How to Use AI for Research Papers: A Practical Walkthrough

The AI Research Paradox: More Output, More Risk

Never before has a tool made academic writing faster — or its pitfalls easier to stumble into. The same technology that helps a researcher draft a literature review in an afternoon can also fabricate a citation so convincing that even the cited author doesn't recognize it.

The adoption numbers are staggering. According to the Stanford AI Index Report, 78% of organizations reported using AI in 2024, up from 55% just one year earlier. In academia specifically, a Frontiers survey of roughly 1,600 researchers across 111 countries found that more than half have used AI while peer reviewing manuscripts. The tools are everywhere, and they are actively reshaping scholarly output.

The productivity gains are real and measurable. An analysis of nearly 2.1 million preprint abstracts from January 2018 to June 2024 found that researchers who use generative AI publish significantly more papers, with LLM-assisted writing showing more complex language and broader citation ranges. Study coauthor Yian Yin of Cornell admitted he was surprised to see "all these consequential patterns happen in the real world" so quickly.

Yet the risks are equally documented, and they cut to the heart of research integrity:

• Fabricated citations — computer scientist Guillaume Cabanac discovered his work cited in a dental journal paper he had no connection to, telling Nature, "I couldn't recognize my own reference" (Nature news).
• Unreliable AI detectors — some institutions have switched off detection tools entirely due to algorithm biases and content obfuscation (peer-reviewed methodology research).
• Overreliance on AI that erodes critical thinking and introduces confirmation bias into analytical work.

Peer-reviewed work frames this tension as the "AI co-piloting-hallucination paradox": optimal outcomes come not from maximizing technology alone, but from jointly optimizing the interaction between human and machine. Meanwhile, Nature reporting notes rising "AI slop" on preprint servers and conferences imposing submission caps to manage volume.

This is why ad-hoc prompting fails. A researcher who asks ChatGPT for a literature review without validation, disclosure, or a structured workflow is gambling with their credibility. The evidence points clearly toward a disciplined process — plan, select tools, execute, validate, disclose — rather than improvisation. The same principle applies to the broader AI tooling ecosystem: at SkillzBench, we're building exactly this kind of verification infrastructure for agent skills, because unvetted AI capabilities carry real risk. The rest of this walkthrough shows how to get the productivity without the pitfalls.

What AI Is Actually Good At (and Where It Fails)

AI can draft a literature review paragraph in seconds, but it can't tell you whether that paragraph actually argues anything. That gap is where most researchers get into trouble.

A systematic review of 24 studies mapped six domains where AI now supports academic writing: idea generation, content structuring, literature synthesis, data management, editing and review, and communication and ethics. ChatGPT was singled out as showing "substantial potential" across those categories. But the same review flagged where the technology still falls short.

The pattern is consistent across the evidence. AI excels at single-paper comprehension and theme identification, but struggles with cross-paper synthesis and complex reasoning. An empirical study testing AI-assisted systematic reviews on 96 scholarly articles found that current tools have "limitations in complex data synthesis and inter-paper comparison." Translation: AI can summarize what one paper says, but connecting findings across a body of literature still requires human judgment.

This is where the "AI co-piloting-hallucination paradox" comes in. As peer-reviewed methodology research puts it, optimal outcomes come from "jointly optimising the interaction between the human and the machine," not from maximizing automation. AI becomes a tool for enhancing, not substituting, human intellectual capacity.

Where AI genuinely helps in a research paper workflow:

• Brainstorming research questions and outlining structure
• Summarizing a single paper's argument or methodology
• Identifying recurring themes across a defined corpus
• Polishing grammar, clarity, and formatting in your own draft

Where it fails, and where you need to stay in the driver's seat:

• Synthesizing conflicting findings across multiple studies
• Generating or verifying citations (hallucinated references are a documented problem)
• Making methodological or interpretive judgments
• Reasoning through novel arguments that require domain expertise

The benchmark data tells the same story. AI capability is climbing fast — MMMU scores rose 18.8 percentage points in a single year, and inference costs for GPT-3.5-level performance dropped over 280-fold between late 2022 and late 2024. Yet the same Stanford AI Index warns that models "often fail to reliably solve logic tasks even when provably correct solutions exist."

For researchers, the practical takeaway is simple: let AI handle the mechanical layers of reading and writing, and reserve the analytical core for yourself. The productivity gains are real, but only when paired with rigorous validation.

The 5-Step Workflow: Plan, Select, Execute, Validate, Disclose

A peer-reviewed framework from MDPI turns AI-assisted research into five deliberate steps: plan, select, execute, validate, disclose. Skip any one of them, and you inherit the risks that dominate the headlines.

Planning comes first, and it means mapping every AI touchpoint before you open a chat window. Decide where AI enters — idea generation, structuring, literature synthesis, editing — and where it never does. A systematic review of 24 studies identifies six domains where AI supports academic writing, and ChatGPT specifically shows substantial potential across them. The same review notes persistent challenges around academic integrity and human-AI balance, so the plan should name your verification rules up front.

Selection means matching tools to tasks, not chasing the newest model. Evidence shows AI excels at single-article comprehension and theme identification but hits current limitations in complex data synthesis and inter-paper comparison. Reserve those tasks for human judgment:

• Idea generation and outlining — strong AI fit
• Single-paper summarization and theme spotting — strong AI fit
• Drafting and line editing — strong AI fit with review
• Cross-paper synthesis and final argument — human-led
• Citation verification — always human-led

Execution is iterative, not one-and-done. Draft with AI, revise with AI, but keep a human in the analytical loop at every pass. The MDPI framework warns that habitual reliance may erode critical thinking and create confirmation bias — its recommended posture is AI as a tool for enhancing, not substituting, human intellectual capacity. Adoption pressure is real: 78% of organizations reported using AI in 2024, up from 55% in 2023, according to the Stanford HAI AI Index.

Validation is mandatory. Fabricated citations are documented in the wild — computer scientist Guillaume Cabanac found his work cited in a dental journal paper he had no connection to. Cross-reference every AI-generated claim against the primary source before it enters your manuscript.

Disclosure closes the loop. Detection tools are unreliable enough that some institutions have switched them off, so transparent self-disclosure — not detector output — is the practical ethical standard.

Agent skills can automate parts of this workflow — browse and test free ones on SkillzBench, which is building free benchmarking, security scanning, and a free MCP connector for agents.

Verify Everything: Catching Hallucinated Citations Before Reviewers Do

When computer scientist Guillaume Cabanac discovered his own name attached to a reference in a dental journal paper he had never touched, he told reporters he "couldn't recognize my own citation." It's a small moment that captures a large problem: AI tools fabricate references convincingly enough that even the cited authors don't notice.

This is why validation isn't an optional cleanup step — it's step four of the peer-reviewed five-step framework for AI-assisted literature work, sitting deliberately between execution and disclosure. The framework's authors are blunt about the stakes: hallucinations can include fictitious citations, invented facts, or misleading information that undermines a paper's credibility and reliability.

Don't trust AI detectors to do this work for you. Detection tools have proven unreliable enough that some institutions have switched them off entirely, citing algorithm biases, the growing sophistication of AI-generated text, and easy content obfuscation. The burden of verification falls on the author, and it should be treated as a mandatory pass before anything leaves your desk:

• Open every citation the AI produced and confirm the paper actually exists, the authors match, and the quoted or paraphrased claim appears in the source.
• Cross-reference AI-generated summaries against the original articles — AI performs well at single-paper comprehension and theme identification, but shows documented limitations in complex cross-paper synthesis, exactly where errors hide.
• Trace any statistic or factual claim back to a primary source before repeating it, since invented facts are a recognized hallucination category.
• Check whether cited sources support the specific claim you're making, not just a nearby topic — the Cabanac case shows citations can look plausible while being entirely wrong.

The timing matters because the volume of AI-assisted writing is climbing fast. An analysis of nearly 2.1 million preprint abstracts found that researchers using generative AI publish significantly more studies, and reviewers are increasingly in the same position — more than 50% of surveyed academics report having used AI while peer reviewing manuscripts. Every unverified citation that slips through now has more chances to propagate.

There's a second reason to keep humans firmly in the analytical loop during verification: overreliance. Research on AI in academic writing warns that habitual dependence on AI content can erode critical thinking and create confirmation biases — the very faculties you need to spot a fabricated reference. The recommended posture treats AI as a tool for enhancing, not substituting, human intellectual capacity.

Finally, close the loop with transparency. Since detectors are unreliable, transparent self-disclosure is becoming the practical ethical standard: state plainly how AI was used in drafting, editing, or synthesis, per your target journal's policy. Disclosure plus verification is what separates AI as a legitimate productivity tool from the undisclosed use now raising integrity concerns across publishing — and it's the difference between a reviewer trusting your paper and flagging it.

For teams building these verification steps into repeatable agent workflows, SkillzBench is building a free directory of agent skills with security scanning and benchmarking, plus a free MCP connector — worth a look if you want your AI pipeline to be as trustworthy as your citations.

Automating the Workflow: From Manual Steps to Agent Skills

Manual steps in a research workflow — tracking new literature, reformatting references, chasing down missing DOIs — consume hours that could go to actual analysis. AI agents and agent skills (reusable instruction packages that plug into coding assistants and chat agents) can automate much of that grunt work, but they introduce a new variable worth taking seriously.

An agent skill can watch a preprint server and flag relevant papers, reformat a bibliography to APA or Vancouver style, and cross-check every citation against a live database. These are exactly the tasks where AI-assisted review research shows tools perform reliably well — single-document comprehension, theme identification, and mechanical formatting — while leaving cross-paper synthesis to you.

The catch: skills are code-adjacent, and roughly one in four carries a real flaw. About 26% of published agent skills contain at least one security vulnerability, which matters when a skill is parsing your unpublished manuscript or touching your reference library. Treat skill selection the way you'd treat a lab protocol — vetted before use, not after.

A practical shortlist for automating a research-paper workflow:

• Literature tracking — a skill that monitors arXiv, PubMed, or SSRN and summarizes new abstracts against your keywords each week
• Citation management — a skill that validates every reference against Crossref or DOI resolvers before it reaches your bibliography
• Formatting and style — a skill that reflows headings, tables, and in-text citations to a target journal's template
• Disclosure drafting — a skill that generates a transparent AI-use statement aligned with the five-step AI-assisted review framework

Vetting is the hard part. A skill's GitHub stars tell you nothing about whether it hallucinates references or leaks prompt context — the same hallucination risk that produced the documented case of a fabricated citation attributed to an uninvolved researcher. You want security scans, pinned versions, and side-by-side benchmarks showing what the skill actually improves.

That's the problem SkillzBench is building to solve. Its free directory crawls and deduplicates skills from GitHub, pins each one to a commit SHA, runs security and license checks, and benchmarks every skill in a sandbox — same model, with and without the skill — to publish a Lift score in percentage points. A free MCP connector then lets you load only the skills you've chosen into Claude Code, Cursor, Codex, or Gemini CLI.

The directory, benchmarks, and connector are all free; the waitlist is live now while the platform is in pre-launch. If you want to extend the workflow above without guessing which skills are safe to run, that's the place to start.