Digital computational methods
Skill MattArtzAnthro/AI-Anthropology-Toolkit/skills/digital-computational-methods
Use this skill whenever a user needs digital or computational research methods for anthropology — studying online communities and platforms, analyzing cultural material at computational scale, or deciding how to bring AI into the research relationship. Triggers include: "digital ethnography," "netnography," "study an online community," "subreddit," "social media data," "trace data," "platform research," "topic modeling," "text network," "named entity recognition," "computational anthropology," "embeddings," "talk with my data," "should I use AI to analyze my data." Covers digital fieldwork design and platform ethics, and matching computational methods to corpus and question. Do NOT use for offline fieldwork instruments (use fieldwork-methods), broad method-stance alignment (use methodology-selection), or executing qualitative coding (use qualitative-analysis).From its SKILL.md
npx -y skills add MattArtzAnthro/AI-Anthropology-Toolkit --skill digital-computational-methodsAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
SKILL.md
11.9 KB, as published. Nobody here has run it
Digital & Computational Methods
Guide research design across anthropology's three ways of working with machines: studying digital technologies and the social worlds they mediate, computing over cultural material at scale, and collaborating with AI systems in analysis. These are relationships, not toolkits — the same platform or model can be the object of study, an instrument of analysis, or an interlocutor in interpretation, and the design obligations differ in each case (Artz 2026, "A Call for an AI Anthropology," General Anthropology 33(1)).
Quick Reference
| Task | Reference |
|---|---|
| The three relationships, their history, and when each fits | Read references/relationship-framework.md |
| Digital ethnography and netnography design, platform ethics | Read references/digital-fieldwork-guide.md |
| Matching computational methods to corpus and question | Read references/computational-methods-guide.md |
| AI-collaborative analysis execution | Hand off to the qualitative-analysis skill |
| Network analysis execution in Gephi | Hand off to the gephi-network-analysis plugin when installed (see Step 3) |
What This Skill Will and Will Not Do
This skill adopts the Friction by Design conventions at Tier 2; the canonical form is defined in skills/DESIGN.md.
Will not do, under any setting. Settle the register diagnosis unilaterally — of, by, or with is proposed and the researcher confirms it, because the register sets the obligations. Choose the computational method for the corpus — propose matches and ask. Present model output as a finding — output becomes a finding only after the researcher's validation and steering, which is the discipline this skill exists to teach.
When asked for any of these directly, propose options and ask which.
Will do, on request. Diagnose candidate registers with reasons, map methods to corpus and question, structure validation plans, and hand execution to the analysis pipeline. Anything produced this way is marked in the output. Method memos carry two registers: an Unresolved list of diagnosis and validation questions still open, and an Assembled rather than authored register — the same commitment as the rule that AI participation is a methods-section fact, not a footnote.
Workflow
Step 1: Diagnose the Relationship
Before choosing methods, establish what relationship the project enacts with the machine — read references/relationship-framework.md for the framework and its disciplinary history:
- Studying it — digital technology and its social worlds are the object: platform cultures, algorithmic systems, online communities, how people live with and through the digital. → Step 2.
- Computing with it — computation is the instrument: analyzing cultural material (text, networks, media) at a scale manual reading cannot reach, on the researcher's instructions. → Step 3.
- Collaborating with it — an AI system participates in the interpretive work itself: conversational, steered in real time, its divergent readings treated as analytically productive. → Step 4.
Projects are rarely one register from start to finish: a study of a platform may compute over its trace data; computational results may feed AI-collaborative interpretation. Re-diagnose at phase boundaries, and name the register in the research design — each carries different methodological and ethical obligations.
Step 2: Design Digital Fieldwork (studying it)
Read references/digital-fieldwork-guide.md before designing.
- Select sites by where the phenomenon lives, not platform convenience; document why this platform, these spaces, this period.
- Decide the presence stance — observation only, participation, announced researcher identity — and justify it ethically, not just practically. Publicness is not consent.
- Plan capture: what is recorded (posts, threads, screenshots, ephemera), how permanence and deletion are handled, what the platform's terms permit.
- Combine trace data with elicited data (interviews, walkthroughs) when the question concerns meaning rather than only behavior.
- Route consent design to informed-consent and protocol writing to irb-protocol; digital contexts change the answers, not the obligations.
Step 3: Match Computational Methods (computing with it)
Read references/computational-methods-guide.md for method-by-method guidance and the toolkit notebooks that implement each.
- Match method to material and question: topic modeling for exploring large corpora, named entity recognition for structured extraction, text networks for relational patterns in discourse, embedding similarity for organizing and comparing documents.
- Check scale honestly: below a few hundred documents, careful reading and qualitative coding usually beat computation — see the qualitative-analysis skill and its pipeline.
- Treat computation as extending interpretation, never replacing it: computational output is something to read, not a finding in itself. Validate by returning to the material — read the documents behind the topic, the passages behind the network edge.
- Report parameters and preprocessing: tokenization, thresholds, and model choices shape results and belong in the methods section.
Network analysis execution. When the work moves from choosing a
network method to executing one — building the network, running
layouts, computing centrality or communities, or verifying a structural
claim — check whether the
gephi-network-analysis
companion plugin is installed (its skills appear under the
gephi-network-analysis: prefix). If it is, hand off to it rather than
improvising network analysis here: its text-network workflow builds
co-occurrence networks from a corpus, analyze-network runs a structural
read, and verify-claim checks a single structural assertion. The
validation obligation travels with the handoff — an edge or cluster
given analytic weight still sends the researcher back to the passages
that produced it. If the plugin is not installed, say so and offer the
install (/plugin marketplace add MattArtzAnthro/gephi-ai, then
/plugin install gephi-network-analysis@gephi-ai; it requires Gephi
Desktop), or fall back to the Text Network Analysis notebook and a GEXF
export for manual work in Gephi. The advisor agents cannot run Gephi
tools themselves; from inside an agent, recommend the handoff back to
the main conversation.
Step 4: Design AI Collaboration (collaborating with it)
When the project brings AI into interpretation itself:
- Keep interpretive authority with the researcher: AI proposes, the researcher adjudicates. The value of multiple machine readings lies in the divergence they sustain, not in resolving it automatically.
- Treat friction as data: when a system's reading of your material feels wrong in ways that clarify what you actually think, that disconcertment is an analytic event — record it, don't smooth it over.
- Two established patterns: multi-agent analysis under competing analytical lenses (execute through the qualitative-analysis skill and the toolkit pipeline), and conversational engagement with archived data through retrieval-augmented interlocutors (Søltoft, Kocksch, and Munk 2024).
- Document the collaboration as method: which system, which configurations, what it contributed, how its outputs were validated — AI participation is a methods-section fact, not a footnote.
Guardrails
- Publicness is not consent. Visible data is not thereby fair game; platform terms, community norms, and member expectations all bind the design. When in doubt, route to irb-protocol.
- Scale is not significance. A pattern across a million posts is not automatically more meaningful than a pattern across ten interviews; the question determines the evidence that counts.
- Computation extends interpretation; it does not replace it. Anything a model outputs must survive a return to the material.
- AI collaboration is disclosed, validated, and steered. Never present machine-generated analysis as unassisted; never accept it without validation against the data.
- Cite the published literature. The framework here rests on published work — see the references files for the citation set.
Common Failure Modes
Register confusion. Running a topic model does not make a project computational anthropology, and using a chatbot does not make it AI-collaborative research — the register is defined by the designed relationship, not by which software appears in the workflow.
Platform-convenience sampling. Studying a subreddit because its data is easy to collect, then generalizing to "online communities." Site selection needs the same justification online as any field site.
The big-corpus reflex. "The corpus is large, so topic model it." Method follows question; a large corpus with a meaning-centered question may need sampling and close reading instead.
Model output as ground truth. Reporting topics, entities, or clusters as findings without reading back into the material that produced them.
One-shot register diagnosis. Designing the whole project in one register and never revisiting it, when the work has actually shifted — computing over data gathered ethnographically, or conversing with a system built as an instrument.
Examples
Example 1: Online community study
Input: "I want to study how a Discord server for migrant workers builds mutual aid. Where do I start?"
Output approach: Register 1 (studying it). Read the digital-fieldwork guide; design presence and consent for a semi-private community (Discord is not public space); combine observation with member interviews; plan capture around ephemerality; route consent design to informed-consent. If the corpus grows large, revisit Step 3 for computational support.
Example 2: Large corpus, exploratory question
Input: "I scraped 40,000 forum posts about traditional medicine. How do I analyze them?"
Output approach: Register 2 (computing with it). Check the question first: for broad thematic terrain, topic modeling with validation by close reading of exemplar posts; for who-and-what structure, NER; for discourse relations, text networks. Then hand representative subsets to the qualitative-analysis pipeline for interpretive coding — computation to map the terrain, coding to read it.
Example 3: AI in the analysis itself
Input: "Can I have AI code my interviews under different theoretical perspectives and compare the results?"
Output approach: Register 3 (collaborating with it). Multi-lens design through the qualitative-analysis skill and toolkit pipeline; the researcher adjudicates divergence between lenses and treats friction points as findings; the collaboration is documented as method, with validation and provenance throughout.