AI Screening: Accuracy, Credits, Workflow

What the AI assistant does

AI screening reads every title and abstract in a chosen scope against your written criteria and suggests include, maybe, or exclude with a short reason for each record. It is deliberately recall-oriented: when in doubt it leans toward keeping a record so a human can look at it.

Two safeguards protect recall on every run, at no extra credit cost. First, a record without a usable abstract is never suggested for exclusion; it is routed to you as a maybe, because a title alone is rarely enough to justify excluding a study. Second, every exclude suggestion is re-read in an independent second model pass whose only job is to catch wrong exclusions; anything that could plausibly meet your criteria on a generous reading is demoted to maybe and sent to you instead of out of the review.

The human decides, always. Suggestions are never applied automatically. Each verdict sits next to an Apply button, and nothing changes your decisions until you click it. You can also bulk-apply one direction at a time with Apply all includes or Apply all maybes, which are the safe directions; exclusions are yours to confirm record by record.

Published benchmark results

On the published benchmark, ~200-record labeled samples from five real published systematic reviews in four research fields, each screened against that review's eligibility criteria, the AI assistant kept 229 of 231 relevant records for human review (99.1 percent recall, 100 percent on three of the five corpora) with a 3.5 percent average false-include rate (zero on the post-traumatic stress disorder corpus). Both missed records were borderline judgment calls on full abstracts. The full methodology is published in the measured accuracy section of the tool page, and the cite page provides ready-to-paste methods text that discloses AI assistance transparently.

Run AI screening

1. Enter your include and exclude criteria first. The AI screens against exactly what you wrote, so specific criteria give better suggestions.
2. Sign in, then click the AI screening button in the sidebar.
3. Choose the scope: undecided records only (the default) or all loaded records.
4. Click the Screen with AI button, which shows the record count and matching credit cost before you start.
5. Review the verdicts. Click Apply on the calls you agree with, or use Apply all includes and Apply all maybes, then handle suggested exclusions one at a time.

Credits and pricing

AI screening is metered at 1 credit per record screened. New accounts get a one-time trial grant of 50 credits so you can try it on a real sample. Subscriptions include a monthly credit allotment that resets each month: 400 on Pro and 1,200 on Team, with larger pools on Institution and Enterprise.

When the included pool runs out, buy a top-up pack from the Top up credits section inside the AI screening panel: 1,000 credits for $25, 4,000 for $100, or 10,000 for $250. Packs are one-time purchases and never expire. The same credit wallet pays for AI translation (1 credit per record) and AI data extraction (2 credits per record). If a record cannot be processed, its credit is refunded.