Introducing: A Strategy for AI in Research
A blueprint for research institutions: authored evidence first, local control of processing, and a reviewed frontier bridge that never takes the archive.
We've just published a strategic blueprint for how research institutions can approach artificial intelligence without trading authorship for speed. It's called A Strategy for AI in Research, and this piece is the short version: what's in it, why we built it this way, and who it's for.
AI can help researchers find evidence, explore methods and question their conclusions. Its value depends on whether that assistance makes the research stronger, not simply faster.
Why does this strategy exist?
The short answer is that most research teams are being asked to have an AI policy before anyone has given them a coherent way to think about what AI does to a claim. So they either lock it out, or let it in with no more structure than a chatbot window and a hope that a fluent answer is a justified one. Neither is a strategy. Both avoid the actual question.
The question the blueprint organises around is simple: what supports this claim? A convincing answer is not enough. Researchers should be able to see what supports a statement, where the evidence is missing, and which decisions remain theirs.
That is not an abstract worry. Messeri and Crockett's 2024 Nature Perspective describes how AI can encourage illusions of understanding and scientific monocultures, in which particular methods and viewpoints displace alternatives. The danger is not that AI is present in the lab. It is that a fluent result becomes a substitute for the intellectual work needed to justify it.
What does authored knowledge at the centre mean?
This strategy puts authored knowledge at the centre. In a knowledge-centric research workflow, approved protocols, literature, datasets and research records provide the evidence context. The system should retrieve relevant material, expose authors and versions, and distinguish what the sources say from what the model infers.
When evidence runs out, the preferred response is 'not found in this collection', followed by a way to inspect the search or extend its scope. Conjecture remains legitimate. It should be labelled as exploration, and verified before anyone accepts it as a finding.
We contrast that with an LLM-centric workflow that has no access to the project archive. Asked which exclusion rule was approved for a study, it can invent a plausible rule and, on a follow-up, invent an amendment that never existed. The researcher still has to check the source. The left-hand behaviour is an acceptance requirement. The right-hand answers illustrate a failure mode.
Does local AI make an answer correct?
No. Local AI gives institutions greater control over their research environment. Keeping documents, prompts, embeddings, retrieved passages, outputs and inference within an institutional boundary can reduce exposure. It does not make an answer correct.
A local model without governed retrieval can still be LLM-centric. A cloud service with verified retrieval can support a knowledge-centric workflow, although its data boundary is different. Local control protects the research boundary. Knowledge grounding makes the evidence inspectable. Human judgement determines whether the claim is justified.
Buhler, Pérez and Boettiger's 2026 Nature Comment argues that science's infrastructure needs differ from those of mass-market AI platforms, and that local, open-weight models can be a more accountable alternative. We map that argument into institutional choices: size services around actual research tasks, keep a local execution option, preserve exit routes, and evaluate the complete system rather than the model in isolation.
How can frontier models help without taking the archive?
Frontier models can contribute without receiving the whole research archive. The frontier bridge prepares a limited, abstract question for the researcher to review before sharing. The response returns to the local environment, where it is checked against the original material before anything is accepted.
The pattern is a controlled round trip. The local Space first establishes what the evidence already supports. It then drafts a minimum-context brief, runs a leak check, and requires approval of the exact payload and destination. Only that approved question crosses the boundary. The returned answer is treated as untrusted advice until a named researcher accepts, partly accepts, or rejects it.
That is the ethical use the title of this piece names: local processing as the default, frontier assistance as a reviewed exception, and authorship remaining with the people who will have to defend the work.
Empathy.AI demonstrates this approach through a bounded research experience. The aim is to help researchers explain, challenge and defend their work while retaining control of its evidence and conclusions. The instrument is Empathy.AI Spaces, running locally or on dedicated private capacity, with the same knowledge-centric pattern we described in A Strategy for AI & Education.
What is this blueprint not?
It is not a sales deck dressed up as research, and it is not a claim that owning hardware automatically produces a defensible paper. We say so directly in the document. Local AI, left ungoverned, carries the same risks as anyone else's chatbot. Privacy describes a controlled relationship. Local describes an execution boundary. Neither word, by itself, proves that a scientific method is valid.
It also is not finished in the sense that matters. The document is a November 2026 discussion draft: strategic guidance, not a commercial offer or a legal opinion. It makes no claim of Oxford endorsement, guaranteed absence of unsupported answers, or a proven research outcome from Empathy.AI. The immediate recommendation is a 12-week bounded pilot, with owners and evidence gates agreed before any wider procurement.
If that sounds like the conversation your research group, library, or integrity office has been putting off, the full document is where we've put the actual thinking. We'd like to have that conversation with you.
References
This article and the strategy document it summarises were drafted with the assistance of AI, then reviewed and directed by a human in the loop. Claims throughout are drawn from the following sources:
- Messeri, L., Crockett, M. J. — 'Artificial intelligence and illusions of understanding in scientific research', Nature, March 2024
- Buhler, C. K., Pérez, F., Boettiger, C. — 'Why scientists should lead the shift away from AI mega data centres', Nature, 11 August 2026
- A Strategy for AI in Research (Empathy.AI, November 2026), the full strategic blueprint this article summarises
- A Strategy for AI & Education (Empathy.AI), the education blueprint this research strategy adapts
- University of Oxford, Policy on the use of generative AI in research
- About Empathy.AI Spaces, product overview and deployment options
Frequently Asked Questions
Why does research need a strategy for AI?
Fluent AI answers can replace the work of deciding which evidence supports a claim. Messeri and Crockett's 2024 Nature Perspective describes this as illusions of understanding and scientific monocultures. A strategy exists so institutions adopt AI that keeps those judgements with the researcher, instead of treating speed as the measure of success.
What does a knowledge-centric research workflow do?
It puts authored evidence at the centre of every claim. The system retrieves approved protocols, literature and records, exposes authors and versions, and distinguishes what the sources say from what the model infers. When evidence runs out, the preferred response is 'not found in this collection', not a convincing guess.
Does running AI locally make an answer correct?
No. Local processing keeps documents, prompts, embeddings and inference inside an institutional boundary, which can reduce exposure. It does not make a claim true. A local model without governed retrieval can still be LLM-centric. Sources, methods and interpretations still need human scrutiny before anything enters the research record.
How can frontier models help without taking the archive?
Through a frontier bridge. The local environment prepares a limited, abstract question for the researcher to review before sharing. Only that approved payload leaves. The response returns to the local environment, where it is checked against the original material. Nothing is accepted until a named researcher decides.
What does Empathy.AI propose for research institutions?
A governed research knowledge environment using Empathy.AI Spaces, locally or on dedicated private capacity. Four capability lines: evidence discovery, methods assistance, research-record stewardship, and communication support. The November 2026 discussion draft recommends a 12-week bounded pilot before any wider procurement.
What does this strategy refuse to claim?
It is not a commercial offer, legal opinion, or Oxford endorsement. It does not claim that privacy guarantees truth, that local AI eliminates unsupported answers, or that Empathy.AI has already proven a research outcome. Local AI, left ungoverned, carries the same risks as any other chatbot.
