will navigate this best are not the ones who use AI least. They are the ones who remain curious about their own thinking — who notice when they are following rather than leading, who maintain the relationships and practices that keep their judgment alive, and who treat AI as a powerful tool in service of a research agenda they have formed themselves.

Related