Hacker News
This manifesto proposes a new research culture where AI is not just a tool but a cognitive environment, intended to deepen inquiry, clarify collaboration, and consciously redistribute cognitive effort among researchers.
AI 生成摘要
這份宣言提出一種新的研究文化，將AI視為一種認知環境而非僅僅是工具，旨在深化探究、釐清協作，並有意識地重新分配研究人員的認知努力。
Instantly share code, notes, and snippets.
(for researchers who don’t merely use AI, but work alongside it)
AI 1 is no longer a tool but a cognitive environment.
AI is increasingly embedded in everyday research practice, shaping how problems are explored, results are produced, and work is coordinated. AI is now part of how research is actually done.
This shift changes not only how work is produced, but how thinking unfolds, how collaboration is structured, and how responsibility is assigned. AI compresses exploratory labor, externalizes parts of reasoning, and alters the tempo of intellectual work. These effects are not neutral.
This document outlines an approach to research that embraces intensive and deliberate use of AI.
The aim is not to replace researchers or automate judgment, but to deepen inquiry, clarify collaboration, and redistribute cognitive effort more consciously.
This is not a universal prescription.
It is an opt-in research culture.
Written in 2026, as AI became unavoidable in research practice.
We believe that AI use can meaningfully expedite research and engineering work.
In practice, AI is used to:
AI expands what can be examined and iterated on within limited time and attention. It makes certain forms of exploration cheaper and faster, and enables lines of inquiry that would otherwise be impractical.
Used deliberately, AI accelerates research and engineering work by increasing both the speed and the scope of exploration, comparison, and refinement.
At the same time, we recognize that AI use introduces new limitations and risks.
AI does not determine what is true, relevant, or important.
It does not supply warrant, responsibility, or judgment.
It does not eliminate the need for care, interpretation, or disagreement.
AI can amplify error as easily as insight, reinforce unexamined assumptions, and produce outputs that appear coherent without being well-grounded. Increased speed and fluency raise the risk of premature convergence and overconfidence.
For these reasons, AI use requires active supervision, skepticism toward fluent output, and explicit stopping criteria. The presence of AI does not relax standards; it makes their enforcement more necessary.
We do not optimize for:
We do optimize for:
AI is not treated as a production engine, but as a pressure vessel for ideas.
Everyone maintains a private AI space.
Prompts, chats, rough iterations, speculative paths, and half-formed ideas are personal. They function like a notebook and are not expected to be shared.
What is shared are consolidated artifacts, such as:
How AI is interacted with is inherently personal and context-dependent.
Prompts, iterations, and conversational paths are rarely stable, transferable, or informative outside the situation in which they were used.
Accordingly, shared work is evaluated by what it makes possible for others: understanding, critique, reuse, and revision.
The guiding rule is simple:
What is shared should support others’ understanding, not document individual effort.
Not everyone uses AI with the same intensity.
Not everyone wants to.
This is not a problem to be solved.
Intensive AI use does create real advantages: faster iteration, broader search, and easier synthesis. It would be dishonest to deny this. But these advantages concern throughput and exploration, not truth, warrant, or authority.
What matters for collaboration is not how work is produced, but whether its claims can be understood, questioned, and defended by others.
Accordingly:
Acceleration changes the process, not the fact that we answer to the world.
AI is welcome before meetings, for preparation and structuring.
AI is welcome after meetings, for synthesis and reflection.
AI is not welcome as a participant in the room.
In discussion:
AI may accelerate thinking, but it must not replace social alignment.
Not all lines of inquiry reach a stable or finished form.
It is acceptable to share work that primarily documents open questions, unresolved tensions, negative results, or reasons an approach did not hold. Such work can be essential for collective understanding, even when it does not resolve into a positive claim.
AI may be used to explore alternatives, surface weaknesses, or probe the stability of emerging ideas. Outputs that converge unusually quickly, appear overly polished, or lack visible points of friction should prompt closer inspection.
Where AI use no longer improves understanding, choosing to stop or revert is appropriate. The decision not to apply AI at a given stage reflects judgment about relevance, risk, or diminishing returns, not failure.
AI does not author work.
People do.
Authorship reflects intellectual responsibility: responsibility for claims made, distinctions drawn, evidence selected, and implications accepted. The use of AI does not alter this responsibility.
Using AI does not outsource thinking. It externalizes parts of it. Judgment, interpretation, and commitment remain human acts.
All claims, interpretations, results, and code behavior remain the responsibility of the listed authors, regardless of how they were produced.
AI may assist with exploration, synthesis, reformulation, and drafting or refactoring code. Responsibility for correctness, framing, omissions, reproducibility, and consequences cannot be delegated.
If a result or code artifact cannot be explained, justified, and defended by its authors without appealing to the system that generated it, it is not ready to be shared.
Ethics is not treated as a checklist, but as a continuing practice of questioning.
Relevant questions include:
Environmental cost
What energy, compute, and material resources does this mode of work consume, and how are those costs distributed across people, institutions, and environments?
Work and employment effects
Which tasks are automated, reduced, or eliminated by AI use, which new tasks are created, and who loses work, gains work, or takes on additional risk as a result?
Concentration of power
Who controls the models, data, infrastructure, and terms of access that make this work possible, and how does that concentration shape who can participate, compete, or meaningfully question dominant systems?
Evidence and challenge
When AI-generated or AI-assisted outputs are used, what counts as acceptable evidence or explanation, and how easily can those results be checked, reproduced, or challenged by others?
Implicit assumptions
What modeling choices, training data, defaults, or abstractions enter the work without explicit acknowledgment, including biases, value judgments, filtering, or content restrictions, and how accessible are these influences to scrutiny, critique, or revision?
Originality and derivation
How is the contribution of the work distinguished from recombination or imitation, and how clearly can its sources, transformations, and novel elements be identified and justified?
Epistemic risk
Where does increased speed encourage premature convergence, automation bias, or erosion of interpretive care?
Scientific responsibility
At what point does AI use stop supporting understanding and begin to compromise the standards the work is meant to uphold?
These questions have no final answers.
They must be revisited as tools, practices, and contexts change.
This way of working is not a requirement for others.
Precisely because methods will differ, collaboration and evaluation cannot depend on shared processes. They depend on what is ultimately put forward for scrutiny.
Accordingly:
AI does not make thinking effortless.
It redistributes effort, compressing some forms of labor while making judgment more visible.
Understanding still takes time.
So does disagreement.
So does learning what can and cannot be trusted.
Those constraints remain.
Here, “AI” refers primarily to contemporary machine-learning systems, especially large-scale generative models (e.g., language, vision, and multimodal models) capable of producing text, code, images, or other structured outputs in response to prompts. More broadly, it also includes the surrounding ecosystem: training data, model architectures, inference infrastructure, fine-tuning methods, evaluation practices, tooling, and organizational workflows that integrate these systems into research and engineering practice. The term is used in the broad, informal sense in which “AI” is commonly used today, rather than as a precise technical or theoretical category. ↩