Session 3 Recap: AI for Literature Review and Research Discovery

From Tools and Projects to Research Workflows

Our third AI Academy session turned from foundations to practice, taking on one of the most universal tasks in academic life: the literature review. Building on Session 1's workflow thinking and Session 2's persistent knowledge bases, we asked a more pointed question: what does it look like to integrate AI responsibly into the discovery, prioritization, and deep engagement stages of scholarly research?

By early 2026, AI is no longer a peripheral aid for drafting or proofreading; surveys from Elsevier and Wiley both report that a majority of researchers now use AI tools regularly, with Wiley's number jumping from 57% to 84% in a single year. The question is no longer whether to engage, but how to do so with rigor.

The Semantic Discovery Landscape

We toured the category of specialized, AI-enhanced literature tools that have matured over the past year: Elicit for structured extraction from Semantic Scholar, ResearchRabbit for citation network visualization and iterative chaining, Consensus for empirical "yes/no" questions with its signature consensus meter, and Paperguide as an all-in-one assistant. The common thread is semantic discovery: instead of Boolean keyword matching, these platforms use vector embeddings to retrieve conceptually relevant work even when terminology varies across disciplines.

Alongside these specialized tools, the major AI providers have rolled out their own deep research capabilities: OpenAI Deep Research, Gemini Deep Research, Claude's research mode, and Perplexity Deep Research. These are not literature tools per se; they are general-purpose agentic research systems that can autonomously plan, search, backtrack, and synthesize over dozens of steps.

When should you reach for which? The answer depends on where you sit in the research process. Specialized tools excel at systematic, reproducible searches and structured data extraction across many papers. Deep research agents excel when you are exploring a new area and do not yet know what to search for, or when your question spans heterogeneous sources: academic papers, preprints, technical documentation, industry reports.

A Hybrid Workflow

The recommended workflow we walked through is iterative and hybrid:

NotebookLM earned particular emphasis here. Because it is grounded exclusively in the sources you provide, responses come with inline citations and effectively zero hallucinations. Its recent integration with Gemini Gems is a genuine step forward: you can now attach an entire NotebookLM corpus to a Gem without re-uploading, dynamically improving the Gem with a living knowledge base.

Discussion: The Wicked Question

The conversation that followed the presentation was the real heart of the session, and it opened up the tension that every faculty member in the room is wrestling with.

Erica raised what Professor Barba called "the wicked question": with these tools, a student could write a competent literature review without understanding anything in it. From a doctoral mentoring perspective, the problem is acute. A lit review is not just a deliverable; it is how students come to know a field: who is working on what, which threads of argument matter, where the disagreements lie. That understanding is forged by struggling through the actual language of the papers. Can the tools deliver their benefits without eroding that deeper learning?

Professor Barba connected this directly to an earlier exchange about a draft policy in biomedical engineering that had proposed banning AI tools from qualifying exams outright. Her position was unambiguous: as a researcher, she wants to use these tools and wants to teach students to use them fully and responsibly. Banning is a knee-jerk reaction; the harder, more honest task is designing for responsible use.

Several threads emerged from the group. One proposal was to assess process rather than product: ask students to document and articulate how they moved from 350 candidate papers down to five, what judgments they made at each stage, and why. That articulation is something the tool cannot supply. Another suggestion was to lean on oral examinations and demands for novelty as complementary checks. A third was the arithmetic analogy: we all use calculators now, but we learned long division first. Is there an equivalent foundational phase for literature work, and if so, what does it look like?

A parallel view, offered from the capstone teaching experience, pressed the other direction: the power these tools bring is simply too great to look away from. Undergraduates who graduate without AI fluency will be behind the curve. The answer is both: automate and amplify the discovery phase, then roll up your sleeves and read the papers that matter. The threshold where automation ends and deep reading begins is where the real learning lives.

The paywall problem surfaced too, as a practical limit on what any of these tools can do. None of the current deep research agents can reach behind institutional subscriptions, which excludes much of the journal-of-record literature in most engineering fields. This remains an open gap, and one worth pushing vendors and libraries to close.

Tools of the Trade

A Note from the Broader Conversation

A colleague reported briefly on a university-level AI discussion held the day before, which had stayed largely at the level of "what can you do with the chat interface?" The contrast with what we do in the Academy, Prof. Barba noted, is exactly the point: moving beyond the chatbox into workflows, knowledge bases, and agentic tools is where the leverage is. Other institutions are asking the same questions about assessment in the age of AI, and nobody has clean answers yet. We are, collectively, stumbling toward wisdom.

The GW Engineering AI Academy is a strategic initiative to position SEAS as an AI-forward institution through systematic faculty development, anchored in the Entrepreneurial Mindset framework of our KEEN partnership.