Mapping the Landscape of AI and Research

Dominik Lukes · AI Competency Centre, University of Oxford

Open slides
Download slides

handouts.fyi/2qq4Share this page with someone else attending.

Mapping the Landscape of AI and Research

Mapping the Landscape of AI and Research

Dominik Lukes · AI Competency Centre, University of Oxford

Introduction

About me

dominiklukes.net

AI Consultant work

ainews.dominiklukes.net

Work on academic productivity

Research in: Translation, Linguistics, Metaphor

Michaelmas 2026 Workshops on Agents

AI and research: A quick history

AI and research is changing

2023
  • AI writes or edits papers
  • AI answers questions
  • AI discovers literature
  • AI hallucinates and makes errors
2026
  • AI makes independent discoveries
  • AI autonomously runs full research workflows
  • AI opens up more possibilities
  • AI finds errors in published literature

Research roles through time

Computers and AI in research Selected publications, 1931–2026 · overlapping research roles. Grouped periods; column widths do not represent elapsed time. Blank cells mean no example selected here. A system can contribute to several roles. Calculate & model: 1931 · Differential analyser; 1949 · Monte Carlo; 1950 · ENIAC weather; 1971 · MACSYMA; 2021 · AlphaFold; 2026 · Intern-S1-Pro* Search & check proofs: 1956–57 · Logic Theory Machine; 1965 · Resolution; 1976 · Four-colour theorem; 1997 · EQP / Robbins; 2017 · Flyspeck Apply specialist knowledge: 1975 · MYCIN; 1978 · DENDRAL / Meta-DENDRAL; 1982 · PROSPECTOR; 2026 · GPT-Rosalind*; 2026 · Claude biology report* Form hypotheses & guide experiments: 1981 · BACON.3; 2004 / 2009 · Robot Scientist; 2020 · Mobile robotic chemist; 2023 · Coscientist; 2026 · Co-Scientist Read, write & coordinate research: 2023 · Scim (paper skimming); 2023 · GPT-4 paper feedback; 2026 · The AI Scientist; 2026 · Robin; 2026 · Paper Assistant Tool Evaluate systems & research outputs: 2025 · Sakana / KOSMOS; 2026 · LABBench2 / LifeSciBench; 2026 · Into the Unknown Computers and AI in research Selected publications, 1931–2026 · overlapping research roles RESEARCH ROLE 1931–1965 1971–1997 2004–2023 2025–2026 Calculate & model 1931 · Differential analyser 1949 · Monte Carlo 1950 · ENIAC weather 1971 · MACSYMA 2021 · AlphaFold 2026 · Intern-S1-Pro* Search & check proofs 1956–57 · Logic Theory Machine 1965 · Resolution 1976 · Four-colour theorem 1997 · EQP / Robbins 2017 · Flyspeck Apply specialist knowledge 1975 · MYCIN 1978 · DENDRAL / Meta-DENDRAL 1982 · PROSPECTOR 2026 · GPT-Rosalind* 2026 · Claude biology report* Form hypotheses & guide experiments 1981 · BACON.3 2004 / 2009 · Robot Scientist 2020 · Mobile robotic chemist 2023 · Coscientist 2026 · Co-Scientist Read, write & coordinate research 2023 · Scim (paper skimming) 2023 · GPT-4 paper feedback 2026 · The AI Scientist 2026 · Robin 2026 · Paper Assistant Tool Evaluate systems & research outputs 2025 · Sakana / KOSMOS 2026 · LABBench2 / LifeSciBench 2026 · Into the Unknown Grouped periods; column widths do not represent elapsed time. Blank cells mean no example selected here. A system can contribute to several roles. * Lab report or announcement. Source links and access limits are retained in the speaker notes.

2021–2026: prediction and connected research tools

2021 · AlphaFold

  1. Prediction
  2. Predict protein structures

2023 · Coscientist

  1. Tool use and experimentation
  2. Connect an LLM to search, code and lab automation

2026 · AI scientist systems

  1. Research coordination
  2. Combine several research activities in one workflow

2026: AI scientists combine different research roles

25 March · The AI Scientist

  1. Code, experiments and writing
  2. Generate ideas, run computational work and write a manuscript

19 May · Co-Scientist

  1. Hypothesis generation
  2. Generate and critique hypotheses for experimental checking

19 May · Robin

  1. Analysis and laboratory feedback
  2. Link literature and analysis agents with experiments

Model knowledge benchmarks are saturated

Recent news

2026: labs report AI for scientific discovery

March–April · Chinese research labs

  1. Scientific models and infrastructure
  2. Intern-S1-Pro report — CAS ScienceOne announcement

16 April · OpenAI GPT-Rosalind

  1. Life-sciences research
  2. A lab announcement and developer evaluations

23 September · Anthropic Claude

  1. Biological discovery
  2. A lab report about an enzyme system

AI in Research Landscape

  • Autonomous discovery
  • Assisted discovery
  • 3Researcher support
  • Project management
  • Expanded methods
  • More types of data
  • More mini experiments
  • Increased ambition
  • More work

Tools for AI in research

Tools for research

Deep Research history

Deep Research tools as of 2025

New paper review services

Everything is different in the age of agents

Dean W. Ball, Among the Agents, 2026

By the end of this year, the least important thing you will be able to do with frontier AI systems will be getting chatbots to answer questions, but this is still how most people will think of “AI.” Expect cognitive dissonance as a result.

Dean W. Ball, Among the Agents, 2026

Timeline of AI

AI as Oracle
  • Capabilities: answer questions, summarise, translate
  • Chatbots: ChatGPT, Gemini, Claude
  • Specialist apps: NotebookLM, Elicit, Consensus
AI as Tool Maker
  • Capabilities: write code, manage code base
  • Outcomes: create scripts, dashboards, simulations and workflows
  • Tools: Cursor, Lovable, Google AI Studio
AI as Tool User
  • Capabilities: plan, work with files, run utilities
  • Outcomes: more ambitious projects, manage data and replicate analyses
  • Desktop agents: Codex, Claude Code, Antigravity

What are agents

What are agents

What makes an agent useful?

A chat can tell me how to fill in an expenses form.
An agent can find the form, search for receipts in emails, calculate totals, process receipts and fill in the form.

The agentic loop

HarnessSends instructions / results to model
ModelMakes a plan
ModelSend commands to harness
HarnessPasses command to computer
ComputerExecutes command and returns to harness

Tools Agents can use

  • Search files
  • Search web
  • Command commands
  • CLI tools
  • APIs
  • Other models
  • Custom scripts
  • Browser
  • Applications 🆕

Move from Oracle to Tool User and Tool Maker

Before 2026
In 2026
Use Deep Research on a website
Use Elicit API to conduct searches
Summarise a paper
Create a visualisation or replication
Format paper for reference manager
Tag and file papers in Zotero
Write an lit review
Make a visualisation of literature

Use external tools

Codex used SketchEngine API to run simple corpus analysis.

Run mini experiments

Codex made a plan to run mini experiment on compression, used local tools to convert images in various sizes and present results for visual inspection as HTML with a table.

Run long-running research loops

Simulate full scientific discovery

What agents enable

  • More control and reproducibility
  • More experiments
  • More data processing

Desk top agents work on local files

AspectAdvantageTrade-off
File handlingFiles stay in your normal folder structureYou need to manage that structure properly
RetrievalOutputs are easier to find laterYou need to save and name things well
ScaleEasier to work across many local filesYou need to know what folder you are giving it access to
PrivacySome tasks can stay mostly localSome tasks still require remote model processing
PersistenceThe work is tied to your project folderIt does not automatically follow you to every machine

Their work can be fully inspected

Full transcripts of all chats are stored as text files on disk. You can ask your agent to inspect them and give you a summary and analyse tool calls and token use.

Context transparency

You can even build an app to analyse them

Literature review example

Elicit (on the web)

Saved view 1 of 2, captured 27 Oct 2026 of elicit.com
Saved view 1 of 2, captured 27 Oct 2026
Saved view 2 of 2, captured 27 Oct 2026 of elicit.com
Saved view 2 of 2, captured 27 Oct 2026
Open elicit.com ↗

Search Elicit via API

Run an Elicit literature survey on how university students use coding agents, then file the papers in a new Zotero collection.

I’ll load the Elicit skill first so I can follow its research workflow and choose the appropriate tools.

load_skill elicit

Tool results

Loaded the Elicit research skill.
Available tools: search papers, screen results, extract findings.

I’ll search for empirical work on coding agents in university study and research, then screen the results for relevance.

elicit.search_papers {"query":"university students using AI coding agents for study and research"}

Tool results

Found 18 candidate papers.

I’ll prioritise studies involving university students and extract the participants, tasks, findings and limitations for a comparable survey.

elicit.screen_and_extract {"criteria":"Empirical studies involving university students","fields":["Participants","Agent","Task","Findings","Limitations"]}

Tool results

Selected 8 relevant papers.
Prepared a survey summary and bibliographic metadata.

The selected papers are ready. I’ll load the Zotero skill to determine how to access the desktop library and organise them.

load_skill zotero

Tool results

Loaded the Zotero skill.
Read the guidance for accessing the desktop library through the local API.

I’ll check that Zotero is running and that the local API supports the operations needed to create and populate a collection.

zotero.local_api.inspect {}

Tool results

Zotero is running.
The library is accessible and collection creation is available.

Before importing anything, I’ll match the selected papers against existing library items to avoid duplicates.

zotero.find_existing_items {"papers":"the 8 selected papers"}

Tool results

Found 3 existing items.
Prepared 5 new items for import.

I’ll create a collection named for the survey, then file both the existing and newly imported papers there.

zotero.create_collection {"name":"Coding agents in university study — Elicit survey"}

Tool results

Created collection A1B2C3D4.

I’ll import the five missing papers, add all eight to the collection, and verify its contents before reporting completion.

zotero.import_and_file {"collection":"A1B2C3D4","existing_items":3,"new_items":5}

Tool results

Imported 5 papers.
Verified that all 8 papers belong to the collection.

Your Zotero collection is created: “Coding agents in university study — Elicit survey”.

It contains 8 papers: 3 already in your library and 5 newly imported.

Open the collection in Zotero: zotero://select/library/collections/A1B2C3D4

Import and tag papers into Zotero collections

Create a visualisation of references for easier review

From paper summaries to visualisations and replications

Open agent-demos.pages.dev ↗

More small experiments

Small experiments: Breakfast TM example

Open languagelog.ldc.upenn.edu ↗

What a small experiment looked like a year ago

Saved view 1 of 4, captured 27 Oct 2026 of viewfromhistory.semanticmachines.fyi
Saved view 1 of 4, captured 27 Oct 2026
Saved view 2 of 4, captured 27 Oct 2026 of viewfromhistory.semanticmachines.fyi
Saved view 2 of 4, captured 27 Oct 2026
Saved view 3 of 4, captured 27 Oct 2026 of viewfromhistory.semanticmachines.fyi
Saved view 3 of 4, captured 27 Oct 2026
Saved view 4 of 4, captured 27 Oct 2026 of viewfromhistory.semanticmachines.fyi
Saved view 4 of 4, captured 27 Oct 2026
Open viewfromhistory.semanticmachines.fyi ↗

A small research project today: Audio Archive Project

Created from 340 hours of audio files over the weekend

Saved view 1 of 2, captured 27 Oct 2026 of radiotimesof.dominiklukes.net
Saved view 1 of 2, captured 27 Oct 2026
Saved view 2 of 2, captured 27 Oct 2026 of radiotimesof.dominiklukes.net
Saved view 2 of 2, captured 27 Oct 2026
Open radiotimesof.dominiklukes.net ↗

Research process

Saved view 1 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 1 of 7, captured 27 Oct 2026
Saved view 2 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 2 of 7, captured 27 Oct 2026
Saved view 3 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 3 of 7, captured 27 Oct 2026
Saved view 4 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 4 of 7, captured 27 Oct 2026
Saved view 5 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 5 of 7, captured 27 Oct 2026
Saved view 6 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 6 of 7, captured 27 Oct 2026
Saved view 7 of 7, captured 27 Oct 2026 of audioarchive.semanticmachines.fyi
Saved view 7 of 7, captured 27 Oct 2026
Open audioarchive.semanticmachines.fyi ↗

Lessons from audio archive project

  • What would have taken a project team weeks can be done in by one person in a few days
  • Agents can act as the research coordinator while researcher acts as PI
  • Smart models can make good choices about research design process along the way
  • It is possible to document the process and inspect all decisions

Impact on research process

Key consequences

  • Impact on the research environment
  • Project team roles
  • Project design
  • Project funding
  • Researcher education

Journals struggle with paper overproduction

Open blog.tmlr.org ↗

Some fields study with glut of results and identity crisis

Open proofsandprompts.com ↗

Existing results are finding errors

Impact on research projects

  • Timelines
  • Aims
  • Roles
  • Costs

Impact on researcher education

  • Speed of change
  • Gaps in supervision
  • Uncertain method impact
  • New challenges to results review
  • Lack of expert trainers

Thank you

Thank you

Dominik Lukes · AI Competency Centre, University of Oxford