Half the findings don't survive a rerun. That's the system working.
July 5, 2026 · Dana Cole, How to Think Straight About Psychology~5 min read
You know the feeling where you read a headline — "study finds that standing in a power pose for two minutes changes your hormones", say, or something about willpower being a finite fuel — and you quietly rearrange a corner of your life around it? I've done it more than once, which is a bit embarrassing for someone who used to teach this for a living. So I want to tell you what a very large study just found, not to make you cynical, but to hand you something more useful: a way of reading the next one.
What the study actually found
This spring, a team led by Brian Nosek — he runs the Center for Open Science — reported the results of the biggest, most systematic replication effort anyone has attempted. Independent groups took a large batch of published findings across psychology, economics and neighbouring fields, and simply reran the experiments to see what came back. The headline, reported by UVA Today and Forbes in April, is that only about half of them replicated. Roughly one in two. The other half either shrank, wobbled, or didn't show up at all the second time round.
Now, before your stomach drops, let me be careful about what that number is and isn't. "About half" is the honest, approximate shape of it — not a precise decimal I'd want you to quote at a dinner party, and itself an estimate from one (very large) effort. And here's the part people skip: a finding that "failed to replicate" is not the same as a finding that's false. It means the effect wasn't confirmed on rerun — it came back weaker, or absent, or only under conditions nobody had pinned down. Not-yet-confirmed, not debunked. The difference matters, and losing it is how you get from a sober result to the lazy conclusion that psychology is all made up.
Why this isn't the scandal it sounds like
Here's the reframe I'd actually stand behind. When a field builds machinery whose whole job is to catch its own mistakes — and then that machinery catches a great many of them, in public, on the record — that is not the field falling apart. That's the field doing exactly the thing that makes it a science in the first place.
Keith Stanovich has been making this point for decades in How to Think Straight About Psychology, and it lands harder than ever this week. Science, he keeps insisting, isn't a gallery of finished truths handed down by geniuses. It's cumulative and probabilistic — a slow, self-correcting grind where knowledge is always provisional, always up for revision. The whole distinction between real science and the pseudoscience that mimics it, he argues, isn't the subject matter. It's the method: public, testable, and willing to be wrong out loud. A replication study is that willingness made concrete. The awkward number isn't a betrayal of the science. It's the receipt.
The smell test you already half-knew
So what do you do with this, standing in the supermarket queue reading a psychology headline on your phone? Stanovich gives you a little hierarchy of evidence, and it's worth carrying around. At the bottom, weakest of all, sits the testimonial and the anecdote — the "I tried it and it changed my life" that self-help runs on. Above that, a single correlational study: X and Y turn up together, which tells you less than it sounds like (correlation isn't causation, and the arrow could point either way). Above that, a controlled experiment. And at the top, the thing you should actually trust: converging evidence — many separate studies, different teams, different methods, all quietly arriving at the same place.
The replication result is basically that hierarchy, proven the hard way. Any lone splashy finding — the one that's new, surprising, perfectly designed for a headline — is precisely the one most likely to be in the half that won't hold up. It's a single, weak signal wearing the costume of a strong one. Whereas a boring, well-worn conclusion that a dozen labs have circled back to from different directions is the one that's earned your weight. The chorus, not the solo.
A 2026 replication effort led by Brian Nosek at the Center for Open Science found that only about half of published social-science findings held up when independent teams reran them. "About half" is itself one estimate, and failing to replicate is not the same as being false. Framework: Keith Stanovich, How to Think Straight About Psychology (converging evidence beats any single study). Popular-science interpretation, not professional advice.
What this means the next time you read a headline
So here's the small, un-heroic toolkit I'd actually trust you to use. Four questions, none of them clever, all of them cheap. First: is this one study, or many? A lone finding — however elegant — is a bet, not a fact; give it the weight of an interesting maybe. Second: what would prove this wrong? That's Stanovich's oldest rule, borrowed from Popper — a claim that can explain any result whatsoever, that no possible observation could dent, isn't telling you anything (that's the difference between a testable finding and a horoscope). Third: expect the effect to shrink. Splashy first results tend to be the high-water mark; the truth, when it settles, is usually more modest. And fourth, the one I'd tattoo somewhere if I were the tattooing sort: distrust the single viral result most of all — the more perfectly it fits the headline, the more it's earned your suspicion, not your share button.
None of this asks you to sneer at psychology, and I'd be sad if that's what you took away. The half that replicated is real, hard-won knowledge, and the machinery that flagged the other half is the reason you can believe the first half at all. What it asks is gentler than cynicism: hold any single finding loosely, and save your confidence for the day the studies start agreeing with each other. That's not a lower standard. It's the only one that's ever worked.
A finding that fails to replicate isn't a lie. It's a promise the field kept — to check its own work in public.
The half that survives is what you get to keep.
Framework from Keith Stanovich's How to Think Straight About Psychology (基思·斯坦诺维奇《对伪心理学说不》) — falsifiability, the hierarchy of evidence (testimonial < single correlational study < controlled experiment < converging evidence), and science as cumulative, probabilistic and self-correcting rather than settled by any lone study. News peg: a 2026 replication effort led by Brian Nosek of the Center for Open Science, reported by UVA Today and Forbes in April 2026, finding that roughly half of published social-science results replicate. "About half" is one estimate; "failed to replicate" means not-yet-confirmed, not false. Popular-science interpretation, not professional psychological advice.
那你排队买菜、低头刷到一条心理学标题时,能拿它怎么办?斯坦诺维奇给了你一个小小的证据等级,值得随身带着。最底下、最弱的,是个人见证和轶事——「我试了,它改变了我的人生」那种,自助读物全靠它撑着。往上一格,是单个的相关研究:X 和 Y 一起出现了,可这比听起来能说明的少得多(相关不等于因果,而且箭头往哪边指都还两说)。再往上,是控制实验。最顶上、真正值得你信的那样东西,叫聚合证据——许多各自独立的研究,不同团队、不同方法,都悄悄落到了同一个地方。
2026 年,开放科学中心(Center for Open Science)的 Brian Nosek 主导的一项大规模复现研究发现:已发表的社会科学结论中,只有大约一半在独立团队重做时仍站得住。「大约一半」本身也只是一个估计,而「未能复现」并不等于「是假的」。框架:基思·斯坦诺维奇《对伪心理学说不》(聚合证据胜过任何单一研究)。本文为科普解读,非专业建议。
じゃあ、レジに並んで、スマホで心理学の見出しを読むとき、どうすればいい? スタノヴィッチが、小さな証拠のランクをくれる。持ち歩く価値がある。いちばん下、いちばん弱いのが、体験談と逸話——「やってみたら人生が変わった」というやつで、自己啓発本はこれで持っている。その上が、一つの相関研究。X と Y が一緒に出てきた、という話で、聞こえるほどには何も言っていない(相関は因果じゃないし、矢印がどっち向きかもまだわからない)。その上が、統制実験。そしていちばん上、あなたが本当に信じていいのが、収束する証拠——別々の多くの研究が、別々のチーム、別々の方法で、そっと同じ場所にたどり着くこと。
2026年、オープンサイエンスセンター(Center for Open Science)のブライアン・ノセック率いる大規模な再現研究は、発表済みの社会科学の知見のうち、独立したチームが追試して持ちこたえたのは約半分だけだと示した。「約半分」という数字自体も一つの推定であり、「再現しなかった」ことは「偽である」ことと同じではない。枠組み:キース・スタノヴィッチ『心理学をまじめに考える方法』(収束する証拠は単一の研究に勝る)。科学解説であり、専門的な助言ではない。