RAGs, Reasoning and Deep Research: What’s new in AI and what might it mean for teaching in 2025?

It's been two years since ChatGPT kicked off the current artificial intelligence (AI) boom, and got all of us talking about how AI affects learning and teaching practice. The conversation has shifted, from some initial calls for an outright ban of AI to defend academic integrity, to now recognising the need to integrate AI across curricula to ensure students build skills in responsible AI use as part of their academic development.

While most of us have probably used OpenAI's ChatGPT, Microsoft's Copilot or maybe Google's Gemini, there are now at least 20 major gen AI models available. When you include all the model variations, open-source options and offerings from smaller players, that number grows to thousands. While the first wave of consumer generative AI technology remains impressive, new AI features released in recent weeks will have changing impacts for university learning and teaching in 2025.

Here are some of the active AI development areas which hint at what might be on the horizon.

Agentic AI and Deep Research

While first-wave gen AI systems were reactive and individual task focused, agentic systems go further and exhibit a degree of autonomy and goal-directed behaviour. 

A recent example of agentic AI is Deep Research, a new feature in Google’s Gemini Advanced (paid version) released in December. When a user submits a question, it generates a multi-step research plan, then dispatches AI agents across multiple online sources. Several minutes later, it returns with a neatly compiled 6–10 page report, complete with references.

Its outputs are interesting, but likely won't make researchers fear for their jobs any time soon. People have noticed how Deep Research can miss vital details and doesn’t discern well between low and high-quality sources and out-of-date information—as it says in fine print “Gemini can make mistakes, so double-check it”.

OpenAI earlier this month released its own competing Deep Research feature for ChatGPT Pro subscribers in the US. Initial reviews suggest it might outperform Google’s version, with one researcher saying it’s “extremely impressive” for literature reviews. However, the US$200 per month subscription cost will likely keep it out of reach for most students, for now.

Retrieval augmented generation (RAG)

By uploading a collection of documents, users can provide an AI model with a reference for generating content. Known as retrieval-augmented generation (RAG), this can improve the accuracy and relevance of AI responses by grounding them in supplied information, rather than relying solely on the model's general training data. RAG isn’t new, but it’s now being implemented in consumer AI tools in interesting ways.

My favourite is NotebookLM, which Google also describes as a personal AI research assistant, but it's very different to Deep Research because it only works with material uploaded by the user. Its simple interface has three main windows: source documents on the left, AI chat in the centre, and outputs on the right.

NotebookLM is geared toward assisting users in exploring, interacting and learning from their research material. Give it up to 50 academic papers or reports (PDFs), web URLs, audio (mp3), even handwritten notes, and it will generate summaries, explanations, FAQs, study revision guides, and basic applied analysis (gap, comparative, sentiment). Understandably it’s gaining popularity among the Australian academic community.