Microsoft Copilot - supporting research

 This article, was published mid-2025, providing an overview of the impact of Copilot researcher and analyst agents on work. 

Having now used both for 6 months plus, my take on these two agents, is that they can be useful, but need to used with care. 

1) working out what is to be done is important. Responses from researcher or analyst agents, arise through the prompts provided. Careful structuring of prompts is therefore essential in obtaining the types of actions required. Otherwise, researcher may go off topic quite easily.

2) providing context is important and delimiting the researcher/analyse agents to specific papers, websites or attachments help ringfence the direction of responses.

3) using the prompt writing agent in Copilot can be useful for tightening prompts.

4) ensuring that 'work or 'web' is selected, along with depth of responses (auto, quick, think deeper) helps again to ensure responses fit expectations.

5) Triangulation is always required, to check the validity of the responses. Limiting resources also help to save time, so that triangulation is restricted and does not have to go all over the place.

6) Draw on the advantages of AI. Summarising key points, comparison of key points across papers/sites, drilling deeper to extend insights, providing 'neutral' perspectives on conceptualisations, frameworks etc.

7) be aware of 'cognitive debt' / 'cognitive atrophy' and how this can come about very easily when your critical thinking is replaced by AI doing the work. You still need to read deeply, cogitate, make your own judgements and come up with your own synthesis. Then use AI to cross check these and see if it provides other insights which are viable.

8) Continue to learn how to manage AI to draw on it's ability to automate some research processes. The key is to use AI as a thinking partner, not to replace your own effortful thinking.

Using AI to support research - Dr Fiona Beals

 Notes from a lunch time presentation by Dr. Fiona Beals (Principal Lead on teaching, innovation and research at - Weltec/Whitireia/Te Pūkenga) on Research & AI: When your friend is your foe.

Welcome from Dr. Lee Smith who looks after the section. 

Presentation began with a overview of the landscape right now. There are a whole raft of resources. What you get from AI, is determined but what data it draws on. Prompts are very important, an open prompt does not usually yield useful outputs. Referencing of AI is fraught as it is not possible to go back to the AI to get the same output - using AI is more like an communication. There have been articles written that names AI as an author but this is not really legal as AI is not a person and unable to take responsibility for their share of the work. It works on a predictive algorithm and cannot predict the unpredictable. Predicated on data and critical literacy on the part of the user.

Moved on to do a show and tell with some of the AI tools which are usable for research. Moved from the ones which she has found to not be relevant/applicable, to the one she now uses most.

1) Perplexity - based on ChatGPT. Described as a search engine on steroids. It generates summary and provides a range of 'references'. Note that with ChatGPT, only mines data before 2021, so not always up to date. Triangulation takes time and feels the output is not as good as what is produced by scholarly databases.

2) Paperbrain - currently does not work. and an example of the rapid flux in the landscape with tools coming up and disappearing.

3) PaperDigest - has the functions to do lit review or be a text rewriter. In the text rewriter, the functions include wikify & justify, paraphase, quesionly and summarise. The lit review searches for papers and provides a summary. Only finds literature that is accessible on the internet, not those behind paywall. Therefore, limited lit review rather than comprehensive.

4) elicit - this is the one she recommends to be useful. Demonstrated how she uses this to overview a topic. Ability to collect and overview many articles quickly. Scholarly database will still yield more reliable results. 

There are Ai tools that will allow up of pdf articles but need to be careful with regards to copyright as once the pdf is uploaded, the database will have access.

5) Litmaps - able to pick up from your ORCID account, your work and trace the whakapapa of your or a piece of work. Provides a visual representation of how various papers are related and bring up patterns of themes.

Demonstrated how ChatGPT could be used to help modify an existing spreadsheet to change its function for purposes of quantitative analysis. Prompt engineering is a key, along with existing knowledge of concepts. Without that, it is difficult to evaluate/ triangulate / or verify the AI generated outputs for accuracy. 

Demonstrated how to use BingChat. Firstly uploaded paper/book via EBcoHost database search and ask BingChat to summarise the pdf/page, or as questions of the content and use it to help work out overall gist of the article. Might not work if database has block in it to prevent articles being gleaned. 

Concluded with caveats including the ethical aspects. Important NOT to enter own personal data or to ask it to do data analysis with it! Also remember the implicit biases etc. that are present in the data sources. 

Good session to see how others use AI tools to support research, especially these initial efforts to test and understand how to use these tools to help our work. 

UNESCO report - AI, ChatGPT and higher education

 A general introduction to AI and ChatGPT released in April by UNESCO.

Provides  'start up' instructions and then progresses into suggestions on how to leverage ChatGPT to encourage critical thinking, along with implications. Table in page 9 and some suggestions for its use to support research are useful.