Why and How to Do Science When AI Is Better at It: A Framework Grounded in Human Understanding and Direction

Introduction

As artificial intelligence begins to match or surpass human researchers on core scientific tasks, the community faces two urgent questions: why should humans continue to do science, and how AI should be incorporated into the scientific process. The authors argue that the answer to the second question depends on a clear articulation of the first. They define the essential purpose of science as the creation of durable, collective human understanding that serves human ends and values. From this premise they derive a practical framework for evaluating AI’s role in research.

The Two‑Part Test: Human Understanding & Human Direction

The framework rests on two complementary criteria:

1. Human Understanding – The scientific community must retain the ability to explain, defend, and contest the claims produced with AI assistance. This includes preserving the conceptual and methodological skills needed to interpret results, assess uncertainties, and integrate findings into the broader knowledge base.
2. Human Direction – Decisions about which questions matter and what findings mean must stay with accountable humans. This ensures that the goals of research remain aligned with societal values and that responsibility for the implications of discoveries does not become diffused.

Both criteria are evaluated for each phase of the research lifecycle, using the Contributor Roles Taxonomy (CRediT) as a scaffold. By asking whether a given task meets the understanding and direction thresholds, researchers can decide whether to delegate it to AI or retain human oversight.

Applying the Test Across Research Phases

1. Conceptualization & Design

When formulating hypotheses or designing experiments, human direction is paramount. AI can suggest promising avenues based on literature mining, but researchers must judge relevance, feasibility, and ethical considerations.

2. Data Collection & Processing

Automation tools excel at gathering and preprocessing large datasets. However, human understanding is needed to validate data quality, recognize biases, and decide which preprocessing steps preserve the integrity of the scientific question.

3. Analysis & Modeling

AI‑driven analytics can uncover patterns beyond human intuition. The test asks whether scientists can still interpret the models, trace causal inferences, and explain outcomes to peers and the public. If not, the task should remain under human control.

4. Writing & Publication

AI can assist with drafting text, formatting references, or suggesting revisions. Yet the final narrative, argumentation, and framing of significance must be authored by humans to ensure accountability and clarity.

5. Peer Review & Education

Even peer review can be augmented with AI‑based checks for methodological soundness or plagiarism. Nonetheless, the ultimate judgment about the work’s contribution and its alignment with community values must be made by human reviewers and educators.

Implications and Reforms

The authors highlight three ways their framework diverges from existing guidelines:

• Purpose‑Driven: It starts from an account of why science exists, rather than from abstract integrity norms.
• Future‑Proof: It remains relevant even if AI eventually outperforms humans on many tasks, because the criteria focus on the effects of AI on understanding and direction, not on current technical limitations.
• Tool‑Agnostic: It evaluates AI use case‑by‑case, across all research stages, rather than prescribing rules based on the type of AI tool.

To operationalize the framework, the paper calls for institutional reforms: clearer attribution of AI contributions in CRediT roles, training programs that emphasize interpretability and ethical decision‑making, and governance structures that maintain human oversight of research agendas.

Practical Takeaway

Researchers can adopt the two‑part test as a checklist for every project milestone. By ensuring that AI assistance does not erode the community’s capacity to understand and steer scientific inquiry, the discipline can harness AI’s power while preserving its fundamentally human purpose.

--- This article synthesizes the key ideas of the paper without adding new quotations or speculative claims.