How Generative AI supports research: from prompts to platforms
Research AI bite: 03.
Key takeaways:
GenAI is increasingly integrated into researchers’ workflows, progressing from simple prompting to embedded assistance, grounded retrieval, agentic coordination, and ultimately toward fully integrated, continuous research workflows.
Structured protocols playing a key role today in enabling clarity, consistency, and reproducibility.
Generative AI (GenAI) is becoming an integral part of the research ecosystem; not by replacing researchers, but by enhancing how they find, analyse, write, and collaborate. The real impact of GenAI is not its novelty, but in its ability to integrate with research workflows, data infrastructures, and standards of reproducibility and evidence.
In my first research AI bite, I warned that GenAI tools can be surprisingly unstable: behaviour may change without warning, like a key suddenly reshaping so it no longer opens your door. These systems are trained on data we give them, and so they systematise our assumptions and amplify our biases. In the second bite, I showed how Dimensions has been built on early implementations of AI, playfully testing our classifiers on Disney songs.
Still, with the right safeguards and clear expectations, GenAI can be more valuable than feared. This post outlines five roles GenAI plays in research today, illustrated with selected tools, literature, and examples (including some of our own work) with more detailed posts to follow.
1. Prompt engineering for research
Most researchers begin here, conversing with a chatbot.
Unstructured prompting
GenAI such as ChatGPT, Gemini, or Claude respond to open-ended questions and commands, offering a simple, informal, and rapid way to get initial insights.
Use cases: Brainstorming questions, summarising research, narrowing searches.
Pros: Flexible, rapid, and intuitive. Some versions are free.
Cons: Not reproducible; prone to hallucination, so definitive answers should be double checked, voluntary and involuntary biases should be expected, and answers should be crossed checked. Latest models are paid.
Examples
The Prompt Report (2024) categorises 58 prompting methods into a usable framework,
Generating visual abstracts,
Summarising or comparing findings across studies,
Assisting in meta-analyses or synthesis tasks.
Structured prompting via protocols
Prompts can be formalised into templates or shared formats, increasing clarity and consistency. I have built a few of these, mostly to classify research related items, which I will soon share in here.
Use cases: Writing support, systematic review screening, query building, or classification. Many protocols should be run through an API to maximise impact.
Pros: Repeatable and auditable.
Cons: This can be expensive and time consuming, biases cannot be fully removed.
Examples
A recent paper in Business Horizons introduced a structured Generative AI Prompt Protocol, grounded in constructivist learning theory,
Smartseeds, our Classification builder (more on this one soon),
Journalscape, our journal classifier (more on this soon too).
2. Embedded AI assistance
These are AI tools embedded directly into research software, offering quiet, task-specific support in existing tools.
Use cases: Improving writing, summarising papers, checking tone or grammar
Pros: Seamless integration; no new interface required
Existing tools
Writefull for Overleaf: Grammar, clarity, and tone feedback in LaTeX
AI Reader Assistant in Papers: Summaries, peer-review checks, and translation for PDFs
AI-assisted query builder in Papers: it turns natural language into structured search queries.
3. Retrieval-Augmented Generation (RAG)
Unlike free-form AI that was trained on general knowledge, RAG systems are trained on domain-specific content, reducing hallucinations and increasing trust.
Use cases: Literature Q&A, evidence synthesis, knowledge discovery
Pros: Factual grounding; transparent references
Existing tools
Dimensions Research GPT: Combines Dimensions' scientific evidence with ChatGPT to answer questions using real citations. There is a free version available for ChatGPT paid users and trained on open research, while the enterprise version is available to Dimensions customers only and trained on all full text available in Dimensions,
Semantic search in ReadCube Papers: Uses embeddings to retrieve conceptually similar articles.
4. Agentic systems
GenAI agentic systems are early-stage tools that help researchers manage multi-step tasks by tracking goals and context, while remaining user-guided and modular.
Use cases: Literature reviews, iterative project development, multi-stage query building, writing across sections,
Pros: Context-aware; increasingly task-aware
Emerging tools
These are not fully autonomous agents yet, but they scaffold continuity, modularity, and user-guided iteration:
DimQuery Assistant (custom GPT, beta version, contact me to become a beta tester): GPT trained on the Dimensions GBQ schema and refined through reinforcement learning. While not fully autonomous, it bridges natural questions and technical schema logic; helping researchers generate structured SQL with context awareness.
DimDSL Assistant (custom GPT, alpha version, contact me to become an alpha tester): GPT trained in Dimensions DSL, supporting writing python and DSL API queries for bibliometric studies. Still not fully autonomous, but it builds the query based on natural language and is bibliometric-aware.
Gems (Gemini): Similar to DimQuery Assistant but for Gemini users.
Artifacts (Claude): Similar to DimQuery Assistant but for Claude users. These can have a standalone interface.
In the literature
A 2024 paper titled Agentic Workflows for Economic Research proposes using LLMs and multimodal AI agents to autonomously conduct literature reviews, data analysis, and interpretation; highlighting efficiency and reproducibility benefits,
The Academy middleware system enables autonomous agents to operate across federated research infrastructures; like HPC clusters and institutional repositories—supporting asynchronous, modular execution of scientific workflows,
ProAgent and the OpenAgents Toolkit showcase early-stage, modular agent orchestration, capable of planning and adapting workflows from user prompts to results.
5. Towards AI-integrated research workflows
GenAI is moving beyond isolated tasks toward integration across full research workflows, supporting continuity from question formulation to synthesis and communication. The aim is a system that maintains memory, coordinates across stages, and reduces fragmentation, making research more reproducible and scalable.
Examples
Multi-stage literature reviews with memory of previous iterations,
Semantic coordination across search, annotation, writing, and visualisation,
Persistent workspaces that track progress, hypotheses, and decisions,
AI agents that propose next steps, suggest variables, or validate models.
In the literature
The Agentic RAG model extends traditional RAG by letting agents dynamically choose retrieval strategies and refine results across iteration, which is ideal for complex review tasks.
Conclusion
Generative AI is reshaping research, not through a single tool, but through a growing ecosystem that spans prompting, analysis, and infrastructure. The question is no longer if we use GenAI, but how. This is not just a technological shift, but a cultural one. To ensure GenAI strengthens rather than undermines research, we must shape it with intent; from the prompt to the platform.