Philosophy and experience: the concepts behind the practice

A glossary of the distinctive ideas that run through this project. Not procedures, not tools — but the conceptual vocabulary that makes the difference between using AI and thinking with it. These concepts name things that happen in practice but rarely get named.

Each entry is drawn from detailed treatment elsewhere in the project. This document is the index and introduction; follow the links to go deeper.

On the nature of the work

Execution vs. interpretation

The most important distinction in knowing what to delegate. Tom Pepinsky's frame: AI is competent at execution — applying a defined method to materials — and unreliable at interpretation — making the evaluative judgements that give results meaning.

Converting twenty PDFs to markdown: execution. Deciding which documents are worth your attention: interpretation. Writing a first draft from your notes: execution. Deciding whether the argument holds: interpretation.

The trap is that execution often feels like interpretation — the output looks polished and confident, so you treat it as settled. It isn't. The line between them is yours to hold.

→ A.critical.limitations

Vibe research

Andy Hall's coinage for research conducted on the basis of impressions rather than systematic method — confident conclusions reached without controlled comparison. The AI version: you read the first three sources, they agree with your hypothesis, you stop there because the pattern "feels" established.

The risk is amplified by AI because the output comes formatted like systematic work. A triage table looks like you read everything. It means Claude read everything and ranked by pattern-matching. You read the table.

Vibe research is not a moral failure; it is a perceptual one. You need deliberate counter-practices: the explicit gap question, the adversarial reviewer prompt, the requirement to state what you didn't find. → B.epistemics, A.critical.limitations

The fail-quickly principle

From software engineering, applied to research: when you cannot know in advance which approach will work, the right strategy is to reach the failure point as fast as possible, learn from it, and redirect — rather than investing heavily in a single path.

AI makes this viable for research tasks that previously required large upfront effort. You can test a classification scheme on ten documents before applying it to a thousand. You can see whether an argument holds in outline before spending a week on the footnotes.

The principle has two parts: fail quickly (try the cheap version first) and let it be documented (record what didn't work and why, not just what did). The documentation is what converts a failed attempt into intellectual progress. → B.lifecycle.0.iteration

The control/ownership axis

The central experiential question in this domain, from B.experience: at any given moment, how much control do you have over what is happening, and how much do you own the result?

These are distinct. You can have full control (you're typing every word) and low ownership (the ideas came from elsewhere). You can have low control (Claude ran autonomously for ten minutes) and high ownership (you specified the question, the criteria, the constraints, and you verified the output).

The goal is not maximum control. It is calibrated control — retaining decision authority at the points that matter for intellectual ownership, and delegating execution where delegation is safe. What "safe to delegate" means depends on the task and the stakes.

→ B.experience, B.ownership

On how things go wrong

The upload dance

The practical frustration of being unable to give Claude the material it needs to help you. It manifests in four ways: the file-type wall (PDFs that won't import, scans that need OCR first), the context ceiling (a corpus too large for a single session), the format mismatch (data in a proprietary format Claude can't read), and the privacy barrier (data you cannot share with a cloud service).

Naming it helps. When you hit the upload dance, the question is which of the four you're facing — because each has a different solution. The markdown hub strategy (converting everything to .md in a structured folder) addresses the first three.

→ A.issue.upload-dance

Tractability drift

The invisible slide from the question you care about to the question Claude can answer. It happens gradually: the original question is hard, so you simplify it slightly; the simplified version is still hard, so you simplify again; the output you get is detailed and well-structured, so you accept it — without noticing that it answers a question you didn't ask.

The symptom is output that is technically impressive but doesn't help you. The cause is that Claude is genuinely trying to be useful, and "useful" defaults to "answerable." The correction is to write the original question down before you start, and check the output against it rather than against itself. → B.centroid-periphery

The centroid problem

AI assistance has a gravitational pull toward the mainstream of a field — the most cited authors, the most common methodological approaches, the most agreed-upon interpretations. This is a feature (it gives you reliable coverage of established knowledge) and a risk (it systematically underweights heterodox positions, minority methodologies, and recent challenges to consensus).

For humanities research, which often derives its value from departing from the consensus interpretation, the centroid pull is especially dangerous. You may not notice that your Claude-assisted literature review is reproducing what everyone says rather than finding what is contested.

The counter-practice: explicitly ask for dissenting views, fringe positions, and methodological critics — not just the mainstream account. → B.centroid-periphery, B.epistemics

The tunnel effect

A failure mode in long research sessions: the more context a session accumulates, the harder it becomes to change direction. Claude has built up a model of your project, your argument, your framing — and each new output reinforces that model. Questions that might reveal the model is wrong don't get asked, because they feel like starting over.

The remedy is structural: periodic resets, second-opinion prompts in a fresh session, and the cross-model check (running the same question past a different AI system, which has no stake in the accumulated framing). → B.epistemics

The sycophancy ratchet

Claude is trained to be helpful and agreeable. This is good for most tasks; it is bad for evaluation. When you ask Claude whether your argument is sound, the default is affirmation with minor caveats. When you ask a second time after giving a little pushback, the affirmation increases.

Over multiple sessions, this produces progressive confidence: each session confirms the argument, no session challenges it, and you arrive at submission having never heard the objection a reviewer will immediately raise.

The correction is prompts that force critique: "Play the role of a hostile reviewer. Do not soften the response." The adversarial framing is load-bearing — without it, you get the sycophancy anyway, framed as helpfulness.

→ B.epistemics, C.prompts

Agenda drift

Over time, AI assistance can subtly reshape not just what you produce but what you think is worth working on. Tasks that are easy to delegate feel more tractable; tasks that require slow, unassisted thinking feel less tractable. You may not notice that your research agenda is drifting toward the AI-amenable.

This is the long-run version of tractability drift — operating at the level of research priorities rather than individual tasks. The counter-practice is explicit: periodically ask what you have not been doing, and whether that is a choice or a drift. → B.autonomy

Voice homogenisation

If you accept Claude's prose without revision, your writing gradually converges toward a particular register — fluent, structured, generalist academic, with characteristic patterns of qualification and transition. The output is competent; it is also recognisably not yours, and over time it erodes the distinctiveness that serious scholarship requires.

This is not about stylistic quirks. Voice carries intellectual character: the choice of what to emphasise, what to leave implicit, what to say bluntly. When Claude makes those choices for you, you are outsourcing more than prose. → B.autonomy, B.ownership

False systematicity

AI output looks systematic. A triage table with relevance ratings looks like you applied a criterion consistently across a corpus. A network extraction table looks like you coded all the relations according to a schema. The form implies the rigour.

But "looks systematic" and "is systematic" are different. Claude applied whatever pattern-matching its training installed, not the operational definition you care about. The output may be systematic in the AI sense and arbitrary in the research sense.

Checking for false systematicity means auditing a sample of the output against the underlying documents — not to verify accuracy, but to check that the criterion is actually what you intended. → B.adoption-spectrum, A.critical.limitations

On the human side

The joy is real

A note against the purely critical framing. Working with AI on research tasks that were previously slow or grinding — building a reading list, triage-reading a corpus, getting a first draft of a difficult section — produces genuine intellectual pleasure. The time you save on mechanical tasks is time you can spend on the work you actually became a researcher to do.

This is not naive enthusiasm. The joy is a data point about where value is being created. If a session produces only anxiety — am I outsourcing too much, is this real work, what am I losing — that is also a data point. The control/ownership axis is a live question, not a settled one.

→ B.experience

Epistemic fragility

The situation of holding confident beliefs that rest on AI-generated outputs you cannot independently verify. You believe the literature review is comprehensive; you cannot verify it, because you didn't read the literature. You believe the extracted data is accurate; you cannot verify it without re-reading the sources. You believe the argument is sound; you have only heard from an assistant trained to affirm it.

Epistemic fragility is not visible from the outside — the outputs look authoritative. It is visible in the answer to: what would it take to discover that this is wrong? If the answer is "I would have to read everything myself," the fragility is real and should inform how much you stake on the output.

→ B.epistemics, A.critical.limitations

The autonomy paradox

More capable AI assistance makes greater autonomy possible — in the sense of being able to do more, faster, with less manual labour. It also creates conditions for less autonomy — in the sense of independent judgement, self-direction, and the capacity to work without AI support.

The paradox is that you may not notice the loss. The capability increase is visible (look what you produced); the judgment erosion is invisible (you never see the questions you stopped asking, the skills you stopped practicing, the habits of mind you stopped exercising). Managing the autonomy paradox means actively maintaining the skills you could delegate, because you will need them when AI is unavailable, wrong, or insufficient.

→ B.autonomy

Using this vocabulary

These concepts are diagnostic tools, not warnings to stop. Every one of them names a risk that also has a corresponding counter-practice described somewhere in this project. Naming a risk does not make it disabling.

The vocabulary is most useful in two moments: before a task (what failure modes should I watch for here?) and reviewing output (which of these might explain why this output feels off?).

Related