AI use and scientific accountability
AI assisted substantial parts of Fire VASE development. It was not the observational data source, an autonomous author, or the final scientific authority.
Plain-language answer
OpenAI Codex/ChatGPT supported implementation, analysis, visualization, documentation, drafting, review, citation work, and quality control. Human investigators directed the analyses, selected claims, reviewed code and outputs, verified calculations and citations where reported, and remain responsible for the manuscript’s integrity and interpretation.
Where AI was used
The documented project record identifies these activities:
- Python code and data engineering for ingestion, lakehouse construction, climate attribution, morphospace analysis, and report production
- statistical and analysis support, including PCA, null models, prediction, matching, and validation workflows
- figure development, rendering, visual checks, and correction of misleading climate-color mappings
- documentation, schemas, manifests, tests, notebooks, and reproducibility handoffs
- manuscript and figure-legend drafting and revision, simulated review, and response-to-review support
- candidate-literature organization, citation checking, and author-guideline compliance work
These categories come from the preserved statement and prompt-log basis. The record documents nine summarized assistance items; it is not a raw, line-by-line transcript of every private prompt.
Where AI was not the scientific authority
Source evidence
Not generated by AI
FIRED/MODIS observations, gridMET, PRISM, and other observational sources retain their own provenance and citation requirements.
Computed evidence
Produced by executable workflows
Tables, models, figures, and diagnostics come from versioned scripts, configuration, declared real inputs, and recorded seeds. Missing required inputs stop the current workflow; no synthetic fallback supports manuscript claims.
Scientific responsibility
Human investigators
Study design, data selection, estimands, analytic choices, claim boundaries, interpretation, citation accuracy, authorship, and final manuscript content remain human responsibilities.
How AI-assisted work was checked
→
→
→
→
Checks documented in the repository include scientific invariants, numerical replays, source and exposure audits, null models, sensitivity analyses, expected-failure controls, citation audits, rendering checks, schema and test suites, and frozen claim-to-source records.
Important qualification. Reproducible computation can show what code did and whether declared checks pass. It cannot by itself establish that an estimand is scientifically sufficient, remove remote-sensing uncertainty, or replace expert judgment. The current manuscript says human scientific review remains necessary.
Can I inspect the record?
Yes. The concise statement and generated report remain unchanged in purpose, and the underlying repository evidence is directly accessible.
- Read the concise AI transparency statement
- Inspect the expanded artifact-level report
- Read the prompt and implementation log
- Inspect the scientific-validation record
- Inspect the submission freeze and claim registry
- Open the notebook index
- See how validators detect deliberately wrong cases
Current artifact-check record
The stored artifact-level reproducibility report records the data lake and 12 derived-statistics comparisons as passing. Four of five PNG pixel comparisons pass; Figure 3 records a small size mismatch. This is disclosed rather than converted into a blanket “all outputs identical” statement.