Trust — for Content and Brands

Yesterday, two interesting things happened — the head of medRxiv emailed me, and Elsevier shared news about its new ScienceDirect AI launch.

Well, ScienceDirect AI really should be called “ScienceDirect AI opens in new tab/window” because that’s how it’s named in the press release, which itself may have been written by an AI that I wouldn’t trust given this section alone:

Not only can you search for articles, you can “chat with a document” — which, no offense, sounds like a really unfulfilling social interaction. Also, this phrasing seems like they are one step away from claiming you can “flirt with a document,” which I will also skip.

I don’t trust AI to summarize research reports in any meaningful way. First, many are flawed, and I am imagining stacking them vertically with their flaws as holes in the pages, and I think the flaws would not line up — that is, each has it’s own unique limitations, data errors, interpretation errors, and biases. How can AI perceive what is solid, elided, or flawed? I don’t think it can.

Also, I’ve seen too many human summaries that miss key facts and misinterpret findings. I always find things reading an actual paper that give me pause or stimulate thoughts.

Why would I outsource my brain to some LLM using a tired inference model?

There is also the “trusted content” proposition, which given the current state of scientific information I’d recast as “preliminary claims, incremental findings, and some really bad stuff we won’t be able to retract for years, if ever.”

I am a big believer in the importance of good claims reporting via independent editorial scrutiny and well-organized peer-review. But let’s remain a little humble about what this is — reporting within specific communities doesn’t necessarily work across communities, especially if an inference engine with no real-world experience is at the heart of it.

“Opens in new tab/window” might describe how these communities should be thought about — separate, distinct, specialized, and worth preserving, especially given the function of claims publication generally, which is to move things forward, not draw stark conclusions.

The medRxiv Brand

On the other end of the spectrum, we have preprint servers — in this case, medRxiv. My post two weeks ago about how a preprint on medRxiv failed to include important commercial disclosures prior to media coverage in the New York Times despite the head of medRxiv being a co-author on the paper seems to have reached said person, Harlan Krumholz.

I received an email yesterday morning with snippets from the post and a request to explain myself.