Permanent record
RIR-3059 · Logged 14 October 2026
Does AI screening mark down research summaries written in a second language?
Large language models increasingly help screen and rate research writing, and they have been shown to prefer text written by other language models. It is unclear whether that preference partly penalises people writing in an additional language. The idea tests whether AI ratings of research abstracts differ between first-language and additional-language South African writers, and whether AI editing closes any gap once underlying quality is held constant.
In plain words
AI tools are starting to help judge research abstracts and proposals. This study would test whether they give lower scores to South Africans who write English as a second language, and whether letting AI polish the text hides the difference.
Primary research question
When large language models rate the quality of research abstracts, do they score South African abstracts written in English as an additional language lower than comparable first-language abstracts, and does AI editing remove the difference once underlying quality is held constant?
Knowledge gap
What remains worth asking
Laurito et al. (2025) found that language models prefer text generated by language models, while Sharma, Joachims and Dean (2026), studying real conference reviews, report that the apparent preference disappears once paper quality is controlled. Neither tested African Englishes or writers using English as an additional language.
Potential contribution
Why it may matter
If AI screening penalises additional-language writing, institutions and funders that adopt it could disadvantage many African researchers without noticing.
Academic placement
Academic fields and topic tags
Study frame: Matched sets of South African research abstracts in first-language, additional-language and AI-edited versions, rated by several language models, with independent expert quality ratings as the control. · Method signals: Controlled experiment, Matched-text comparison, Statistical modelling
Possible study pathways
One question, different levels
Honours
Rate a small matched set of abstracts with two language models and compare scores before and after AI editing.
Research master’s
Build a larger matched corpus with expert quality ratings and model the language effect.
Doctoral
Extend to several African Englishes and to real screening settings such as grant or admission review.
- originality
- High
- methodology
- Moderate
- Data access
- Moderate
- ethics
- Moderate
Quality and audit
87/100 quality
Safeguard-based audit was not recorded under the prior standard.
- Named population and an explicit quality control
- Engages a published counter-finding
- Practical relevance for African researchers
This older record can be checked against the current standard on request. It changes only if every academic gate and independent safeguard passes.
Sign-in required · uses one AI creditProvenance
Research Idea curation
- DOI and bibliographic metadata independently resolved through Crossref
- The source reports the AI–AI preference that the idea tests
- The additional-language comparison is an inferred direction
The public contributor code contains no name or account email.
APA 7 source
Laurito, W., Davis, B., Grietzer, P., Gavenčiak, T., Böhm, A., & Kulveit, J. (2025). AI–AI bias: Large language models favor communications generated by large language models. Proceedings of the National Academy of Sciences, 122(31), Article e2415697122. https://doi.org/10.1073/pnas.2415697122
Discussion and limitations
This is an AI-inferred direction. The source authors did not state it as a recommendation.
Open source ↗