AI can help screen a large set of abstracts, but the quality of the workflow starts with how those records were found. When an AI assistant searches the open web, it can surface press releases, reviews, and other secondary sources alongside original research. Even an accurate summary of a press release offers a different basis for screening than the original study abstract. Start with a documented search of bibliographic databases and preserve the source record for each decision.

Screening introduces a different risk. A model may infer that a study enrolled the population of interest or measured a particular outcome when the abstract never says so. That inference can sound convincing. Ask the model to assess each eligibility criterion against the supplied text, cite the passage it relied on, and flag missing information. At the abstract stage, an unclear criterion can be a reason to advance a record to full-text review.

One prompt can ask a model to screen and check its answer, but that self-check is part of the same assessment and can repeat the same error. Separate the initial decision from verification. A second assessment can compare the proposed decision and supporting passages with the original abstract and eligibility criteria. Disagreement is a signal for review, and agreement can still hide a shared mistake.

Humans should define the protocol, resolve ambiguous cases, and remain accountable for what gets excluded. The goal is to focus their time on disagreements, missing information, and borderline cases. Before relying on that approach, compare the workflow with independent human screening and measure how often it would exclude a record a reviewer would advance. If performance supports selective review, continue auditing a sample of routine exclusions. The workload benefit depends on how much review can be safely avoided and how much time verification adds.