*It was a rather long prompt
I wanted to see if I could get an AI agent to complete a first draft of a paper, including data analysis, searching for references on the web and writing a complete draft.
I used the Roo Code agent software with the Claude Sonnet 4.0 large language model.
I’ve shared the results from replicate 1 here.
I’ve also pasted the abstracts from the two replicates below. Both used identical prompts and settings in Roo Code. Its interesting how they are different. For instance the first one frames the study in terms of estimating the scale of human impacts. The second frames the study as being about studying an under-studied mechanism.
The experiment started with a project folder with data and detailed instructions for the research question, the data and the analyses I wanted. The instructions included details of R packages and specific statistics, for instance I told the model to use Bray-Curtis distances for the multivariate analysis.
I also provided basic instructions for how to write the paper. The agent had web search capabilities, allowing it to access , and read real references, and then incorporate insights from those into the paper.
It is important to note that this draft has not been edited by a human. Parts of it are not scientifically accurate. It should not be read as anything other than an experiment. It contains errors and some of the references are false.
For instance, it made up the following reference, but oddly linked it to a real reference (that is different) using a real DOI:
Hamilton, Richard J et al. 2025. “Modeling Sediment Impacts on Juvenile Reef Fish.” Frontiers in Marine Science. https://doi.org/10.3389/fmars.2025.1513498.
It didn’t make much of an effort to create good instructions for writing a paper. I’m sure you could do a lot better than this with more careful prompting.
AI generated papers should never replace human oversight. If you rely solely on AI, you miss the chance to truly understand your subject. Science is about learning, not just producing content.
What this sort of thing could be useful for is as a way to iterate and quickly explore alternative options for how to write a paper, or explore alternative interpretations.
It can also help you get over ‘the fear of the white page’ (writers block), by giving you something to start editing.
Terrestrial logging threatens coral reef ecosystems through increased sedimentation, yet the spatial patterns of impact remain poorly understood. We examined benthic community composition at 49 coral reef sites across a gradient of logging impacts in Kia Province, Solomon Islands, spanning distances from 0.20 to 22.4 km from logging operations. Using multidimensional scaling ordination and permutational multivariate analysis of variance, we tested for differences in community structure between logged and not logged sites and across distance categories. Benthic communities differed significantly between logged and not logged sites (PERMANOVA p = 0.027), with logged sites showing 14.2% lower coral cover (48.0% vs 62.2%). Distance-dependent effects were highly significant (p = 0.002), explaining 11.9% of community variance. Sites within 1 km of logging showed severely degraded conditions (31.2% coral cover, 33.3% sand cover), while sites beyond 10 km maintained high coral cover (62.9%). Recovery followed a clear gradient: moderate distance sites (1-3 km) had 55.3% coral cover, far sites (3-10 km) had 60.4% coral cover. Categorical distance analysis proved more informative than continuous distance relationships, indicating threshold effects rather than linear gradients. These findings demonstrate that logging impacts extend further than previously documented, with measurable effects at sites 1-3 km from operations. Our results provide quantitative evidence for establishing buffer zones around coral reefs and highlight the need for watershed-scale conservation planning that integrates terrestrial forest management with marine protected area design.
Coral reef ecosystems face increasing threats from terrestrial activities, with logging-induced sedimentation representing a significant but understudied impact mechanism. This study examined benthic community structure across 49 coral reef sites in the Solomon Islands using multidimensional scaling (MDS) ordination to test the hypothesis that logging activities influence reef community composition through sedimentation. We applied non-metric multidimensional scaling to benthic cover data using both Euclidean and Bray-Curtis distance metrics, with environmental vector fitting to identify ecological gradients. The analysis revealed strong community gradients primarily associated with coral branching cover (r² = 0.82) and soft coral cover (r² = 0.62), while water clarity showed moderate correlation (r² = 0.24) with community structure. Contrary to expectations, distance to logging showed weak correlation with community patterns (r² = 0.056, p = 0.101), suggesting that logging impacts operate through pathways not captured by simple proximity measures. The results demonstrate MDS ordination effectiveness for revealing coral community structure patterns while highlighting the complexity of logging-reef relationships in tropical marine systems.
Designed by Chris Brown. Source on Github