AI & Technology

PubMed adds 4,600 papers a day. AI has to prove it read the real ones

According to the National Library of Medicine, PubMed added roughly 1.7 million citations this past year, taking the database past 39 million records.

That works out to more than 4,600 new papers every day. For a medical affairs team building a publication plan or a launch narrative, any one of them could shift what the evidence says, and nobody on that team has time to read more than a sliver of them.

For years the workaround was patience. For instance, a 2018 case study cited in a 2025 Frontiers in Pharmacology review put the average systematic review at 66 weeks, so most companies settled for narrower searches, older snapshots and analysts reconciling sources late into the night before a congress deadline. GenAI looked like a way out, since it could read faster than any team and draft a summary in seconds.

However, it also carried a new problem into the evidence base. In April, a Nature analysis estimated that tens of thousands of 2025 publications probably contain invalid references generated by AI.

A month later, an audit published in The Lancet checked 97 million references across roughly 2.5 million biomedical papers, seeing the sharpest climb of citations of studies that didn’t exist in 2024, aligning to the adoption of AI tools.

In medicine those papers feed clinical guidelines, so a phantom study at the base of the evidence chain carries into whatever gets built on top of it.

Speed has stopped being the hard part. Buyers now want to know whether every sentence can be walked back to a real paper, and what the system does when it can’t find one.

That is the gap Prezent Vivo is aiming at with Literature Review, a new capability built on Astrid, the company’s AI agent tuned for life sciences.

According to Rajat Mishra, Prezent Vivo‘s CEO and co-founder, “Every scientific story, strategy, and publication plan starts with the literature. But scientific literature never stands still.”

Astrid searches trusted databases such as PubMed and Europe PMC and assembles a structured review of the published data on a topic, whether a team needs the latest findings or the older record. The design choice Mishra stresses sits in the output: “Every claim is traced to its source, and anything Astrid can’t support is flagged for review.”

Surfacing an unsupported claim for a human, instead of smoothing it into fluent prose, is the behaviour that protects a team when its work reaches regulators and the physicians it briefs.

The payoff, in Mishra’s telling, is time handed back to specialists. Prezent has said senior medical communications experts can spend up to 40% of their hours on production work. Literature Review, he says, gives teams “a faster, more thorough starting point for the work only they can do,” which he describes as shaping the narrative, informing strategy and identifying the gaps that guide the next evidence generation plan.

A review that marks clearly where the published data runs thin also tells a company where its next study should go.

In a year when the biomedical record itself is picking up invented references, the promise buyers weigh most may be the second half of Mishra’s pitch: that what comes back has been checked against its source.

Author:

Related Articles

Back to top button