AIscreenR 0.5.1, published on CRAN on November 13, 2026, changes the package’s default screening model and adds optional model-reported confidence scores. For researchers using it in systematic reviews, the release also introduces compatibility changes that can stop older GPT screening scripts. The release date and version are listed in the CRAN package record.
What changes in existing screening scripts
The 0.5.1 changelog identifies three changes relevant to existing title and abstract screening workflows. Their practical implications are set out below.
| Change | Documented behavior | Implication for existing scripts |
|---|---|---|
| Default model | GPT screening functions now default to gpt-6-luna. | Review calls that omit an explicit model. |
| Older GPT models | GPT-4 and GPT-3.5 models are deprecated within AIscreenR; GPT screening functions stop when these are selected, including with custom_model = TRUE. | Check model settings before rerunning an older workflow. |
| Sampling argument | top_p is removed from GPT screening calls and their output; screen_analyzer() now accepts data without it. | Review function arguments and downstream code that expects that column. |
These are package-level changes. The removal of older models from AIscreenR does not establish that their provider has retired every corresponding model.
What the confidence output means
Setting confidence = TRUE requests a score from 0 to 100 describing how confident the model says it is in its decision. With repeated screening, the aggregated output includes mean_confidence. The release also adds gpt_confidence to report() so confidence can appear in its reports, according to the changelog.
The 0.5.1 reference manual specifies that confidence defaults to FALSE. It applies to the default function-calling setup and is ignored when users provide their own tools. That distinction matters for teams with customized screening integrations: the argument alone does not guarantee that their output will contain confidence scores.
The appropriate interpretation is a model’s self-assessment. A high score does not establish a validated probability that an inclusion or exclusion decision is correct, and averaging those scores does not establish screening accuracy. Assessing observed errors remains a separate task; precision and recall describe performance against reference labels.
Why the release matters for review teams
The maintainers’ project documentation describes AIscreenR as supporting systematic-review screening, including use as a second screener or reducing the references humans need to screen. The new output gives teams another field to inspect alongside decisions, rather than evidence by itself that a review can dispense with human assessment.
For university evidence-synthesis teams and research offices supporting these workflows, the practical response is to document the upgrade as a methodological change. Recording the package version, explicit model, prompt and relevant settings supports reproducibility when comparing runs. The manual documents model, submodel and run information in screening outputs; those records help distinguish a changed configuration from a change in the literature being screened.
Sources
- CRAN: AIscreenR package record. Checked 2026-11-15.
- CRAN: AIscreenR release changelog. Checked 2026-11-15.
- CRAN: AIscreenR 0.5.1 reference manual, Oregon State mirror. Checked 2026-11-15.
- AIscreenR: Maintainers’ project documentation. Checked 2026-11-15.
