Verona Declaration on the Use of Generative AI in Computing Research

15 November 2026 By In IFIP TC12 Blog

Section 01

Preamble

Science advances by creating knowledge that can be critically examined and trusted. Generative AI is changing how research is conducted, but not the standards that yield credibility of scientific results.

Generative AI can produce convincing but incorrect results, reinforce users' assumptions, and generate output faster than researchers can critically evaluate it. Production of material that is of potential scientific interest may become boundless, but what we can accept as scientific results is bounded by the capacity of human scrutiny.

Scientific responsibility remains human. Researchers must be able to justify their claims, assess the evidence, explain their methods, and stand behind their conclusions. We call this epistemic control: retaining the understanding and judgment needed to direct and critically assess one's own research, including work supported by generative AI.

This is a living declaration. Stable, citable versions will coexist with an open invitation to comment, challenge, and improve its principles as research practices and generative AI capabilities evolve.

Section 02

Purpose and Boundaries

This declaration addresses the use of generative AI in computing research, including computer science, computer engineering, AI, software engineering, information systems, data science, and related fields. It covers generative models and generative AI-based agents that support or automate research. Its principles apply to research offered as a contribution to scientific knowledge, whether conducted in academia or industry. They can also guide confidential and internal research, with transparency adapted to legitimate constraints.

The declaration does not prohibit generative AI use, prescribe a single workflow, or constitute binding regulation. Delegation of scientific tasks to generative AI is legitimate only where researchers can exercise meaningful responsibility. Developing the capacity to do so is also an essential goal of research education.

 

Section 03

Our Unique Position: The Dual Role

The computing research community both develops generative AI systems and uses them in its own research. This gives us a dual responsibility: (i) to design systems that enable scientific scrutiny, provenance, reproducibility, and meaningful human oversight, and (ii) to use these technologies responsibly. We should share what we learn with other research communities.

 

Section 04

Why Scientific Integrity Matters

What distinguishes science is not access to powerful tools, but the assurance that claims are supported by sound methods and evidence, open to scrutiny, and backed by accountable researchers. This assurance underpins society's trust in science and science's continued legitimacy.

Generative AI can help produce and verify results, but researchers must do more than approve its outputs: they must continue to shape scientific questions, understand the work, and take responsibility for its conclusions. The appearance of science has become cheap to produce. Science has not.

 

Section 05

Our Values

These five values describe the assurances that should accompany every scientific claim, with or without generative AI.

01

Scientific Integrity and Validity

Scientific claims should rest on sound methods and evidence, plausible presentation without these is insufficient. These standards apply to established methodologies and to human and generative AI-assisted work alike. Fabrication, falsification, and misleading representation are unacceptable, regardless of whether they stem from negligence or intent. Researchers remain responsible for all parts of their publications.

02

Responsibility, Attribution, and Accountability

Scientific contributions must be fairly attributed. Researchers remain accountable for the scientific quality and ethical aspects of their work. Generative AI assistance neither transfers responsibility nor substitutes for substantive human intellectual contribution.

03

Transparency, Verifiability, and Reproducibility

Researchers should disclose generative AI use in proportion to its influence on scientific claims. Where a generative AI system forms part of the evidence for a claim, it should be reported as a research instrument. Reproducibility should be pursued where feasible, while recognizing limitations of changing or inaccessible systems and legitimate privacy, security, and confidentiality constraints.

04

Scientific Understanding and Judgment

Science should not merely produce outputs, but also develop human understanding and independent judgment. Researchers and research communities should retain authority over scientific questions, methods, and evaluation standards. Generative AI may inform scientific judgments, but consequential decisions about validity, peer review, funding, hiring, and publication acceptance must remain under meaningful human authority.

05

Responsible Innovation and Societal Trust

Generative AI-assisted computing research should advance understanding and innovation while critically examining the limitations and risks of its methods and results. Claims of effectiveness, safety, and societal benefit should be proportionate to supporting evidence.

 

Section 06

Opportunities for Science

Generative AI can accelerate discovery, broaden participation, support verification and documentation, and expand the questions researchers can investigate. It can augment human creativity and free up time for formulating questions, designing studies, and interpreting evidence. These opportunities are real possibilities, but gains are not automatic: better and more research requires preserving the practices that make knowledge trustworthy.

 

Section 07

Risks to Science

The following risks are not inevitable outcomes of generative AI. Some are emerging, others can still be prevented. Several are structural and cannot be addressed through individual diligence alone.

01

Plausibility Without Validity

Generative AI can produce convincing papers, proofs, code, architectures, and analyses without sound evidence or reasoning. Invalid results can propagate through the scientific record and undermine trust. Public attention and commercial promotion can also amplify generative AI research claims beyond what independent evidence and scientific scrutiny justify, in turn distorting research priorities and expectations.

02

Overwhelmed Scientific Evaluation

Research output can grow faster than experts can critically assess it. Scarce reviewer capacity or attention may turn polished presentation into a substitute for substantive evaluation.

03

Automation Bias in Scientific Decisions

Confident AI assessments could be accepted without independent scrutiny, reducing oversight to nominal approval in research, peer review, funding, hiring, and supervision.

04

Distorted Incentives and Unequal Access

Generative AI may favor questions that are easy to automate over those that matter most. Funding and evaluation practices may reward enthusiasm for generative AI, or penalize justified criticism, rather than assess scientific merit. Unequal access to models, data, and compute can concentrate both research opportunities and the ability to verify published claims.

05

Loss of Provenance and Attribution

Generative AI-mediated synthesis can obscure the origins of ideas, evidence, software, and data, weakening recognition and traceability of scientific contributions.

06

Dependence and Structural Irreproducibility

Reliance on opaque or externally controlled systems can undermine scientific independence and make findings difficult to verify when models change, become inaccessible, or are withdrawn.

07

Erosion of Expertise and Diversity

Excessive delegation can weaken the skills needed to supervise generative AI-assisted work. Reliance on similar systems may narrow methodological diversity, independent thinking, and originality.

 

Section 08

The Essential Trade-off

Scientific progress depends on three goods: efficiency (doing more with the same resources), effectiveness (producing knowledge that is valid), and resilience (the scientific system's capacity to detect and correct its own errors). Generative AI most readily increases efficiency while, as outlined by the risks above, potentially eroding effectiveness and resilience. Incentives that reward volume, such as paper counting, may lead to an imbalanced emphasis on efficiency. The practices that sustain effectiveness and resilience (review, replication, maintained human expertise, and independent inquiry) require time and resources. Where efficiency conflicts irreducibly with effectiveness or resilience, efficiency must give way.

 

Section 09

The Delegation Principle

To operationalize the above insights, we frame the use of generative AI as delegation. Under the established principle of delegation, execution can be transferred, but accountability cannot. This remains true when delegating to generative AI. Human researchers remain answerable for the methods, evidence, claims, interpretations, conclusions, and artifacts they present. Generative AI may extend researchers' capabilities, but must not replace their independent scientific judgment. Responsible delegation therefore requires epistemic control. This means that researchers themselves determine whether the outputs of delegated work can support a scientific claim, and can explain why. Disclosure or nominal approval of these outputs does not establish such control.

01

Competence

Researchers should delegate only work they have sufficient understanding to assess. The respective judgment must be developed in training and maintained through practice.

02

Verifiable Tasks

Delegated work should be structured in such a way that consequential outputs, assumptions, and evidence can be meaningfully verified. This might require intermediate checks when end-result verification is insufficient.

03

Proportionate Oversight

Verification should reflect a task's influence on scientific claims and the consequences of error. Oversight must remain feasible given the complexity and volume of output.

04

Traceable Decisions

Researchers should retain checkpoints, records of consequential decisions, and the ability to question, redirect, or stop generative AI-assisted work.

 

Section 10

Shared Commitments Across the Scientific Ecosystem

Responsibility for trustworthy research is shared. It begins with individual researchers and teams, but because several risks are structural, it extends to actors whose decisions shape research practice.

01

Researchers and Research Groups

Research groups and individual researchers should select generative AI systems for their scientific suitability, expected benefits, and verifiability. Using generative AI is not an end in itself. Researchers and research groups should preserve competent oversight, verify consequential generative AI-assisted work, disclose material generative AI use, protect confidential information, and develop students' independent expertise.

02

Institutions, Publishers, and Scientific Venues

Institutions, publishers, and venues should reward quality over volume, provide secure and governed generative AI access, establish proportionate disclosure and reproducibility expectations, and safeguard meaningful human judgment in peer review, funding, hiring, and promotion.

03

Funders and Policymakers

Funders and policymakers should evaluate scientific merit independently of whether proposals promote or criticize generative AI, and support open and durable research infrastructure, compute capacity also available for public verification, equitable access, and diverse research methods, including work that generative AI does not readily accelerate.

04

AI Providers and Industry Partners

AI providers and industry partners should enable independent scrutiny through stable model versions, research access, and support provenance by preserving model histories, while protecting researchers' freedom to publish and criticize results, including unfavorable ones.

 

Section 11

A Call to Action

Responsible generative AI-assisted research requires collective action. We invite the global computing research community to put these principles into practice, challenge and refine them, and shape a future in which generative AI strengthens scientific discovery without compromising what makes science trustworthy.

Read 8 times
Login to post comments

Tag Cloud

Newsletter

captcha 

Contact us

Send us an email, or use the contact form.

Connect with us

Facebook Twitter Google plus  LinkedInfeed
You are here: Home Blog Verona Declaration on the Use of Generative AI in Computing Research