I read a lot of papers. Whitepapers at lunch, preprints at night, the occasional supplemental methods section when I'm feeling reckless. After a few years of this I've developed a rule: a paper is a lead, never a verdict. The Global Investigative Journalism Network just published a tipsheet on verifying scientific claims aimed at reporters, but it's genuinely useful for anyone who touches science writing, which in 2026 is basically everyone with a feed.

The framing is refreshing. GIJN doesn't say papers are untrustworthy. It says a published paper is the start of an investigation, not the end of one. Which is exactly how I treat a GitHub PR from a stranger: interesting, possibly brilliant, possibly broken, and definitely worth a review pass before you merge it into your worldview.

Here's the distilled version, with my own developer-brain annotations baked in.

1. Read the actual paper. Not the abstract, not the press release, the paper. Press releases are marketing. Abstracts are trailers. The methods section is where the truth lives and where most people stop reading first.

2. Check the journal, and whether it's peer reviewed. Predatory journals exist and they'll publish almost anything for a fee. A journal with a real editorial board, a real impact factor history and a track record of retractions (yes, retractions, they're a sign the process works) is a better bet.

3. Look at the sample size and study design. A headline that says "compound X cures Y" built on 12 mice is not a headline. It's a hypothesis. Cell studies, animal studies and human trials are different rungs on a very tall ladder, and conflating them is the single most common science-reporting error.

4. Check who funded it. Funding doesn't automatically invalidate results, but it changes your prior. Industry-funded trials publishing favorable industry outcomes deserve extra scrutiny, the same way you'd squint at a benchmark run by the chip vendor itself.

5. Look for conflicts of interest beyond funding. Does the author hold a patent tied to the result? Do they sit on the company's scientific advisory board? Disclosures are usually buried at the end, in small type, where hope goes to hide.

6. Compare against the existing literature. One paper contradicting 50 years of work is either a revolution or an outlier, and the base rates heavily favor outlier. A finding that hasn't been replicated is a claim, not a fact.

7. Talk to independent experts. Not the authors. Scientists in the same field who can tell you whether the methods are sound and whether the conclusions overreach. This is the step that separates real journalism from aggregation, and it works for blog posts too.

8. Be suspicious of press-release science. If the claims in the university's press release are much bigger than the claims in the paper, someone in a communications office got creative. That gap is the story.

9. Check preprint status. A lot of what flies around as "new study finds" hasn't been peer reviewed at all. Preprints are valuable, they accelerated vaccine research enormously, but they come unvetted and should be labeled as such.

10. Watch for statistical red flags. P-hacking, cherry-picked endpoints, switching outcome measures mid-study. If a paper measured 40 things and highlights the two that came out significant, that's not a breakthrough, that's a slot machine that paid out.

What I like about the GIJN sheet is that it treats skepticism as workflow rather than cynicism. Every tip is a checkable step, not a vibe. That's exactly how I think about debugging: you don't trust the code or distrust it, you write tests and find out.

I've leaned on versions of these checks myself. When I covered the viral genome generation story last week, the difference between "AI designed a working genome" and "lab confirmed specific functions in a partial construct" was enormous, and you only see that gap inside the methods. Same with the fusion plasma record out of Oxfordshire: the qualifiers matter, and the qualifiers never make the headline.

The uncomfortable truth is that verification is slow and sharing is instant. Every incentive pushes you to skip the slow part. Don't. The next time a study blows up your feed, ask one question before you hit share: who checked it, and against what?