AI Tools for Population Researchers

This guide was created in September 2026 for the BRIDGE workshop. Because AI tools change quickly, some of the information here may have changed or become outdated by the time you read it. If you notice anything like this, please email the team at cand3 [at] mcgill.ca with your suggestions, so we can keep the guide up to date and useful for the network.

“AI Tools for Population Research” was prepared by Clara Bolster-Foucault, CAnD3 training program facilitator and PhD candidate in Epidemiology at McGill University. Clara created the guide for Bridging Minds: AI Literacy for the Next Era of Research. That event was the BRIDGE proof-of-concept meeting, held in Montréal on September 10–11, 2026, and it focused on critical engagement with generative AI in research. The guide is meant to help graduate students and researchers in the CAnD3 network use AI tools thoughtfully and responsibly in their own work.

The AI Tool Overview table below provides an overview of tool performance across 7 key tasks, using icons to indicate the suitability of each tool for a given task: 

• ✓ : the tool is designed for this task and does it well
• ~ : the tool can do this task but another tool on the list does it better
• ✗ : the tool is not suited to this task

AI Tool Overview

Please note that prices, model versions, and data handling terms change frequently; we recommend confirming current terms with each platform before proceeding.

1. General-purpose assistants

These are conversational assistants built on large language models, offering a single interface for drafting and editing, summarizing and reasoning over long documents, writing and running code, and answering questions about material you provide. They are well-suited to administrative tasks and are a typical entry point for researchers new to integrating AI into their workflow. None of the tools in this section has direct access to a scholarly index, which makes them poorly suited to literature discovery.

These systems generate plausible text which may contain plausible errors, so numbers, citations, dates, and quotations require verification against verified sources. The consumer and institutional versions each product also differ in how input is handled and retained, so it is important to confirm which account you are signed into before entering anything confidential.

Claude

Oriented toward sustained work on documents and code. Long inputs are the main strength. Full manuscripts, a set of documents, or a codebase can sit in one conversation and be reasoned over together.

• Projects hold a fixed set of documents and instructions that persist across conversations, which supports ongoing work against a stable set of sources.
• Artifacts render drafts, tables and code beside the conversation, where they can be revised in place rather than regenerated each time.
• Runs code in a sandbox and produces files directly, including Word, Excel and PowerPoint output.
• Connects to Google Drive, Gmail and Calendar, and to other services through the Model Context Protocol.
• Claude Code, covered in section 4, applies the same models to work in a terminal or editor.

Cautions: Multiple models are integrated into a single interface, and users must switch between them. Model choice affects skillset and quality of output. Long context invites uploading whole document sets and users should consider confidentiality constraints. Web search draws on the open web rather than a scholarly index.

ChatGPT

A general-purpose assistant covering drafting, analysis, image generation, voice and scheduled work through one interface.

• Deep research compiles long referenced reports from web sources over several minutes.
• Data analysis runs Python on an uploaded spreadsheet and returns figures and cleaned files with nothing installed locally.
• Custom GPTs package instructions and reference files for reuse or for sharing with colleagues.

Cautions: Multiple models are integrated into a single interface, and users must switch between them. Model choice affects skillset and quality of output. References in deep research outputs come from web sources of unverified quality and are not filtered for peer review.

Gemini

A general-purpose assistant that sits inside the Google Workspace applications, so it operates on documents in Drive, threads in Gmail and calls in Meet without anything being uploaded. Very large inputs are handled well, which suits long reports and large document sets.

• Available in the side panel of Docs, Sheets, Gmail and Drive, working on the file already open.
• Handles PDFs, images, audio and video in the same conversation as text.
• Deep research compiles long referenced reports from web sources.
• Canvas provides a working surface for drafts and code beside the conversation.

Cautions: It can be challenging to navigate between personal and institutional Google accounts when signed in on the same machine. Integration also removes the moment of deliberate upload into a chat or app, so material can be processed without the user’s awareness.

Gemini Notebook

A research assistant restricted to sources deliberately added to a notebook. Answers are drawn only from the uploaded set of sources and responses have inline citations pointing to the passage it was sourced from, which facilitates verification.

• Holds a large collection of documents, slides, web pages and video transcripts as a queryable body.
• Answers only from the sources provided, and states when the material does not contain the answer.
• Generates study guides, briefing summaries and timelines across the whole collection.
• Produces an audio overview, a synthetic spoken discussion of the source material.

Cautions: grounding in the uploaded sources removes fabricated references but not misreading, so a citation can point to a real passage that does not support the claim attached to it. Nothing outside the notebook is consulted, which means gaps in the uploaded set become gaps in the answer without being flagged.

Microsoft Copilot

A general-purpose assistant embedded in the Microsoft 365 applications, which can run inside an institutional tenancy so prompts and files remain within agreements the institution already holds. Copilot has access to institutional content, and can use content from emails, files, chats, and meetings in the user’s account as well as content from the open web.

• Works inside Word, Excel, Outlook and Teams on the file or thread already open.
• Searches across institutional content, which makes it useful for retrieving where something was decided or who wrote what.
• Produces meeting recaps in Teams, covered in section 2.
• Drafts and summarizes in the same interface as the surrounding work, with no upload step.

Cautions: Integration with Microsoft account data everything the account has permission to open, which in most organizations is broader than people assume, so material can appear in a summary that the user did not know was reachable.

Perplexity

A search interface that returns a written answer with numbered citations to the pages it drew from. Answers are assembled from live web results at the time of asking, which makes it suited to current information and to questions where the source matters as much as the answer.

• Every claim carries a numbered link, so the underlying page is one click away.
• Follow-up questions retain the thread of the search.
• An academic mode restricts results to indexed papers.
• Spaces keep collections of searches and uploaded files together on a topic.

Cautions: the academic mode is a filter over web search rather than a substitute for a literature database, and its coverage and ranking are not documented well enough to support a systematic search. Citation to a page is not verification of the page, and low-quality sources are cited in the same format as good ones.

2. Meeting notetakers

These tools record meetings, transcribe the discussion, and return a summary with action items. In many institutions, these tools are already available within existing videoconferencing platforms. Tools that join the call as a visible participant announce themselves, whereas tools that capture system audio leave the disclosure up to you. Summaries tend to underestimate disagreement and can misattribute positions. They should therefore be reviewed before being circulated as a record.

Built-in notetakers (Teams, Zoom, Meet)

The notetaker feature built into widely used videoconferencing platforms. Recordings stay inside the environment an institution has already approved. AI features are tied to a specific licence tier, so availability depends on plan.

• Announces itself to all participants, either as a visible attendee or an on-screen indicator, so disclosure is handled by the tool rather than by the person recording.
• Leaves a searchable transcript attached to the meeting record alongside the summary.
• Summaries and action items are generated automatically once the meeting ends.
• Zoom's assistant can be invited into Teams and Meet calls, where it appears as a guest and produces the same output as on a native Zoom call.

Cautions: Summaries are built from the transcript and omit anything shown on a screen or whiteboard. Transcription assumes a single meeting language, and switching between French and English is handled poorly.

Granola

A notetaker that transcribes audio directly from the computer rather than joining a call. Nothing enters the meeting, so it works the same way across platforms and in a room with no call at all. Summaries are built around notes typed during the conversation rather than from the transcript alone.

• Captures online, hybrid, or fully in-person conversations.
• Typed notes determine what the summary treats as important.
• Templates keep recurring meetings consistently structured.
• Questions can be asked across past meetings.

Cautions: Nothing announces the capture to other participants, so disclosure rests entirely with the user. Audio is processed by services outside institutional agreements, which rules it out for participant data. Summaries depend on what users type during the meeting, so little is gained from meetings attended passively.

3. Literature discovery

These tools search academic literature and help organize results. Retrieval-focused tools search an index of real papers and return results drawn from it. Errors with retrieval tools appear as gaps in coverage or as poor relevance ranking. Generative tools produce summaries of relevant literature. As the summaries are AI generated, the citations within it can contain references that do not exist or references that exist while saying something other than what the text claims. Every citation should therefore be verified against the source.

The indexes underlying these tools skew heavily toward English-language journal publication from high-income countries, shaping both what an AI-assisted search returns and what it omits without explicit indication. This is most consequential for research on marginalized populations, in low- and middle-income settings, and in non-anglophone health systems.

Documents uploaded to these platforms are transmitted to third parties for processing and storage, and on consumer tiers are generally used for model training unless the setting is turned off. Confidential material, including unpublished work, manuscripts sent for peer review, and grant applications, should not be uploaded.

Elicit

Searches Semantic Scholar, a free index of roughly 138 million papers maintained by the Allen Institute for AI, which aggregates publisher metadata, PubMed, and preprint servers. Results come back as a table, with one row per paper and columns the user defines. Sample size, population, design and findings can be pulled into those columns across a large set of papers at once.

• Custom extraction columns produce a comparable table across studies.
• Screens large candidate sets against stated inclusion and exclusion criteria, with a recorded decision for each paper.
• Report mode assembles a narrative summary across a set of sources.
• Uploaded PDFs can be added to a search-derived set, so material outside the index can be included.

Cautions: Extraction depth is a function of access rather than relevance, since full text is available for open access papers while paywalled papers often only access the abstract. Extraction performs well on structured empirical papers and less well where a construct is operationalized differently across studies. Screening decisions remain the user's responsibility and require checking against the papers.

Consensus

Answers directional (effect of X on Y) questions from published literature and lists the supporting papers with answers. Most relevant for health and biomedical questions. The corpus draws on Semantic Scholar, OpenAlex, Consensus's own crawl of the scholarly web, and publisher full-text papers, giving roughly 220 million peer-reviewed papers (updated weekly). The full PubMed corpus is searchable. Medical mode restricts results to papers and clinical guidelines from medical journals.

• Shows how retrieved studies distribute across supporting, mixed, and opposing findings.
• Each result carries a one-line statement of what that paper found.
• Deep search runs multiple queries and returns a longer synthesis.
• Study design and journal metadata appear alongside results, and results can be ranked by design and citation counts.

Cautions: Retrieved studies are counted without weighting by design, sample size or risk of bias, so results represent a tally and not a literature synthesis. Coverage is thinner in the social sciences than in the health sciences and effectively absent for the humanities. Full text is not available for every paper, and where no publisher agreement or open access copy exists the analysis rests on the abstract alone.

Research Rabbit / Litmaps

Citation mapping tools that start from one or more known papers and display what those papers cite and what has since cited them as a visual graph. Research Rabbit draws on Semantic Scholar and PubMed. Litmaps ingests Crossref, Semantic Scholar and OpenAlex, giving roughly 270 million records.

• Surfaces connected work that keyword search misses, including papers using different terminology for the same construct.
• Collections update as new citing work appears, with alerts.
• Both sync with Zotero so an existing library can seed a map.
• Author and journal views show where a cluster of work is concentrated.

Cautions: Litmaps only indexes records whose metadata has been openly deposited, so publications for which metadata was never deposited is absent from the database. Maps are built from citation relationships and reproduce citation practice, so highly cited work dominates while recent and peripheral work is underrepresented. Neither tool reads the papers and relevance still has to be judged by the user.

SciSpace

Covers the full sequence from finding papers to producing manuscript text. It searches its own index of published work alongside Google Scholar and PubMed, explains individual papers passage by passage, holds the results in a reference library, and generates literature review prose from them.

• Chat alongside an open PDF explains passages, notation and methods in context.
• Search results can be arranged into an extraction table.
• Includes a reference library, an AI writer and a paraphrasing tool.
• Specialized agents cover particular domains including biomedicine.

Cautions: The generative writing features can result in fabricated and mismatched references, and text from the writer or paraphraser be revised against the cited source. Paraphrasing raises questions of integrity since reworded text from a source without attribution remains plagiarism. Retrieval spans several indexes in a way that is not documented well enough to support a reproducible search.

4. Data analysis & code

These tools write, explain, and debug code within the environment you already use (e.g. R, Python). They are strongest on translation between languages, debugging errors you can clearly describe, and interpreting unfamiliar output. They are weakest on study design and specification and will tend to over-state confidence. This section also covers the case where a model performs the analytical work itself, such as classifying open-text survey responses.

Some tools suggest code in your editor and send the surrounding code to a provider as context. Some execute code on the provider’s servers, so any data you attach will be uploaded. Others are called from inside a script and send only what you pass into them. A model running on your own hardware does not upload any files. Uploading data to a browser-based tool transfers it to a third party, which most administrative and health data agreements prohibit. Where this applies, code can be developed against synthetic or public data, or a model can be run locally.

Suggested code often runs without error and can execute something slightly different than what was prompted because of implicit assumptions made without checking with the user. Although these models are fluent in the syntax of statistical procedures, they are much less reliable about when those procedures apply. Specification choices, variable construction, and missing data handling warrant particular attention. Suggested code should therefore be read carefully for substantive accuracy before trusting output. Where a model is classifying or coding text, treat it as you would a human coder and validate against a hand-coded subsample.

Finally, suggestions accepted in an editor leave no record of its origin, and analyses carried out inside a chat interface cannot be rerun unless you keep the code and move it into a script. Extra attention should be paid to annotating code to explain decisions and ensuring reproducibility.

IDE assistants (GitHub Copilot, Positron, VS Code)

Code completion and chat functions that operate directly inside the editor. Suggestions appear as the user types, drawing on the surrounding file and the open project for context, and a chat panel answers questions about the code without leaving the environment.

• Completes lines and blocks in place, which suits repetitive transformations and plotting code.
• Chat can explain an error message, a regression output or an unfamiliar function in reference to the actual code in the file.
• Positron's assistant is aware of the session, including loaded data frames and variable names.
• Works across R, Python, SQL and the shell in the same project.

Cautions: Surrounding code is sent to the provider as context, which includes whatever is visible in the file, so a script with a database password or an embedded participant identifier transmits those along with everything else. Suggestions are accepted with a keystroke and leave no record of their origin, which means that the only opportunity for review is at the point of acceptance.

Claude Code

Works on a project environment from the terminal or an editor. It reads files across a repository, edits multiple files, runs code, inspects the output, and revises based on its conclusions.

• Executes a task across several files at once, including refactoring and restructuring an analysis pipeline.
• Runs the code it writes, reads errors and iterates without prompting at each step.
• Reads a project as a whole, so it can locate where a variable was constructed even in a different script.
• Works with version control, producing commits and branches.

Cautions: Running commands and editing files autonomously can result in errors propagating before they are noticed, so users should implement and monitor version control. The scope of what it reads is the project directory, which frequently contains data files alongside code, so confidentiality should be considered.

ellmer (R)

An R package for calling language models from inside a script, used to turn text into variables. Text is classified or has specified fields extracted, with results returned as columns in a data frame ready for analysis.

• Codes open-text survey responses into a category scheme across thousands of rows.
• Returns structured output conforming to a specified schema, so results arrive as usable columns rather than prose.
• Works with several providers, including local models, through one interface.
• The call sits in the script, so the prompt, model and parameters are versioned with the analysis.

Cautions: Whatever is passed in the call is transmitted to the provider, and a data frame of open-text responses may contain identifying detail that participants included. Model outputs vary between runs and model versions, so the same script can produce different results at different times. Users should record the model version and validate results against a hand-coded subsample.

Ollama

Downloads a language model onto a computer and runs it there. This allows it to work without an internet connection or anything being transmitted, which is advantageous when data cannot leave a machine.

• Downloads and runs models locally through a single command.
• Exposes a local API, so it can be called from R or Python in the same way as a hosted provider.
• Model files can be held on institutional hardware or a secure workstation.
• No account, no per-token cost, and no network dependency once the model is downloaded.

Cautions: Performance depends on available memory. As a result, locally runnable models are meaningfully weaker than web-hosted models (particularly for deep reasoning tasks).

5. Transcription & translation

These tools convert speech to text or text from one language into another. Transcription accuracy varies with audio quality, overlapping speech, and language. Machine translation performs well between well-resourced language pairs, including French and English, but less well on specialized terminology. Material intended for publication or for participants should be reviewed by a fluent speaker.

Whisper (local and API)

An open speech recognition model that transcribes audio into text. It can be downloaded and run on a computer, in which case the recording never leaves it, or it can be called as a paid service, in which case the audio is uploaded into a server. Local use requires an interface such as MacWhisper or WhisperX.

• Handles long recordings and poor audio better than most alternatives, including overlapping speech and background noise.
• Transcribes around a hundred languages and can translate speech directly into English.
• WhisperX adds speaker labels and word-level timestamps, which matters for interview data.
• Free when run locally, with no per-minute charge and no account.

Cautions: The local and paid/server versions have different data handling implications (only the local form satisfies an agreement that prohibits transmitting recordings). Local transcription of a long interview is time-intensive and benefits from a recent machine. The model occasionally produces fluent text for passages where there is no clear speech, so users should check the result against the audio input.

Otter / Sonix

Cloud transcription services with a browser interface. Audio is uploaded and returns as an editable transcript synchronized to the recording, with speaker labels applied automatically.

• Editing can occur alongside playback, so corrections can be made in reference to the audio.
• Speaker identification and searchable transcripts across a whole project.
• Sonix supports transcription in a wider set of languages and translation of the resulting transcript.
• Otter joins calendared meetings directly and produces live transcription.

Cautions: Audio is uploaded to a commercial service outside institutional agreements. Automatic speaker labels drift when voices are similar or speakers interrupt each other, so labels need verification before the transcript is used analytically.

DeepL

Machine translation of text and documents, with a reputation for handling French and English more naturally than general-purpose alternatives. Whole documents can be translated with formatting preserved (paid tier).

• Translates Word, PowerPoint and PDF files while keeping layout intact.
• A glossary fixes how specified terms are always rendered, which holds technical vocabulary consistent across a document.
• Alternative phrasings can be selected within a translation.
• A write function edits text within a language rather than across languages.

Cautions: Text is transmitted to the service, so users translating anything confidential should verify the terms of service. Fluency can mask errors in specialized terminology for non-native speakers.

Prepared by Clara Bolster-Foucault (McGill University) for the BRIDGE Proof-of-Concept · September 2026