Awareness of AI tools

• Mishra et al. (2024): 87.6% of clinical researchers were aware of large language models; those aware had a higher number of publications (p < 0.001).
• Mohammadi et al. (2026): 73.1% of publishing academics had heard “a lot” about generative AI tools; only 3.4% reported no familiarity.
• Chugunova et al. (2026): 42.4% reported being very or rather familiar with AI tools; a further 44.0% had used them a few times.

Frequency of use

• Chugunova et al. (2026): 25.9% use AI tools at least once a day; 22.2% never use them for any work task.
• Mohammadi et al. (2026): 54.7% use generative AI at least monthly for academic purposes (28.2% never). For research specifically, 62.5% use it at least monthly and 13.4% never.
• Mishra et al. (2024): among aware respondents, 18.7% had used LLMs in a publication; 81.3% had not.
• Van Noorden & Perkel (2023): in 2023, none reported using AI regularly for research, and 17.4% used generative AI weekly or more often.

Comparing the 2023 and 2024 surveys

Chugunova et al. (2026) report a shift from 2023 (Van Noorden & Perkel, 2023):

• In 2023, regular research use of AI was essentially absent; by mid-2024, a quarter of researchers used AI daily (Chugunova et al., 2026).

Reported non-use moved in the opposite direction:

• Chugunova et al. (2026) attribute this to the 2023 sample being skewed toward researchers already interested in AI.

Six domains of academic AI use

Khalifa & Albadawy (2024) group the literature into six domains:

Reported uses across the research process

Most common research uses

• Chugunova et al. (2026): the two most frequent uses are piloting and testing (47.9%) and writing code (43.2%); 50.6% use AI for two to five distinct tasks.
• Mohammadi et al. (2026): the most common research uses are translating text (13.5%), proofreading drafts (13.0%), synthesising initial drafts (12.5%) and literature reviews (12.5%); supporting data analysis is lower (6.4%). These tools were mainly used for academic writing rather than data analysis.
• Mishra et al. (2024): among those who had used LLMs, the most common stages were grammatical errors and formatting (64.9%), writing (45.9%) and revision and editing (45.9%).
• Van Noorden & Perkel (2023): in 2023, the single most common use was creative fun unrelated to research.

Disclosure of AI use

• Mishra et al. (2024): among researchers who had used LLMs in a publication, 40.5% did not acknowledge that use.
• Mishra et al. (2024): 58.1% thought journals should allow AI use in research, and 78.3% thought regulations (for example modified journal policies, AI review boards, or detection tools) should be in place.

Role, field and age

• Chugunova et al. (2026): familiarity and use decline with age and grow with education; familiarity is highest in the social sciences. Clustering researchers as leaders, builders and analysts, “leaders” are the most positive toward AI.
• Mohammadi et al. (2026): PhD students are the highest users, followed by assistant professors and instructors. The social sciences show the highest adoption, then engineering. Highest adoption by country: Taiwan, South Korea, India, Iran; lowest: United States, United Kingdom, China, Russia.
• Mishra et al. (2024): only number of publications was significantly associated with awareness; age, region and field of practice were not.

The gender gap

• Chugunova et al. (2026): the gender gap in AI use for research is almost entirely explained by observable characteristics, with familiarity the dominant factor (71% of the explained gap, 99% of the total gap). Among non-users, women rated AI as more helpful than men did. 45.9% of women versus 34.8% of men had not consulted, and did not intend to consult, resources such as prompt engineering guides.
• Mohammadi et al. (2026): women were 10% less likely to use generative AI frequently (daily or weekly) for research, 4% less for teaching, and 3% less for administrative tasks.

Prompting as a skill

Perceived future impact

• Van Noorden & Perkel (2023): more than half expected AI tools to become “very important” or “essential” in their field within a decade.
• Chugunova et al. (2026): 69.2% expect AI to transform or revolutionise their field within a decade.
• Mishra et al. (2024): 52.0% expect a major overall impact. Areas seen as most impacted: grammatical errors and formatting (66.3%), revision and editing (57.2%), writing (57.2%) and literature review (54.2%). Least impacted: methodology, journal selection and study ideas. 50.8% anticipated a positive future impact; 32.6% were unsure.

Perceived benefits

Mohammadi et al. (2026), share agreeing or strongly agreeing:

• Personalised tutoring or instruction: 37.4%
• Improving problem-solving abilities: 35.2%
• Enhancing students’ learning: 34.5%
• Fostering creativity and ideation: 31.2% agree vs 35.9% disagree
• Accuracy and reliability of AI content: 20.6% confident vs 48.7% disagree
• Consistency of AI content: 20.0% agree vs 50.1% disagree

Concerns about AI use

• Van Noorden & Perkel (2023) (machine learning): 69% more reliance on pattern recognition without understanding; 58% entrenching bias; 55% fraud easier; 53% irreproducible research. On generative AI: 68% misinformation; 68% easier plagiarism; 66% mistakes in papers.
• Mohammadi et al. (2026) (moderate or extreme concern): inaccuracy 67.8%; plagiarism 65.0%; discouraging critical thinking 61.7%; lack of transparency 59.2%; lack of explainability 57.8%; intellectual property 52.2%; data privacy 49.0%; bias 41.3%.
• Mishra et al. (2024): misinformation 66.7%; unintended bias 65.7%; copyright infringement 63.1%; impaired creativity 60.1%; lack of accountability 58.1%; data security 50.5%.

Barriers and regulation

• Chugunova et al. (2026): the most common top-two barriers are legal uncertainties (17.6%), lack of knowledge (17.4%) and suitable tools not available (16.6%).
• Chugunova et al. (2026): on where guidance should come from, 58.7% look to supranational bodies (for example the EU), 51.3% to research societies, 49.0% to professional associations.
• Mishra et al. (2024): 78.3% thought regulations should be put in place; 58.1% thought journals should allow AI use.