Scientists' Perspectives on the Potential for Generative AI in their Fields

The paper examines how advances in Generative AI—particularly large language models and multimodal systems—may transform scientific practice across life, physical, and social sciences. The research question centers on whether and how Generative AI can add value to scientific work beyond accelerating discovery, including education of future scientists and communication of findings. The context is rapid progress in AI capabilities and growing interest in AI-for-Science applications. The study aims to synthesize domain experts’ perspectives on benefits, risks, and trust requirements to guide responsible, human-centered AI tool design for scientific education, inquiry, and communication.

The related work situates Generative AI within broader AI-assisted science. It highlights the impact of the transformer architecture enabling powerful foundation models for text, images, audio, and video, while noting limitations such as hallucinations, bias, memorization, and training-data provenance concerns. Prior AI (non-generative) has advanced discovery across disciplines (e.g., AlphaFold for protein structure, astronomy, climate, conservation biology, drug discovery, and medical diagnostics). Emerging domain-adapted generative and foundation models (e.g., ClimaX for weather/climate, GenSLM for genomics, Minerva for quantitative reasoning) illustrate early applications; the brief public release and retraction of Galactica underscores factuality risks. Reports and initiatives (e.g., Nobel Turing Challenge, National Academies, Stanford AI Index) reflect momentum in AI-for-Science. The contribution is a human-centered synthesis of scientists’ perspectives on opportunities and pitfalls of Generative AI across education, discovery, and communication.

Design: Qualitative, semi-structured interview study with inductive analysis. Participants: N=20 scientists (7 female, 13 male) spanning physical sciences (e.g., physics, geology, chemistry), life sciences (e.g., immunology, bioinformatics, ecology), and social sciences (e.g., sociology, anthropology, behavioral economics). All held Ph.D. degrees. Computer scientists were excluded by design. Affiliations: All participants were affiliated with Alphabet for confidentiality; nine full-time Alphabet employees and eleven university or research institute faculty with part-time roles at Alphabet (e.g., consulting, sabbatical). All were employed in the United States. Recruitment: Snowball sampling among Alphabet-affiliated scientists and through Alphabet University Relations lists of visiting faculty. Procedure: One researcher conducted all interviews over video (Google Meet) across late January to early February 2023. Sessions lasted 30–45 minutes, were semi-structured, and followed a script with flexibility for probing. Notes were taken in real time; most sessions were recorded and auto-transcribed (four not recorded due to error). No compensation was offered. Interview topics: Background and demographics; discipline challenges and tasks; brainstorming applications of Generative AI to education, literature review, hypothesis generation, experimental design, data collection/analysis, and dissemination; concerns and trust requirements for AI in science. Analysis: A single experienced qualitative researcher used an inductive approach with open coding, iteratively developing themes from notes/transcripts, followed by affinity diagramming to organize and refine themes. Verbatim quotes were marked explicitly; other responses summarized. Findings are organized around emergent themes (education, data, literature reviews, coding, discovery, communication, trust).

• Education: Opportunities include AI-generated lesson materials, intelligent tutoring, scalable Q&A support, jargon translation, and writing assistance benefiting English learners. Concerns focus on AI-enabled cheating, erosion of critical thinking, unequal assessment, and exposure to hallucinations or biased summaries. Some suggest evolving assessments (e.g., oral exams) and teaching responsible AI use.
• Data: Major efficiency gains anticipated in dataset assembly, cleaning, merging heterogeneous sources, labeling, and pattern discovery at scale. AI can aid qualitative data analysis standardization, create synthetic data, generate human-subjects prompts, and extract structured data from literature. Challenges include potential extraction/labeling errors, loss of human interpretive reflection in qualitative analysis, limited applicability to small-data domains, and risks of AI-generated fake data (especially in less-open fields).
• Literature reviews: Desired capabilities include natural-language discovery of relevant papers with direct source links and rationales, multimodal and level-tailored resources, author/venue network mapping, DEI-aware visibility, synthesizing summaries or on-demand review articles, trend identification, gap analysis, impact prediction, and cross-disciplinary bridging beyond jargon silos. Concerns include hallucinations (e.g., fake citations), biased or selective summarization (“politics of summarization”), reinforcement of citation silos, and unclear benefits over current tools without validated metrics.
• Coding: Anticipated value in AI-assisted code generation, environment setup, documentation, API guidance, legacy language support (e.g., FORTRAN), statistical scripting, and guardrails for appropriate analyses. Benefits include speedups (estimates up to 30–40%), improved code quality/reuse, and broadened participation. Core concern is correctness and subtle bugs; IP/memorization risks could complicate ownership of AI-generated code.
• Discovery: AI could guide hypothesis prioritization, optimal sensor placement, experimental design feedback within constraints, and simulation choices; propose next methods, generate lab planning details, and accelerate materials discovery. Novel research question generation by AI was viewed skeptically by most; human creativity remains central. Social science opportunities include LLM-simulated participants, conversational data collection, and societal-scale analyses. Concerns include amplifying research fads, factuality/replicability risks, training on noisy literature, limited generalization to future regimes (e.g., climate), and dual-use harms.
• Communication: Strong interest in AI-drafted first versions of sections (Related Work, Introductions), grant proposals, administrative communications, formatting (condensing abstracts, citation styles), figure/table generation, audience adaptation, and reviewer support (pre-filtering, drafting reviews). Risks include publication spam, misinformation/disinformation, reviewer overload, and feedback loops contaminating future training data.
• Trust: Scientists require source citation with granular references, factuality, confidence estimates, top-N outputs with uncertainties, explainability or interactive validation, clear disclosure of AI use, and demonstrated track records via benchmarks or human baselines. Domain experts should remain in the loop, especially for high-stakes applications. Trust may be earned through repeated accuracy and utility over time.

The study addresses the central question of how Generative AI can augment scientific work by synthesizing experts’ views across disciplines. Findings suggest AI’s near-term role is complementary: accelerating education, data workflows, coding, literature synthesis, experimental planning, and scholarly communication, while preserving human oversight for creativity, critical reflection, and high-stakes decisions. Trust, transparency, and grounding in verifiable sources are pivotal to adoption. The results underscore the need to balance productivity gains with risks such as cheating, biased or hallucinated outputs, publication spam, and dual-use harms. The perspectives highlight that responsible design—confidence metrics, citations, explainability, and human-in-the-loop—will be essential. The potential workforce implications (e.g., displacement of mid-level roles) signal broader societal impacts even as scientific progress may accelerate. Overall, integrating domain expertise into AI development and evaluation is critical to align tools with scientific values and maintain public trust.

Interviews with twenty scientists across life, physical, and social sciences reveal broad potential for Generative AI to enhance scientific education, data handling, literature synthesis, coding, discovery workflows, and communication. Scientists envision AI as an assistive partner rather than a replacement, with human judgment remaining central to ideation, evaluation, and ethical oversight. Realizing benefits requires addressing factuality, bias, spam, and dual-use risks, embedding citations, confidence measures, explainability, and human-in-the-loop safeguards. Future work should co-design tools with domain experts, establish rigorous benchmarks and comparative baselines, and develop protocols for transparency and responsible use to ensure AI augments scientific rigor and public trust.

• Sample restricted to 20 U.S.-based scientists; many disciplines not represented.
• All participants affiliated with Alphabet, likely increasing familiarity with and positivity toward Generative AI relative to peers; potential selection bias.
• Exclusion of computer scientists by design limits perspectives from a field central to AI development.
• Speculation about future capabilities is inherently uncertain; many suggestions may reflect incremental extrapolations from current systems.
• Rapid advances in AI may have outpaced parts of the study (interviews conducted Jan–Feb 2023, preceding GPT-4 release).
• Some capabilities discussed are not unique to Generative AI and may be achievable via other ML methods.