Scholarlink

<p>Large language models (LLMs) can accelerate academic writing, but unmanaged use can weaken evidentiary traceability, blur accountability, expose confidential material, and introduce unsupported claims or misleading figures. This article presents Five-Phase Writing (FPW), a tool-agnostic, human-governed framework for ethically responsible LLM-assisted research writing. FPW structures work into evidence discovery and verification; argument-led drafting; figure development and integrity checking; critical revision and responsible paraphrasing; and language polishing, disclosure, and submission audit. Each phase produces an auditable artefact and ends with a human-controlled exit criterion. The framework was evaluated using 120 paired writing tasks across health and life sciences, science, technology, engineering and mathematics, and social sciences and humanities. Relative to the conventional workflow, FPW reduced completion time by 5.84&#xa0;h (95% confidence interval − 6.24 to  −  5.43), improved manuscript quality by 6.37 points (95% confidence interval 5.79–6.96), increased evidence traceability by 16.19 points (95% confidence interval 15.22–17.17), reduced verification errors by 35.3%, and increased disclosure compliance by 25.0 percentage points. Perceived novelty was unchanged (<i>p</i> = 0.080). A component ablation analysis showed that median standardised utility declined from 1.34 for complete FPW to 0.43 when verification was removed and 0.14 when assistance was limited to polishing. The findings support FPW as an evidence-based governance framework that links research efficiency to accountability, transparency, non-maleficence, privacy, and human agency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Five-phase writing (FPW): an empirically evaluated framework for ethical, traceable, and human-governed research writing with large language models

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Hamza Shahbaz,
  • Badar Dad,
  • Safdar Saud

摘要

Large language models (LLMs) can accelerate academic writing, but unmanaged use can weaken evidentiary traceability, blur accountability, expose confidential material, and introduce unsupported claims or misleading figures. This article presents Five-Phase Writing (FPW), a tool-agnostic, human-governed framework for ethically responsible LLM-assisted research writing. FPW structures work into evidence discovery and verification; argument-led drafting; figure development and integrity checking; critical revision and responsible paraphrasing; and language polishing, disclosure, and submission audit. Each phase produces an auditable artefact and ends with a human-controlled exit criterion. The framework was evaluated using 120 paired writing tasks across health and life sciences, science, technology, engineering and mathematics, and social sciences and humanities. Relative to the conventional workflow, FPW reduced completion time by 5.84 h (95% confidence interval − 6.24 to  −  5.43), improved manuscript quality by 6.37 points (95% confidence interval 5.79–6.96), increased evidence traceability by 16.19 points (95% confidence interval 15.22–17.17), reduced verification errors by 35.3%, and increased disclosure compliance by 25.0 percentage points. Perceived novelty was unchanged (p = 0.080). A component ablation analysis showed that median standardised utility declined from 1.34 for complete FPW to 0.43 when verification was removed and 0.14 when assistance was limited to polishing. The findings support FPW as an evidence-based governance framework that links research efficiency to accountability, transparency, non-maleficence, privacy, and human agency.