Vibe Researching as Wolf Coming: Can AI Agents with Skills Replace or Augment Social Scientists?

This paper argues that AI agents equipped with specialized skills can augment, but not fully replace, social scientists by executing codifiable research tasks autonomously through "vibe researching," while highlighting the enduring necessity of human theoretical originality and tacit knowledge alongside the profession's emerging risks of stratification and pedagogical crisis.

The Gist

Imagine you are a chef trying to create a new, award-winning dish. In the past, you had to chop every vegetable, measure every spice, and stir every pot yourself. If you wanted to try a complex recipe, it took you weeks of hard labor.

Now, imagine you have a super-intelligent kitchen robot (an AI Agent) that can do all the chopping, measuring, stirring, and even plating for you. It can read 10,000 cookbooks in a second, check if the ingredients are fresh, and even write the recipe card for you.

This paper, titled "Vibe Researching as Wolf Coming," asks a scary but exciting question: If the robot does all the cooking, what is left for the human chef to do?

Here is the breakdown of the paper in simple terms, using some creative analogies.

1. The "Wolf" is Already Here

The title references the fable of "The Boy Who Cried Wolf," but with a twist: The wolf is real, and it's at the door.

In the past, computers were just calculators (Wave 1) or data collectors (Wave 2). They could do math or find facts, but they couldn't think.

• The Old Way: You asked the computer to run a number, and it gave you a result. You had to figure out what the result meant.
• The New Way (AI Agents): The computer is now a "research assistant" that can do the whole job. It can find the books, read them, write the essay, check the grammar, and even pretend to be a critic reviewing your work.

The author calls this "Vibe Researching." It's like "vibe coding" (where you tell a computer what you want, and it writes the code without you reading the details). In research, you tell the AI your idea, and it builds the whole study for you.

2. The Tool: "Scholar-Skill"

The author built a specific tool called Scholar-Skill. Think of this as a Swiss Army Knife for researchers that has 26 different blades.

• It can write the introduction.
• It can find the data.
• It can run the complex math.
• It can format the paper for specific magazines (journals).
• It can even simulate a panel of critics to tell you how to fix your paper before you submit it.

The author used this tool to write the paper you are reading right now. It did the heavy lifting, but the human (the author) still had to steer the ship.

3. The Big Problem: The "Cognitive Boundary"

The paper argues that we shouldn't just ask, "Can AI do this task?" Instead, we need to ask, "Does this task require a human soul (or 'tacit knowledge')?"

The author creates a map to separate tasks into two groups:

• The "Recipe" Tasks (High Automation): These are like following a strict recipe. "Add 2 cups of flour," "Run this specific math formula," "Check for spelling errors."
  • The AI is amazing here. It never gets tired, never makes a math error, and can read every book in the library in seconds.
• The "Gut Feeling" Tasks (Low Automation): These are like knowing which dish to cook for a specific guest, or knowing that a certain ingredient tastes weird in this specific season. This requires "field knowledge"—the unwritten rules, the politics, the intuition of what matters right now.
  • The AI struggles here. It can mimic the style, but it doesn't truly understand why an idea is brilliant or why a specific theory is dead. It can't feel the "vibe" of the academic community.

The Danger Zone: The paper warns that the boundary isn't between "Step 1" and "Step 2." The boundary cuts through every step. Even when writing a paper (a "Recipe" task), you need "Gut Feeling" to know if the argument makes sense. If you let the AI do the "Gut Feeling" parts, you might publish nonsense that looks perfect on the surface.

4. The Three Big Risks

If we let the AI take over too much, three bad things happen:

1. The "Fragile" Chef: If you let the robot chop all the vegetables, you forget how to hold a knife. If the robot makes a mistake (and it will), you won't have the skills to catch it. You become a "manager" of a process you don't understand.
2. The Rich Get Richer (Stratification): Only researchers at fancy universities with money can afford the best AI tools. Researchers at smaller schools or in other languages might get left behind. This creates a bigger gap between the "haves" and "have-nots" in science.
3. The Student Crisis: If we teach students to just "press the button" and let the AI do the work, they won't learn how to think. When they graduate, they might be great at using the tool, but terrible at solving problems when the tool fails.

5. How to Survive the Wolf (The 5 Rules)

The author doesn't say "Ban the AI!" Instead, he says, "Use it wisely." Here are the 5 rules for "Responsible Vibe Researching":

1. Be Honest (Disclose): Tell people, "I used an AI to help write this." It's not cheating; it's just being transparent.
2. Double-Check Everything (Verify): Never trust the robot blindly. If the AI writes code, you must run it yourself. If it cites a book, check if the book actually exists.
3. Keep Your Skills Sharp (Maintain Skills): Don't stop doing the hard work. Sometimes, write a paragraph by hand or run a math problem without the AI. You need to keep your "muscles" strong so you can spot when the AI is wrong.
4. Keep the Big Ideas Human (Protect Originality): The AI can give you 100 ideas, but you must choose the one that is truly new and important. The "spark" must come from you.
5. Share the Tools (Design for Access): Make sure these tools are available to everyone, not just the rich and powerful.

The Bottom Line

The paper ends with a powerful analogy: Aviation.
Pilots don't just "fly" anymore; they use autopilot. But they still train for years, check the instruments, and know how to take over if the autopilot fails.

AI is the autopilot for social science. It can fly the plane faster and smoother than ever before. But we still need to be the pilots. If we let go of the controls completely, we might crash into a mountain we didn't see coming.

The Wolf is here. The question isn't whether to let it in (it's already inside). The question is: Will you let it drive the car, or will you keep your hands on the wheel?

Technical Summary

Based on the paper "Can AI Agents with Skills Replace or Augment Social Scientists?" by Yongjun Zhang (March 2026), here is a detailed technical summary covering the problem, methodology, contributions, results, and significance.

1. Problem Statement

The paper addresses a paradigm shift in social science research driven by the emergence of AI Agents (systems with persistent state, tool access, and multi-step reasoning capabilities) versus traditional chatbots.

• The Core Question: As AI agents gain the ability to autonomously execute entire research pipelines (from literature review to manuscript submission), what constitutes the researcher's distinctive contribution? Is the researcher still the author, or merely a curator of AI-generated options?
• The Risk: The paper identifies a "vibe researching" phenomenon (parallel to "vibe coding"), where scholars describe a goal and accept AI execution without deep engagement. This raises concerns about:
  • Deskilling: The erosion of human capacity to verify or understand the work.
  • Epistemic Integrity: The inability of AI to handle tacit knowledge (field politics, editorial culture, theoretical novelty).
  • Stratification: The widening gap between researchers with access to advanced AI tools and those without.

2. Methodology

The paper employs a conceptual framework development grounded in a concrete operational case study.

• Case Study System: Scholar-Skill

  • Architecture: A plugin for Claude Code (Anthropic) installed locally on the researcher's machine.
  • Scope: Comprises 26 specialist skills organized into 13 stage-groups covering the full research lifecycle (Idea $\to$ Submission).
  • Orchestration: An orchestrator manages 18 phases with 53 quality gates and 5 hard stops (e.g., data safety, ethics compliance) to prevent pipeline advancement without meeting standards.
  • Key Technical Features:
    • Multi-Agent Panels: Uses panels of specialized agents (theorist, methodologist, editor, devil's advocate) to stress-test ideas and drafts.
    • Verified Citation Pool: A zero-tolerance mechanism against citation fabrication. It builds a pre-approved pool from the user's local library (Zotero/Mendeley) and cross-references against APIs (CrossRef, Semantic Scholar, OpenAlex) before and after drafting.
    • Analytic Action Dispatch (AAD): Automatically classifies and executes code-based tasks (regressions, power analysis) in parallel, leaving only external tasks (e.g., IRB approval) for humans.
    • Dual Execution Modes: Operates in "data-available mode" (executing real analysis) or "code-template mode" (generating runnable scaffolding for future data).

• Analytical Framework:

  • The author develops a Cognitive Task Framework classifying research activities along two dimensions:
    1. Codifiability: Can the task be decomposed into explicit rules?
    2. Tacit Knowledge Requirement: Does the task require unarticulated field knowledge (politics, timing, intuition)?
  • This framework is used to map the delegation boundary between human and AI.

3. Key Contributions

The paper makes three primary theoretical and practical contributions:

1. Definition of "Vibe Researching": It names and analyzes the shift from "chatbot-assisted" writing to "agent-executed" research pipelines, distinguishing the current wave of automation (Wave 4: Multi-step Reasoning) from previous waves (computation, data collection, text analysis).
2. The Cognitive Delegation Boundary: It argues that the boundary between human and AI is cognitive, not sequential. It does not fall between stages (e.g., "AI does analysis, humans do theory") but cuts through every stage.
   • Delegable: High codifiability, low tacit knowledge (e.g., running regressions, formatting citations).
   • Non-Delegable: Low codifiability, high tacit knowledge (e.g., theoretical originality, field judgment, identifying "live" vs. "settled" debates).
3. A Framework for Responsible AI Integration: It proposes a taxonomy of four task types (Execution, Communication, Planning, Formulation) to guide researchers on where to delegate and where to maintain human oversight.

4. Results and Findings

• AI Capabilities (Strengths):
  • Speed & Coverage: AI agents can synthesize thousands of papers in minutes and execute complex methodological pipelines (e.g., staggered DiD, causal forests) instantly.
  • Methodological Scaffolding: AI lowers the barrier to entry for advanced methods, allowing researchers to attempt complex designs they might otherwise avoid.
  • Quality Control: The scholar-skill system successfully prevents citation fabrication through its 5-tier verification pipeline and ensures paper-to-code correspondence in replication packages.
• AI Limitations (Weaknesses):
  • Theoretical Originality: AI excels at recombining existing frameworks but fails to generate genuinely new mechanisms or paradigm-shifting concepts (e.g., Granovetter's "weak ties"). It is an articulator, not a creator.
  • Tacit Knowledge: AI lacks access to field politics, editorial preferences, and the "metaknowledge" of what is currently considered interesting or politically fraught in a subfield.
  • The Verification Gap: If a researcher delegates a task entirely without possessing the underlying skill, they cannot verify the output. This leads to a "fragile augmentation" where the human loses the capacity to judge the AI's work.
• Stratification Risks: The paper identifies four axes of inequality: Cost (subscription fees), Language (English bias), Technical Skill (command-line proficiency), and Field (calibration bias toward specific journals).

5. Significance and Implications

The paper concludes that AI agents are not merely tools but qualitative shifts in research infrastructure. Its significance lies in:

• Pedagogical Crisis: Traditional PhD training focuses on execution skills (running regressions, coding interviews) which AI can now automate. The paper argues for a shift in training: teaching methods as foundations for evaluation rather than just production, and prioritizing deep theoretical training where AI is weakest.
• Normative Shift: The profession must move from "can AI do this?" to "how do we govern this?" The paper proposes five principles for Responsible Vibe Researching:
  1. Disclose: Transparent reporting of AI assistance.
  2. Verify: Human review of all code and citations.
  3. Maintain Skills: Deliberate practice of delegated tasks to preserve judgment capacity.
  4. Protect Originality: Ensuring the research question and theoretical contribution remain human.
  5. Design for Access: Promoting open models and tools to prevent academic stratification.
• Future Outlook: The "wolf is at the door." The paper suggests that the future of social science depends on using AI to amplify human reach while rigorously protecting the human capacity for original inquiry and tacit judgment. It draws an analogy to aviation: safety comes not from banning automation, but from preserving human judgment at critical decision points through rigorous training and governance.