The seven shifts of the last 6 months that most change how lecturers can use AI in their research work.
AI research tools have moved from 'autocomplete with citations' to genuinely agentic collaborators in under two years. For lecturers, the practical question is no longer whether to use them, students and reviewers already do, but how to use them rigorously enough to defend in a methods section, a viva, or a peer review.
GPT-5.5, Gemini 3.1 Pro, and Claude's extended-thinking modes (Opus & Fables) now plan before they answer. And all now comes with the size of 1 million tokens. But the trade-off is latency and cost (limits). Use reasoning models for synthesis, critique, and methods work; use fast models for triage and reformatting.
Gemini Deep Research, ChatGPT's research mode, and Claude's research-grade Projects can plan a multi-step investigation, browse 50+ sources, and return a structured citation-backed brief. This is the largest single productivity shift for literature work since Google Scholar.
A new product category: tools that refuse to answer beyond the sources you upload, with click-through citations to the exact passage. Hallucination drops sharply. Best for assigned-reading work, focused literature reviews, and qualitative analysis where provenance is non-negotiable.
Frontier models now read PDFs (including math), interpret figures, transcribe long-form video, and reason about tables. This unlocks systematic review of figure-heavy literatures, automated alt-text for accessibility, and analysis of qualitative video data. Workflows that were impractical 6 months ago.
Elicit, Consensus, SciSpace, Undermind, and Research Rabbit specialise in scholarly search with claim-level evidence. They complement, not replace, generalist chatbots. Use them when you need traceable claims across a defined corpus rather than open-ended synthesis.
Phi, Gemma, and Llama variants run locally on a modern laptop. For sensitive data (interview transcripts, unpublished drafts, anything covered by ethics approval), local models are increasingly viable for first-pass coding, summarisation, and redaction.
The skill that separates good and great AI users is no longer prompt wording. It is what context you assemble around the model: which sources, which examples, which role, which output schema, which refusal rules. The rest of this workshop is, in effect, applied context engineering for three tools.