Generative AI in Public Policy Research

The most widely discussed ethical risks include data privacy and confidentiality (especially when files are uploaded for analysis), hallucinations and fabricated references, model-propagated biases, insufficient traceability of sources, and the impact on reproducibility. Recent editorial policies remind authors that AI outputs may appear authoritative yet contain inaccuracies; it is the author’s duty to verify and properly cite the literature—never the AI itself—as a source (ICMJE, 2024; Nature, 2025). Recommended mitigation measures include: using only anonymized or authorized data; opting for enterprise environments with data safeguards when necessary; requiring search and citation logs; recording versions and prompts; and submitting generated excerpts to human peer review and similarity-checking tools.

Among popular tools, ChatGPT stands out for its data-analysis capabilities with file uploads, table and chart generation, and its ability to perform exploratory data analysis within the environment (“Advanced Data Analysis”). In 2025, OpenAI also introduced enhancements to Search within ChatGPT and announced Pulse, an experience that conducts proactive research and delivers updates based on user preferences—useful for preliminary literature screening and hypothesis scouting, which must later be verified in primary sources (OPENAI, 2025). In scientific research, ethical uses include drafting protocols and checklists, outlining analysis plans, generating annotated code, and suggesting argumentative structures, always with human oversight and citations to original sources (never to the model itself).

Perplexity was designed with an emphasis on search with citations, and more recently introduced a Deep Research mode that performs multiple queries, reads hundreds of sources, and produces reports with full references, enabling traceability of the information path—valuable for rapid reviews and exploratory mappings. The platform also offers a Pro/Research mode and a library for organizing findings (PERPLEXITY, 2025). Responsible use entails: (i) inspecting all references provided; (ii) excluding unreliable websites; (iii) grounding conclusions in peer-reviewed articles and official documents; and (iv) documenting all filters and eligibility decisions applied by both the system and the researcher.

Gemini integrates natively into the Google ecosystem, featuring multimodal chat, Deep Research, and incorporation into Workspace (Gmail, Docs, Sheets, Drive), including enterprise functionalities with expanded rollout in 2025. For research, this means workflow integration across documents and spreadsheets, PDF summarization, guided notes (NotebookLM/Audio Overviews), and context-aware collaboration—useful for managing literature reviews and experimental data (GOOGLE, 2024–2025). From an ethical standpoint, it is essential to configure institutional data policies (e.g., disabling training outside contractual scopes) and maintain reference repositories with stable document versions.

Grok (xAI) distinguishes itself through Live Search and integration with real-time data from the X platform (formerly Twitter) and web sources, claiming strong real-time retrieval capabilities. In scientific contexts, it may be useful for trend monitoring, identifying preprints mentioned on social media, and tracking emerging public policy debates—provided that credibility criteria and triangulation with official or peer-reviewed sources are applied. Because social networks contain noise and misinformation, researchers should define source-eligibility rules and maintain logs of queries and results (xAI, 2025).

A prudent protocol for using generative AI in research can be summarized as follows:

1. Clarity of purpose and alignment with UNESCO/OECD ethical frameworks;
2. No AI authorship and explicit disclosure of use in accordance with COPE/ICMJE/editorial policies;
3. Data protection and risk assessment before any upload;
4. Source traceability (logs, links, snapshots, DOIs) and manual verification;
5. Reproducibility (saving prompts, versions, seeds, and scripts); and
6. Critical evaluation of bias and societal impact.

Tools such as ChatGPT, Perplexity, Gemini, and Grok can accelerate parts of the research cycle, but they must serve scientific integrity—never replace it.