Field Kit Science
OverviewBundlesLicensesSearchHow it works
AI in research · what moves, what breaks, a way of working

Every claim traces backto its page.

A research corpus that grows every night behind a license gate, read by language models that must show their sources. Not a product to install: a way of working, built from open parts.

Claim 18✓Supported

Beyond accuracy, several authors warn of an epistemic cost.

↳ Quote · p. 49 · curator's highlight

“The proliferation of AI tools in science risks introducing a phase of scientific enquiry in which we produce more but understand less”

↳ Paper
Messeri 2024Reddoi:10.1038/s41586-024-07146-0
The problem

Language models are already in the literature.Almost nobody says so.

13.5%

of 2024 PubMed abstracts were, at a lower bound, processed with an LLM

Kobak 2025 · trace →

40%

the same lower bound in some disciplines, countries and journals

Kobak 2025 · trace →

~0.1%

of 75,000 papers published since 2023 disclosed any AI use

He 2026 · trace →

The trouble is not that researchers use these tools. It is that a fluent summary carries no address: which sentence in which paper it rests on, whether you were allowed to read that paper, or whether it has since been retracted.

What is moving · 1 of 3

Models of the language of life

Foundation models trained on sequences, structures, images and sound from across the tree of life. In evolution and biodiversity research, generative AI is already part of the method.

  • Green

    The reference point everyone knows: structure from sequence.

    Jumper 2021 · Nature: Highly accurate protein structure prediction with AlphaFold

  • Yellow

    Evolution read as a language: a functional GFP only 58% similar to known variants.

    Hayes 2025 · Science: Simulating 500 million years of evolution with a language model

  • Green

    Modelling and design from molecular to genome scale, in one DNA language model.

    Nguyen 2024 · Science: Sequence modeling and design from molecular to genome scale with Evo

  • Yellow

    All domains of life, including the effects of variants.

    Brixi 2026 · Nature: Genome modelling and design across all domains of life with Evo 2

  • Green

    From prediction to design: new functional genes written by a genomic language model.

    Merchant 2025 · Nature: Semantic design of functional de novo genes from a genomic language model

  • Green

    Taxonomy meets vision models: a direct line to Linnaeus.

    Stevens 2024 · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings: BioCLIP: A Vision Foundation Model for the Tree of Life

  • Green

    Acoustic biodiversity monitoring at scale.

    Kahl 2021 · Ecological Informatics: BirdNET: A deep learning solution for avian diversity monitoring

  • Green

    The Uppsala anchor: constraint and innovation across hundreds of placental mammals.

    Christmas 2023 · Science: Evolutionary constraint and innovation across hundreds of placental mammals

Notes are the curator's reading. Classes come live from the license gate: on green papers a note can be checked against the full text indexed here, on the others it rests on the abstract.

What is moving · 2 of 3

Agents that do research

Not a chatbot that answers, but systems that plan, call tools, run experiments and write up. Almost all of this is from the last three years.

  • Green

    An agent that plans and runs real chemistry experiments.

    Boiko 2023 · Nature: Autonomous chemical research with large language models

  • Red

    A team of agents with an AI as principal investigator, designing nanobodies.

    Swanson 2025 · Nature: The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies

  • Greensources disagree

    Hypothesis generation, validated in vitro.

    Gottweis 2026 · Nature: Accelerating scientific discovery with Co-Scientist

  • Green

    AI mirrors experimental science: a mechanism of gene transfer crucial to bacterial evolution.

    Penadés 2025 · Cell: AI mirrors experimental science to uncover a mechanism of gene transfer crucial to bacterial evolution

  • Green

    The whole research cycle automated, including an accepted workshop paper.

    Lu 2026 · Nature: Towards end-to-end automation of AI research

  • Green

    A general-purpose biomedical agent.

    Huang 2025 · bioRxiv (Cold Spring Harbor Laboratory): Biomni: A General-Purpose Biomedical AI Agent

  • Green

    New this autumn: an agent environment for collaborative academic research groups.

    Wu 2026 · bioRxiv (Cold Spring Harbor Laboratory): An agentic AI environment to support biomedical research in collaborative academic environments

Notes are the curator's reading. Classes come live from the license gate: on green papers a note can be checked against the full text indexed here, on the others it rests on the abstract.

What is moving · 3 of 3

Where they fall short

The same literature measures the gap. Read these before you trust the ones above.

  • Green

    Where models still trail experts on practical biology research tasks.

    Laurent 2024 · arXiv (Cornell University): LAB-Bench: Measuring Capabilities of Language Models for Biology Research

  • Green

    Agents scored on end-to-end biomedical machine learning.

    Miller 2025 · bioRxiv (Cold Spring Harbor Laboratory): BioML-bench: Evaluation of AI Agents for End-to-End Biomedical ML

  • Green

    What models know, and do not, about plant molecular biology.

    Burda 2025 · bioRxiv (Cold Spring Harbor Laboratory): What Large Language Models Know About Plant Molecular Biology

  • Green

    The limits of multimodal models in chemistry and materials research.

    Alampara 2025 · Nature Computational Science: Probing the limitations of multimodal language models for chemistry and materials research

  • Green

    How much faster, really: the limits of acceleration in biomedical research.

    Hebenstreit 2026 · Scientific Reports: What are the limits to biomedical research acceleration through general-purpose AI?

  • Green

    Verification as the bottleneck of AI-driven discovery.

    Cornelio 2026 · Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences: The need for verification in artificial intelligence-driven scientific discovery

Notes are the curator's reading. Classes come live from the license gate: on green papers a note can be checked against the full text indexed here, on the others it rests on the abstract.

What breaks

What it does to the evidence base

Fabricated references, summaries that overreach, text nobody disclosed, and papers retracted for what a model wrote. The figures in the problem section come from the same bundle.

  • Green

    55% of references fabricated with GPT-3.5, 18% with GPT-4.

    Walters 2023 · Scientific Reports: Fabrication and errors in the bibliographic citations generated by ChatGPT

  • Red

    Illusions of understanding: AI tools can make a field feel it knows more than it does.

    Messeri 2024 · Nature: Artificial intelligence and illusions of understanding in scientific research

  • Green

    Summaries that generalise beyond what the study found.

    Peters 2025 · Royal Society Open Science: Generalization bias in large language model summarization of scientific research

  • Red

    A measurable rise in LLM-modified text across fields.

    Liang 2025 · Nature Human Behaviour: Quantifying large language model usage in scientific papers

  • Orange

    LLM-generated text in peer review.

    Liang 2024 · arXiv (Cornell University): Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews

  • Yellow

    Journal AI policies have not curbed the rise in AI-assisted writing.

    He 2026 · Proceedings of the National Academy of Sciences: Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing

  • Green

    Delegating a task to AI can increase dishonest behaviour.

    Köbis 2025 · Nature: Delegation to artificial intelligence can increase dishonest behaviour

  • Red⚠ Retracted

    Retracted: the introduction opened with “Certainly, here is”.

    Zhang 2024 · Surfaces and Interfaces: RETRACTED: The three-dimensional porous mesh structure of Cu-based metal-organic-framework - Aramid cellulose separator enhances the electrochemical performance of lithium metal anode batteries

  • Green⚠ Retracted

    Retracted for AI-generated figures. Its license lets it be stored; the retraction keeps it out of synthesis.

    Guo 2024 · Frontiers in Cell and Developmental Biology: RETRACTED: Cellular functions of spermatogonial stem cells in relation to JAK/STAT signaling pathway

Notes are the curator's reading. Classes come live from the license gate: on green papers a note can be checked against the full text indexed here, on the others it rests on the abstract.

The idea

Don't trust the summary.Trace it.

Treat verification as a way of working, not a tool. The system reads at night; you judge in the morning, and every judgement has somewhere to stand.

01

Gate

Decide what you may do with a text before you store it. Unknown means closed.

02

Trace

Every claim points to a quote, a page and a DOI. A claim without a source cannot be saved.

03

Check

A second model judges each claim. A plain string match confirms every quote is really there.

04

Remember

Know when the system is saying the same thing again in new words.

How a claim is made

The model proposes. The rest checks.And your method goes in at every step.

Step
What happens
Where your method goes in
01Choose the reading
the system, from your bundle
Up to 80 numbered excerpts: anchors first, then what is new, then the rest in turn. Open full text, abstracts where allowed, your Zotero highlights. Never retracted or closed text.
Your anchors and highlights decide what is read. Change the centre, the size, what counts as evidence.
02Write
Kimi K3
Up to ten claims, each citing one to four excerpts with a word-for-word quote. It is told what earlier nights concluded and asked what is new.
The instructions are plain text: ask for contradictions, methods, your theoretical lens. Sensitive material can stay on a local model.
03Check the quotes
no model
Every quote must appear word for word in the excerpt it cites. A source that fails is dropped; a claim with none left is never saved.
The one step not to soften.
04Verify
GLM-5.3, another model family
Reads each claim against the full text it cites: supported, partly supported or not, with a reason.
Choose the checker, sharpen what the verdicts mean, or make a colleague the checker.
05Remember
no model
Ties each claim to the first one that said the same, and marks it new, a repeat, or extended by a new paper.
Decide what counts as saying the same thing.
06Judge
you
Read the trace before you repeat a claim. Take a paper that does not belong out of the bundle, and say why.
This is the method. The rest makes it possible.

The model never searches. It reads what the system and you have chosen, and every sentence it writes has to point back into that. So far: 524 claims checked, 429 supported, 94 partly, 1 not.

One claim, traced

One claim, all the way down.

Claim 18 · research-integrity · 30 Sept 2026

Beyond accuracy, several authors warn of an epistemic cost. Messeri & Crockett 2024 caution that AI may lead science to produce more but understand less. Bietti & Bangerter 2026 argue that outsourcing writing risks decoupling it from thinking and eroding core reasoning skills.

✓SupportedWritten by Claude Opus 5.5
  1. E8 · Annotation · p. 49Red

    A closed paper. The system holds none of its text: this is the curator's own highlight from Zotero, with page and comment.

    “The proliferation of AI tools in science risks introducing a phase of scientific enquiry in which we produce more but understand less”

    Curator: This is one of the comments I hear most often when speaking to colleagues about use of AI tools.

    Messeri 2024: Artificial intelligence and illusions of understanding in scientific research

  2. E15 · Full textGreen

    An open paper, read from its full text, indexed on our own server. This text carries no page numbers.

    “it threatens the development and exercise of core cognitive skills involved in scientific reasoning”

    Bietti 2026: Will the widespread use of large language models in scientific writing undermine scientists’ critical thinking?

Memory: the synthesis has made this claim 19 times since 30 Sept 2026, each time in new words. It is counted as one line, so only a new paper extends it.

Where it chafes

Four things that went wrong,and what they taught.

The license was green. The PDF was not.

A paper was correctly green for its preprint. The PDF the curator had in Zotero was the publisher's version, through the library. It was stored and indexed, and six claims cited it.

Changed: The Zotero fallback now requires the gate itself to find the published version open, with no conflict. The text and its index were removed; the claims kept their other sources.

A license belongs to a version, not to a DOI. The gate has to know which version it is holding.

Every check passed. The paper didn't belong.

In one bundle, 33 of 40 discovered papers came in because they cite a single, very widely cited anchor. Marketing, supply chains and the game of Go scored as close to the topic as the papers that belonged.

Changed: Each bundle can define its own centre, no anchor may bring in more than a share of a night's picks, and removals are recorded with a reason so they are not rediscovered.

Machine checks verify the chain from claim to page. Whether a paper belongs is still a human judgement.

The synthesis kept saying the same thing.

Night after night the same claims came back in new words. One line, about the share of LLM-processed abstracts, was made eleven times. The model was fine; it was being shown nearly the same papers every night.

Changed: Claim memory links rewordings to one line. The evidence rotates, so papers that have waited longest are read next, and the model is told what has already been said.

A fluent new paragraph is not new knowledge. Ask what is new since last night.

Every job was green. Nothing was coming in.

For a week the nightly jobs all reported success while almost no new paper arrived. Five of six bundles had reached their target, and discovery only looked at the most-cited works citing each anchor, so last week's papers could never qualify.

Changed: A second lane takes work from the last 30 days, newest first, about ten papers a week across the bundles. It does not count against a bundle's target.

A green tick says a job ran, not that it did anything. Watch what arrives, not only what runs.

What it asks of supervision

Training researcherswhen their tools are agents.

Master's and PhD students already use these tools, often without saying so. The research on what it does to learning and to the supervisory relationship is young, and growing every week.

  • Orange

    Better inside the frontier, 19 percentage points worse outside it.

    Dell’Acqua 2023 · SSRN Electronic Journal: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality

  • Green

    Doctoral students use AI quietly, as a substitute for supervision.

    Boyd 2025 · Buildings and Cities: Generative AI: reconfiguring supervision and doctoral research

  • Yellow

    The tension between productivity and academic norms.

    Lai 2025 · Australasian Journal of Educational Technology: The impacts and tensions of generative AI on doctoral students’ supervisory and peer dynamics: An activity theory analysis

  • Yellow

    Generative AI without guardrails can harm learning.

    Bastani 2025 · Proceedings of the National Academy of Sciences: Generative AI without guardrails can harm learning: Evidence from high school mathematics

  • Green

    Cognitive offloading, and the calibration it takes to stay accountable for the result.

    Gao 2026 · Frontiers in Psychology: Cognitive offloading and metacognitive calibration in generative AI-mediated doctoral learning: a grounded theory study of accountable reconstruction

  • Green

    When the thesis passes but the mind doesn't.

    Singh 2026 · Transformation in Higher Education: When the thesis passes but the mind doesn’t: Generative artificial intelligence, proxy regimes and epistemic agency in universities

  • Red

    Published 21 October 2026: doctoral researchers' AI use and their integrity concerns.

    Zhuang 2026 · Educational Studies: Doctoral researchers’ Generative AI use and research-integrity concerns: a cross-context study

Notes are the curator's reading. Classes come live from the license gate: on green papers a note can be checked against the full text indexed here, on the others it rests on the abstract.

What it asks of supervision

What to ask of a student's AI-assisted work.

01

Show the chain

Every claim you bring me points to a quote, a page and a DOI. If it cannot, it is a lead, not a finding.

02

Say who checked

Which parts did a model write, which did a model check, and which did you read yourself?

03

Formulate the task first

Write down what you are asking before the agent does it. The jagged frontier is easiest to see from there.

Questions to take with you
  • Where would you not trust a connection like this in your own research?
  • Where is the line between a connection you would not have seen and one that should not be there?
  • How much should your own judgements steer what the system shows you next?
  • What would you have to do every week for this to stay honest?
Sources

The parts are open.The method is yours.

Every building block here is free to use. What makes it work is the way of working around it: your anchors, your reading, and a person who judges.

Papers cited in this talk
  • RedMesseri 2024: Artificial intelligence and illusions of understanding in scientific research
  • GreenBietti 2026: Will the widespread use of large language models in scientific writing undermine scientists’ critical thinking?
  • GreenKobak 2025: Delving into LLM-assisted writing in biomedical publications through excess vocabulary
  • YellowFortenbach 2026: Large Language Model Authorship in Ophthalmic Publications
  • YellowHe 2026: Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing
  • GreenLee 2025: The role of large language models in the peer-review process: opportunities and challenges for medical journal reviewers and editors

And the 37 papers in the overview, each with its license class: science.andhumans.se/story#sources

Built on
  • OpenAlex ↗
  • Crossref + Retraction Watch ↗
  • Unpaywall ↗
  • Europe PMC ↗
  • Zotero ↗
  • pgvector ↗
  • bge-m3 ↗
  • Model Context Protocol ↗
Appendix

The system in detail.For the discussion.

Why now

Why it is possible now.

Three things that were not true a few years ago.

Open metadata

OpenAlex (CC0), Crossref, Unpaywall and Europe PMC say, per paper and per version, where a legal copy is and under which license.

Retractions in the open

Since 2023 the Retraction Watch database is openly available through Crossref, so checking every paper every night costs nothing.

Models that can cite

Current models can write against numbered evidence and be checked by a second model. Agent protocols let you do it from a conversation.

The whole flow · live

From a paper to a traced claim.

Each box reads this instance as the page loads. ↳ marks where a person decides.

Ways in
Zotero group
every 5 min
5 Nov, 15:07 ✓
↳ you: highlights, comments
Claude and the CLI
on request, via MCP
Citation search
02:15
5 Nov, 02:15 ✓
New work, last 30 days
03:00
5 Nov, 03:00 ✓
License recheck
03:30 · every 30 days
5 Nov, 03:30 ✓
→
One door
Intake queue
786 processed · 19 failed

29 new papers in the last 7 days. Every way in writes only here; one worker does the rest.

→
The worker
Metadata
OpenAlex · Crossref · DataCite · Europe PMC
License gate
214 green · 86 yellow · 21 orange · 68 red
↳ you: 16 conflicts to settle
Retraction check
2 retracted · kept, never used
Full text and index
green only · bge-m3 · hourly
5 Nov, 14:52 ✓
→
The corpus
Bundles
389 papers · 8 bundles
↳ you: choose the anchors
Curator's reading
4 highlights from Zotero
↳ you: read and annotate
Removals
out of a bundle, with a reason; never rediscovered
↳ you: prune what does not belong
→
Synthesis · 04:30
Write
Kimi K3
5 Nov, 04:33 ✓
Check
GLM-5.3 + quote string match
524 claims · 429 supported
Remember
new · repeat · extended
→
Read
Web view
read-only, for anyone
Claude via MCP
trace, search, list claims
The trace
claim → quote → page → DOI
↳ you: open the sentence before you repeat it
Private notes→ a separate system; only counts cross over →Nothing private enters this corpus, so it never reaches a model or the web view.
One night · live

What runs while you sleep.

Stockholm time. Each mark shows when the job last ran and whether the last seven days went well.

00:00
01:00
02:00
03:00
04:00
05:00
06:00
02:15 Citation search
5 Nov, 02:15 ✓
03:00 New work, last 30 days
5 Nov, 03:00 ✓
03:30 License + retraction recheck
5 Nov, 03:30 ✓
04:30 Synthesis and check
5 Nov, 04:33 ✓
05:10 Counts from the private system
last 5 Nov 2026
Every 5 min · Zotero sync
5 Nov, 15:07 ✓
Hourly · search index, intake retry
5 Nov, 14:52 ✓
Papers added per day, last 30 days
Papers added per day · peak 12
0128 Oct5 Nov
Table view
8 Oct1
9 Oct0
10 Oct0
11 Oct0
12 Oct0
13 Oct1
14 Oct0
15 Oct4
16 Oct3
17 Oct3
18 Oct0
19 Oct0
20 Oct0
21 Oct0
22 Oct0
23 Oct4
24 Oct3
25 Oct2
26 Oct0
27 Oct0
28 Oct0
29 Oct0
30 Oct4
31 Oct2
1 Nov0
2 Nov2
3 Nov12
4 Nov1
5 Nov8
Papers
389
8 active bundles
Claims checked
524
429 supported
License conflicts
23
16 waiting for a human
Retracted
2
kept, never synthesised
New work · research-integrity · ranking of 14 Oct 2026

A connection you would not have seen,or one that should not be there?

Recent papers that cite the bundle's anchors, ranked by geometry, not by a model: closeness to the bundle's centre, plus a bonus for each further anchor cited. 44 cleared the floor. The lane takes 3 a week.

RankScoreAnchorsPublishedTitle
10.831423 Sept 2026From disclosure to verification: a risk-differentiated framework for the ethical use of AI in academic researchpicked tonight
20.77924 Oct 2026AI detectors, academic freedom, and scientific communication: A transition from prohibition to regulated disclosure
30.759224 Sept 2026Navigating AI-Accelerated Science: An Intellectual Virtue Approach to Responsible Inquiry
40.756119 Sept 2026Responsible Generative AI in Doctoral Education: Balancing Innovation, Academic Integrity, and Researcher Autonomy
50.748115 Sept 2026Pressures, politics, possibilities: how AI might transform the research ecosystem
60.74516 Oct 2026AI-assisted peer review: efficiency for reviewers, burden for authors—a narrative review
70.740123 Sept 2026Artificial intelligence in tissue engineering and regenerative medicine: from algorithmic capability to intelligent integrationdoes it belong?
… 130.728123 Sept 2026Generative Artificial Intelligence in Oncology Publishing: Vocabulary Adoption Without Gains in Publication Volume, 2018–2025plainly belongs

Rank 7 is a review of AI in tissue engineering. “AI in science” sits close to this bundle's centre, so it scores 0.740 on one anchor.

Rank 13 measures LLM vocabulary in oncology papers, the bundle's own subject. It ranks lower.

At three a week, rank 7 arrives within weeks and passes every machine check. Then it is the curator's call: keep it, or remove it with a reason.

Boundaries

Rules that holdeven when the code is wrong.

A bug in the intake worker should not be able to leak a closed text or publish an unsourced claim, so the important rules live in the database too.

  • The gate lives twice

    As a pure function with an offline test suite, and as database triggers: full text only for green and yellow, the search index only for green.

  • Unknown means closed

    If no source can say what a paper's license is, it is treated as red until someone finds out.

  • Retracted is kept, never used

    Retracted papers stay visible and marked. The database refuses them as a source for any claim.

  • No claim without a source

    The database will not commit a claim that has no source attached.

  • Writes go through one door

    Every way in only adds to the intake queue. This site and the public agent profile can only read.

  • Nothing private crosses over

    This instance has its own database. From the curator's private system it receives counts, never content.

The way of working

You don't need this code.You need the way of working.

What it asks of you
  • Choose the anchorsA bundle starts from a handful of papers you know well. They decide what the system goes looking for.
  • Settle the conflictsWhen sources disagree about a license, the gate flags it. You open the PDF and read the license line.
  • Read traces, not summariesOpen the sentence before you repeat the claim. The trace is there so that you use it.
  • Prune, with a reasonWhen a paper does not belong, take it out of the bundle and write down why. The record is part of the method.
  • Expect repetitionMost nights add little. The useful question every morning is what changed, not what was said.
Where to start

Build your own, around your own reading.

01

One bundle

Ten papers you know well, on one question you care about.

02

The gate first

Before you store any text, decide what you are allowed to do with it.

03

The trace before the synthesis

Make it impossible to save a claim without a page and a DOI, then let a model write.

04

A month of nights

Run it every night and read the mornings before you trust it with anything.

What you build
  • Open metadata

    OpenAlex, Crossref with Retraction Watch, Unpaywall, Europe PMC, arXiv and bioRxiv. All have free APIs.

  • The reference manager you already use

    Zotero with a shared group: one collection per bundle, and your highlights become evidence.

  • A database that can search by meaning

    Postgres with pgvector, and an embedding model you host yourself (here bge-m3).

  • Two language models

    One writes, a different one checks. Here they are open-weight models at night, Claude in conversation.

  • A scheduler and an agent protocol

    Jobs that run while you sleep, and MCP so you can add a paper or trace a claim from a chat.

  • A small server

    This runs in about 1–1.5 GB of memory, next to other things on a modest box.

Introduction01 / 22✕

Read-only. Public bibliographic data only; no visitor data is collected. Metadata from OpenAlex (CC0), Crossref, Unpaywall, Europe PMC, DataCite and bioRxiv. Full text is held and indexed only where the license allows it (see Licenses). Embeddings: bge-m3, self-hosted.

Built by Fredrik Sjöberg as a demonstration of research knowledge infrastructure. This instance has its own database; nothing from private notes is shown here.