Michael Weissman on Lab Leak and 'Science'
Source: Michael Weissman on Lab Leak and ‘Science’
Publisher: Econ Journal Watch | Author: James Robbins
Published: March 28, 2026 | Archived: July 17, 2026
‹ Previous episode | Next episode ›
Posted 29 Mar 2026
An influential article by Jonathan Pekar and 28 other authors published in Science in 2022 claimed that Bayesian analysis of the molecular phylogeny of early SARS-CoV-2 cases indicated that the likelihood that two successful introductions to humans had occurred was greater than the likelihood that just one had occurred. Michael Weissman explains his EJW article, which discusses a fundamental error hiding in plain sight and initially pointed out by Angus McCowan. Weissman uses a simple analogy to explain the error. Correcting the error using the data, model, and simulations of the original paper reverses the implication of the analysis-the single-introduction likelihood becomes greater than the two-introductions likelihood. That undermines the article’s supposed support for natural origin. Weissman is interviewed by James M. Robins of Harvard University. Weissman and Robins discuss the editorial practices at Science, which, they suggest, ought to retract the article.
‹ Previous episode | Next episode ›
Transcript
0:00 James Robbins: So this is Jamie Robbins interviewing Michael Weissman about his recent paper on errors in an important paper concerning the early part of the pandemic. Michael, will you describe what the paper was that you were criticizing in this article?
0:19 Micheal Weissman: Yes, they argued this was a paper in Science magazine in 2022. There was one of a pair of papers that established or were alleged to establish that Covid spilled over into human beings at the Huanan Seafood Market in Wuhan. This particular paper claimed that looking at the early sequences, sequences of the early cases showed that there were two separate lineages that almost certainly spilled over separately into human beings. That is, they didn’t. It wasn’t the one spilled over into human beings and then branched out with the mutations later.
1:10 Micheal Weissman: It was that it started in another animal, done some mutating, and then spilled over twice, and that these two spillovers gave the two successfully propagating lineages. So that’s. That was what the paper claimed.
1:29 James Robbins: And these lineages were named A and B, or at least in the paper. Yeah, And I noticed that you said that all the cases from the market were B.
1:44 Micheal Weissman: That’s correct.
1:45 James Robbins: So how did A get into this?
1:48 Micheal Weissman: Well, the cases from the market, the 12 cases that were sequenced, were all B. But as you looked around the rest of Wuhan, or even popping up in Washington State or various other parts of China, there were nearly as many A cases, A cases as well. And these, these two lineages were very close to each other. They only differed by two nucleotides, which even can happen occasionally in one transmission.
2:20 James Robbins: And what is it about the paper that, as I understand it, was seen as evidence for zoonotic origin rather than lab leak? How would they claim to establish have.
2:38 Micheal Weissman: That’s a little bit obscure and complicated and I think maybe somewhat overhyped. So maybe we can come back to that because I think it’s first important to establish what are the. There is a well defined question, was there one successful spillover or two separate successful spillovers? And that actually has a relatively simple answer that I can describe. And then we can go back over.
3:12 Micheal Weissman: What does that mean? Because that gets more complicated.
3:16 James Robbins: Okay, so what is the key criticism? You have their argument that there were.
3:23 Micheal Weissman: Okay, so they claim to do a Bayesian analysis of what sort of patterns would you get in the sequences if you had just one spillover versus if you had two spillovers? And they compared some features of the sequence collection and the likelihood under those two scenarios, the probability you would get those if there was one spillover or if there were two, they initially claimed that the likelihood ratio was 60:1, favoring tooth spillover. But, you know, very close to definite answer. Then at that time, pseudonymous, now we know his name is Angus McCowan, a patent expert went over the code and found multiple big coding errors. And just fixing the coding errors, which they did in response to his insistence, reduced the odds to four or so to one, still favoring two spillovers over one.
4:33 Micheal Weissman: But Angus McCowan had pointed out almost exactly three years ago that there was a fundamental logic error in their approach and that if you correct the fundamental logic error, the odds actually favor that there was just one spillover. And here’s the fundamental logic error. When you’re doing these conditional probabilities, you pick some features.
4:59 James Robbins: Let me interrupt. What do you probabilities have to do with anything?
5:03 Micheal Weissman: Well, they’re using a Bayesian argument. So you look at whether various observed pieces of data, what’s the conditional probability given one hypothesis or given the other hypothesis, that you would see those data. So if the conditional probability says, yeah, you’re pretty likely to see those data for one hypothesis, and you’re unlikely to see those data for the other hypothesis, then those particular data favor the hypothesis with the high conditional, which gives a high conditional probability. So that’s the standard Bayesian thing. It’s as you know, a completely conventional logical method of approaching questions like this, comparing hypotheses.
5:50 Micheal Weissman: But here’s the thing. They took two of the features that they’d observed in the real data. They asked what the conditional probability was of seeing those. If there was just one spillover and it wasn’t very high, and they just didn’t ask for two spillovers. In effect, they assumed that the conditional probability was 1, that it was certain you would get those features for two spillovers, which it absolutely wasn’t.
6:18 Micheal Weissman: And you know, it’s just a basic error in the fundamental application of a well defined logical system. It’s like a sort of 2/2 equals minus 4 type error.
6:33 James Robbins: You had a very good example.
6:35 Micheal Weissman: Yes, I’ll give the.
6:37 James Robbins: For a layperson of what their error was.
6:41 Micheal Weissman: Yes.
6:42 James Robbins: Without requiring any knowledge of virology or.
6:45 Micheal Weissman: Of COVID Because this is fundamentally a logic problem. The virology, which might be interesting, I didn’t deal with and it’s not necessary to understand the logic. Here’s, here’s my example. Let’s say there’s been a burglary and you have two suspects and you’ve, you saw a car leaving the burglary, so you’re pretty sure that that was the burglar. So the car happened to be a blue Toyota for one suspect.
7:14 Micheal Weissman: Say suspect one, you say, I don’t know much about them, but what’s the probability they would drive a blue Toyota? Well, it’s probably not very high. You don’t know much about them and most people don’t drive blue Toyotas. For the other suspect, you say, what’s the probability would drive a car? Well, that’s pretty high. Most people drive cars. But of course you would. So that would just totally put the thumb on the scales for saying that it was the. The suspect that you only required that they drive a car, not the suspect that you said it’s not very likely that they drive a Go Toyota. So for here, for the single spill hypothesis, they required the analog of blue and Toyota.
8:03 Micheal Weissman: And for the double spill hypothesis, which is the one they wanted to favor, they required the analog of car.
8:11 James Robbins: That does sound like a pretty serious error. How did you or more or Angus McCowan happen on this error and why before that? Why had it been missed if it was so fundamental by the referees and so forth?
8:32 Micheal Weissman: I can answer the first question better. The second one, Angus had the patience to go through the whole thing and it turned out it wasn’t it. This particular one, unlike all the big coding errors, wasn’t hidden in the code. It was right there, written out in the supplement. The what? What they required. And you could read that they just didn’t require these two features, that of their favorite hypothesis that they required of the unfavored one. Who knows if I even read the supplement that carefully. I didn’t catch it until Angus pointed it out. Partly because you just don’t believe that there’s going to be an error like that.
9:15 Micheal Weissman: You just don’t think anybody could do that. It’s so blatantly wrong that, you know, you’re looking for things that are more subtle that may say something about why the other people didn’t miss it. But what it doesn’t explain is why after he pointed this out, it’s almost exactly three years ago now he pointed it out on Pub Here he wrote some of the lead authors and the undoubtedly the editors are aware of it. Why nobody has done the obvious thing, which is if a paper contains an error and you know how to correct the error, and correcting the error reverses the conclusion. So when you do a minimal correction of this error, it ends up favoring a single spillover rather than a double spillover.
10:04 Micheal Weissman: The ordinary thing to do would be to retract the paper the core qualitative result of the paper has the wrong sign. It’s upside down from what their own analysis, using their own model and their own data would give. If you just take out the math error and nothing has happened, although they fixed the coding errors. No response on this.
10:28 James Robbins: Now, if you say the honest analysis that used the correct Bayesian logic reversed the evidence from their, what they claim to two spillovers to evidence for one spillover, why should they retract? I thought science was supposed to be neutral and the paper was just as good. It was just evidence for one spillover.
10:54 Micheal Weissman: Oh well, you could do that, but once things are messed up enough, maybe it’s simpler to retract. Yeah, good question. One could correct the paper, but then it would involve a bunch of work because there’s other things other than the central qualitative conclusion. So anyway, that’s their problem. But. But leaving it up in the provably incorrect form is everybody’s problem.
11:23 James Robbins: Before we get further into that, do you want to describe the specific.
11:29 Micheal Weissman: Yes.
11:30 James Robbins: Pieces of evidence that they left out of one likelihood, but included in the other?
11:36 Micheal Weissman: Just for completion, I should say that these two lineages were very similar to each other. They differed by two nucleotides, that is two of these little spots in the genetic code out of 30,000 spots. So really similar. It’s like a couple weeks worth of accidental mutation though could happen in a single transmission. So, so they required that the single spill hypothesis have a two nucleotide difference between the, between the sequences of the lineages.
12:11 Micheal Weissman: And they didn’t put any requirement on the sequence difference for the two spill hypothesis. The other thing was that we mentioned the number of cases in each of these two lineages was about the same, roughly the same, although in the market it was all lineage B. And so they required of the single spill hypothesis that it give the two lineages have about the same number of cases. And in the two spill hypothesis they just put no requirement on that at all. Zero.
12:47 Micheal Weissman: No, just. And what’s particularly interesting is you can just look at their data on single spell hypotheses, outcomes simulations. They do this by simulation and see that different single spills give a huge range of different sizes. And so you say, well, if you’ve got two of them, the chance that those two are about the same size is low, tactically, comes out under 20%. And they just left that factor out.
13:19 Micheal Weissman: So they left out these two factors and each one is, you know, more or less one chance in five, something like that. And that’s enough to swing the odds from favoring two spills to favoring one spill.
13:34 James Robbins: As I understand it, even in the. Is it in the one spill case, did they not even take it into account at all? And in the two spill, they took essentially,.
13:47 Micheal Weissman: In the one spill, they required both of these and they saw how often they happen in their simulations. In the two spill, they just dropped the requirements, didn’t pay any attention.
14:01 James Robbins: I see.
14:02 Micheal Weissman: And again, that’s equivalent to assigning a probability of 1 instead of seeing what the actual probability would be.
14:08 James Robbins: Okay.
14:08 Micheal Weissman: So my contribution that went beyond what Angus had done. Angus had redone all the simulations with fixing the code. And it was very complicated and people were having trouble understanding it. And I showed you could really simplify the argument and do it only using the content of what the original authors had prepared, only using their simulations, not doing anything new.
14:34 James Robbins: And do they simulate both one spill and two spills?
14:40 Micheal Weissman: Two. No, no. They only simulated one spill. And then they just sort of made up a story about two spill. This was, I know from talking to people in the field that this was. Everybody noticed this right away, like, what’s, what’s your model? The two spills. And nobody got a straight answer. And so took somebody with patience, namely Angus McCown, really going through and realizing what they’d done, instead of having a model of the two spill hypothesis.
15:11 James Robbins: I see. And your calculation of the probabilities from the two spills, if they’d only done a simulation for one spill, can you explain how?
15:21 Micheal Weissman: Yes, because at a high level you.
15:23 James Robbins: Were able to calculate.
15:24 Micheal Weissman: Yes, they’re very explicit that their two spill hypothesis is of two independent single spills. So since they have data on what comes out of the simulations for single spills, you can easily compare the sizes of any random pairs of their single spill hypothesis. And they have a broad range of sizes. And so you can say, well, how often are they as close as they demanded they be. And it’s not very often for, I.
16:01 James Robbins: Mean, as they demanded to be for.
16:02 Micheal Weissman: The single spill case, correct for the sequence difference of two nucleotides. There’s also a sort of simple argument. If you go back to the most recent common ancestor before the spill, there’s going to be some evolutionary time before the two spills. And it’s, you know, the mutations occur at random. That’s according to both common sense and their own model. And if the most recent common ancestor was just the right length of time before the spills, you get a certain probability that it could be 2 nucleotide difference. Of course, it could also be 0. It could be 1, could be 3. There’s a what? A Poisson distribution of possible values.
16:53 Micheal Weissman: Even if you tune the time for those mutations to accumulate to be just the best to make two, as likely as you can get it, you still only get a 27% chance. And if you do a more realistic standard thing where you say, oh, it’s about long enough for there to be two, you get more like a 20% chance.
17:13 James Robbins: And how did they analyze that differently in the one spill versus the two spill case?
17:19 Micheal Weissman: That bias for two spill? They just didn’t say. They in effect assumed that it was always exactly two mutations different and there’s just no model for one spill. They look at what comes out of their simulations and sometimes, sometimes there just aren’t these two separate lineages. Sometimes, you know, it’s fairly rare that you get two big lineages that are that close to each other, or rather even that far from each other in sequence, even though it’s not very far, because if you let something run for a while, it gets a head start on the other one.
17:58 Micheal Weissman: And usually if they’re about the same size, they have to have to happen fairly soon. So it’s not super common, but it turns out to be even less common in their two spill model.
18:11 James Robbins: So what’s the lesson you took away from this? And what do you want the reader to take away?
18:18 Micheal Weissman: Okay, well, now here’s where it gets complicated because there’s multiple lessons. One of them is just, you know, the particular technical lesson. The argument that somehow two spills favored the Huanan Seafood Market story to the extent that had any weight that it should has to go the other way because it probably wasn’t two spills. It looks like just using their own argument, it was more likely one spill. So it’s another blow to the Huanan Seafood market story.
18:53 Micheal Weissman: And part of that’s because the one spill model tends to give earlier initial spills because those extra mutations happened in people, not before. And that makes the pandemic start even earlier before cases start up in the market. And also the lineage B that was found in the market in all the market cases that were sequenced is less like the, well, I don’t know if they’re ancestral, the related natural sequences. So that makes it look like the market cases descended from a spillover of something that was more ancestral looking.
19:34 James Robbins: I see.
19:36 Micheal Weissman: So that’s the most technical.
19:38 James Robbins: And what, in terms of the alternative hypothesis of a lab week, what are the implications for that?
19:48 Micheal Weissman: Well, a lab week is quite Consistent with a single spill. After one spill, people might wise up and try and avoid a second one, though it’s also consistent with a couple spills like the Marburg leak was multiple spills. The Sverdlov anthrax leak was multiple strains. So for lab leak versus zoonosis, the implications would move a little bit more toward lab leak. But it’s maybe not such a huge deal for statistics education.
20:29 Micheal Weissman: I think there’s a big implication which is people learn to use these extremely fancy, complex programs. But, you know, it’s more important just, you know, to learn to distinguish your ass from a hole in the ground and be able to do the basics right. And I think that there should be way more emphasis on, you know, getting the basics right and not on fancy programs for big data for editors, reviewers. Well, you know, the implications are sort of obvious. This was in Science magazine, which is a American association for the Advancement of Science publication.
21:07 Micheal Weissman: A few years ago they dropped the category of technical comments, where if somebody just screwed up a paper, you could send in a technical comment and show that it was wrong. They’ve got to restore that. I mean, how else do you fix errors like this?
21:23 James Robbins: How could an important journal like Science do that? I thought the basis of Science was that mistakes are corrected through the scientific literature and process. But what possible reason could Science Magazine have had for making that decision?
21:42 Micheal Weissman: I can only speculate. I do know that they made that decision before, just before COVID Hit. So that’s not the explanation of it. It was horrible. I got one of the last technical comments into a AAAS publication. But I was warned at the time, you’re lucky we’re dropping this. And they had no explanation. And I had no idea how science works without that process. Otherwise, you know, you build error on error and it’s completely unstable structure. One possible thing would be more journals like Econ Journal, watch that take up the slack and publish critiques of other journals when those journals don’t fulfill their obligations to publish the corrections.
22:29 James Robbins: So you, you think the journals would rather that than be seen as serious journals that self. Correct.
22:37 Micheal Weissman: I don’t know what they would rather do but. But right now they’re getting away with it. And so that’s question if there, if the incentives were changed because there were serious journals, several serious journals, maybe specializing in physical sciences and specializing in biological sciences and so on, devoted to catching serious errors in other journal maybe would change the incentive structure. There’s something special of course for anything Covid related though this can’t account for why AAAS changed its policy. And that’s that there’s this enormous embarrassment for the whole scientific community that probably in some sense collectively were responsible for Covid that it was not individually, but probably it was a spillover from scientific activity.
23:31 Micheal Weissman: And that makes big motivations in addition to the usual ones to just not want to revisit these arguments.
23:41 James Robbins: And what, what positive evidence is there for a possible spillover? And did the people who worked in the laboratory where they worked with these kind of viruses, have they made a statement about this?
23:58 Micheal Weissman: They. No, the people who work in the library, in the laboratory have said they didn’t do it. They did. When asked were you working on stuff like this? Der Spiegel asked Xi Zhengli. She said, I don’t want to answer that question. So most trivial positive piece of evidence is again a Bayesian piece of evidence. If there were a natural spillover, it could happen anywhere in Southeast Asia, especially where the natural relatives live around southern Yunnan, Laos, northern Vietnam.
24:35 James Robbins: By relatives you mean the bats carrying these viruses?
24:40 Micheal Weissman: Yeah, viruses that are carry that are of this family carried by bats. And they’re well located. This actually is not controversial. But even if you ignore where they’re from, it could happen anywhere in China or Southeast Asia. So that’s a big area, well over a billion people. Wuhan is a big city, but it’s only got 10 million. So it’s less than 1% of the potential area. So the conditional probability that if there was a lab spillover it would be in Wuhan is very high because that’s where we know people were working on viruses, you know, in this family and playing around with modifying the genomes. So if you say where would it happen? You’d say, oh, probably Wuhan for lab leak.
25:31 Micheal Weissman: If you said what’s the chance? If it was some sort of natural thing would happen in Wuhan, it’s less than 1%. So using the basic Bayesian reasoning already shifts whatever your prior guesses were by a factor of 100 toward Lab League. And then there’s, there’s other features. But that is a long story on which I’ve written a long non peer reviewed blog.
25:56 James Robbins: People were interested in reading that.
25:59 Micheal Weissman: Where would they go to find your Weissman substack? Okay, Google Michael Weissman substack. And it’s called an inconvenient probability. It’s the most recent version of my and inconvenient probability.
26:13 James Robbins: And papers I you wrote had something like 28 or 29 authors.
26:19 Micheal Weissman: This one had, the one I’m critiqued here had 29 authors. Yes. 29 Authors, so. And it’s been cited over 300 times. He was cited in the New York Times before it even came out, and various other foreign affairs, so on. So it had a. A big impact. And it’s just wrong. And these places are not acknowledging that it was just plain wrong.
Write a comment