Scaled up, I expect defaults to reduce meat consumption by ~1-2 pp.

But increased convenience can lay the groundwork for more transformational changes.

A lot of folks are excited about plant-based defaults, a choice architecture strategy that sets up a meal so that the easy, no-action option is vegan and/or vegetarian.

• Cass Sunstein, when asked, “What behavior change strategies have you seen be most effective in helping people adapt to eating more plant-based meals?” replies, “Defaults…If you are defaulted into something... chances are it’s going to stick... it is highly suggested that that’s the best nudge.”

• Jennifer Channin says defaults “bypass the false dichotomy of individual versus government intervention when it comes to our food system.”

• Björn writes “I can say confidently that plant-based defaults are the initiative that excites me most. It’s that perfect combination of palatable, scalable, and impactful work that can really move the needle…most interventions to reduce meat consumption are not this consistent, nor this powerful.”

• Katie Cantrell, CEO of Greener by Default, tells Michael Grunwald: “What’s so exciting is the way [a default] retains freedom of choice. Nobody’s forced to go vegan…The idea is to make the sustainable choice the path of least resistance, without getting into the big political and cultural fights.”

I think Katie’s comment nicely illuminates defaults’ appeal. Persuading people to eat less meat is hard and sometimes causes backlash, veganism is rare and perceived to be a social risk, and meat is culturally ascendant. If we can sidestep all that while reducing meat consumption and helping institutions meet their climate and budget targets to boot, that’s a win-win-win.

My view: I expect defaults, when scaled up — meaning, deployed by institutions who aren’t mission-aligned and without careful supervision by advocates and researchers — to reduce meat consumption about as effectively as everything else: by about 1-2 percentage points (henceforth, ‘pp.’). This post explains why I think that.

First, I’ll adduce evidence across domains that scaling up is hard and typically reduces effect sizes. Second, I’ll explain why choice architecture literature in general seems unusually prone to inflated estimates of impact. Third, I’ll argue that plant-based default evaluations to date provide no credible, direct estimates of net effects at scale. Fourth, I’ll make the case that the “automaticity” paradigm — the idea that many domains of human behavior are explained by unconscious or unreflective processes — isn’t a good model for eating meat. Fifth, I’ll sketch out an alternative theory: defaults, by increasing the convenience of eating plant-based (no small thing!), lay the groundwork for sticking to dietary changes, and therefore work nicely in conjunction with arguments about food and self-identity, and with concerted efforts to make vegan food appealing and nutritious.

This post is longer than my usual and more heavily annotated. Emphasis added throughout.

Most stuff doesn’t scale

Consider Peter Rossi’s Iron Law of Evaluation: the “expected value of any net impact assessment of any large scale social program is zero.” His Stainless Steel Law: the “better designed the impact assessment of a social program, the more likely is the resulting estimate of net impact to be zero.”¹ Examples:

• DellaVigna & Linos (2022), studying nudges, find that the average effect size from two representative meta-analyses was about 8.7 pp., but when deployed by Nudge Units in governments, it was 1.4 pp. This difference was largely explained by selective publication and low statistical power in academic papers.

  • Along the same lines, nudges to get incoming college students to apply for more financial aid are generally evaluated in “relatively small-scale studies done in partnership with a local organization serving hundreds or thousands of students” and are viewed as “promising because it seems possible to scale them at low cost” (Bird et al., 2019). However, when these nudges actually are scaled to hundreds of thousands of students, researchers “consistently find no effect [of] these messages on student enrollment or financial aid outcomes. This null finding is consistent across samples, content, timing, visual presentation, and offers of personalized help. Large sample sizes allow us to rule out very small effects of these interventions.”

• Agostinelli et al. (2020), studying the push and pull between parental and peer influence: “interventions (e.g., busing) that move children to a more favorable neighborhood have large effects but lose impact when they are scaled up because [of] parents’ equilibrium responses.”

• Stevenson (2023), studying criminal justice reforms: “most reforms and interventions in the criminal legal space are shown to have little lasting effect when evaluated with gold standard methods… Stabilizing forces push people back toward the path they would have been on absent the intervention. Cascades—small interventions that lead to large and lasting changes—are rare. And causal processes are complex and context dependent, meaning that a success achieved in one setting may not port well to another.”²

• Vivalt (2020), also discussing development: “Government-implemented programs also had smaller effect sizes than academic/NGO-implemented programs, even after controlling for sample size.”

• Bidisha Barooah, a senior evaluation specialist at 3ie: “taking the program to scale meant that the quality of implementation changed, because the institutions that were responsible for taking the program to scale had different capacities than what was there in the pilot phase… The challenge typically comes when the agency that is doing the pilot is different than the agency that is doing the scale-up.”

And so on (Al-Ubaydli, List, & Suskind 2017). I’m afraid this is just how the world works. At scale, equilibrium effects dominate; implementing institutions aren’t motivated to get the details right the way a supervising grad student was; and pilot sites are often chosen deliberately for qualities that make them poor models of the world at large (Alcott, 2015).

The nudge literature in particular is a mess

First, some of its (former) superstars are fraudsters. My belief, to quote Alex Tabarrok, is that the rot is deep and not yet uprooted. Second, and related: a lot of contemporary papers still cite those fraudsters. I asked Andrew Gelman about this, and he speculated that folks still believe that the theories are credible even if the work isn’t. To my eyes, it suggests credulousness and a lack of attention to detail. Third, rosy results are more likely to get surfaced: a critical review (Maier et al., 2022) found “no evidence for the effectiveness of nudges remains” when “publication bias is appropriately corrected for.”

Fourth, publication bias and fraud are intertwined. What happened, I think, was that unscrupulous folks published huge effect sizes, and the fame and glory they got for that — TED talks and the like — set a tone where questioning those results, either by interrogating their provenance or by publishing conflicting findings, was considered a dead-end. As a former grad student (Zoé Ziani) told the New Yorker:

When she expressed her doubts, the adviser snapped at her, “Don’t ever say that!” Members of Ziani’s dissertation committee couldn’t understand why this nobody of a student was being so truculent. In the end, two of them refused to sign off on her degree if she did not remove criticisms of Gino’s paper from her dissertation. One warned Ziani not to second-guess a professor of Gino’s stature in this way. In an e-mail, the adviser wrote, “Academic research is like a conversation at a cocktail party. You are storming in, shouting ‘You suck!’... [Joseph] Simmons told me that he could name countless people who had similar experiences. “Some people are hurt by this stuff and they don’t even know. They think they’re not good enough—‘It must be me’—so they leave the field,” he said. “That’s where I started to get angry.³ How many Zoés are there?””

The plant-based defaults literature doesn’t provide credible estimates of long-term net effects

While preparing a meta-analysis, I read (as far as I know) every plant-based default evaluation published by December 2023.⁴ Our paper didn’t include any default studies, however, because none met our inclusion criteria: at least 25 people in both treatment and control (or at least 10 days/meals in total for cluster-assigned studies), random assignment, and measurements of consumption of meat and/or animal products taken at least a single day after treatment began.

In our sample, the closest thing to a default study looked at randomly repositioning items on a cafeteria billboard menu to put the vegetarian option first or not (Andersson & Nelander, 2021). The authors found that “placing the vegetarian option at the top of the menu on average decreases the share of meat option sold that day by 5.5 percentage points,” but the paper actually counts meat and fish as separate categories, and the “reduction in meat dishes is in turn spread out over the vegetarian and fish options available, increasing each of these albeit not significantly so.” The net displacement to vegetarian meals is — you guessed it — about 1 pp.

Many default studies didn’t qualify for our paper because they measured meal choices right away (Boronowsky et al. 2022; Friis et al. 2017; Hansen et al. 2021; Radnitz et al. 2018, 2023; Zhang et al. 2022, 2024). Typically these studies look at reducing meat consumption at catered events, and measure consumption at those events only. If there were regression to the meat at a subsequent meal, we wouldn’t know. Others were excluded because they featured hypothetical outcomes (Betz et al. 2022; Campbell-Arvai et al. 2012; Erhard et al. 2023; Nijeboer 2023; Rutenfrans 2023); didn’t assign treatment (fully) randomly (Saulais et al. 2019; Vandenbroele et al. 2018⁵); or were underpowered (Hohle 2014, Taufik et al. 2022, Gravert & Kurz 2019).⁶

I think our exclusion of hypothetical and immediate outcomes is pretty intuitive, but our hurdles about random assignment and statistical power could use some explanation. In both cases, we’re thinking about underlying differences between treatment and control groups that are unrelated to treatment. Consider Ardesch et al. (2025), who studied plant-based defaults in a university dining hall. It took place in a “field setting, with measurements of real behaviour,” and “data was collected at three time points across a period of 4 months.” However, treatment was not random. Instead, baseline meat consumption was measured for 10 days four weeks before the default, 10 days during the default, and 10 days at follow-up eight weeks later. The study found defaults to be “highly effective at changing food choices, with more than twice the number of vegetarian items sold relative to baseline.”

For this to be an unbiased causal estimate, the reader must assume that nothing else changed over the semester that also influenced meat consumption, or that any such changes were trivial. I doubt this. I remember college as a time when diets were in flux. A big reason why is that many students are managing their own portions for the first time and getting used to a new institution’s cooking, so they tend to gain weight. Perhaps most of that surplus eating takes place at the beginning of the semester, which leads to a big initial surge in eating, inclusive of meat, which later declines. That’s just one story; even if the authors can rule it out, we could tell others. The fundamental limitation of non-randomized designs is that it is impossible in principle to rule them all out. Moreover, the truth/existence of an alternative theory is not contingent on my ability to tell a story about it. Perhaps there is no clear alternative story, but many small factors that cumulatively add up to a very biased estimate.⁷ We don’t know.

Likewise, if you only have two clusters in treatment and two in control, even if treatment was assigned randomly, I don’t take for granted that all systematic differences actually do balance out, even if they do in theory.⁸

About Ginn & Sparkman (2024)

Ginn & Sparkman (2024) is a prominent counterexample. That paper randomly assigned plant-based defaults at three different universities over the course of a semester. On average, successfully implemented plant-based defaults “raised plant-based dining by 58.3%;” they lowered meat consumption by an estimated 21.4–57.2%” at affected stations; and they “reduced total servings taken by 26.1% at the station[s] intervened on.” This study is, to date, our best available evidence on the dietary impacts of plant-based defaults. However, for two reasons, I’m hesitant to treat its results as estimates of the effects of scaled-up defaults.

1. They measure treatment on the treated, but at one of three schools, staff undermined the implementation and didn’t deliver treatment. (Apparently they told students that meat actually was available, and in some cases steered them towards it.) When you scale things up, you can expect exactly this kind of implementation snafu. The “net effect” is inclusive of places where things don’t go right.⁹

2. There’s no measure of net meat consumption at the level of the dining hall. We know what happened at particular stations, but not what effect that had on meat consumption at large. I’ve discussed the “leakage problem” elsewhere — plant-based default days served fewer meals, suggesting patrons simply ate elsewhere to avoid the intervention — and the authors deal with this as well as the data allow for: “[n]otably, even if we assume 100% of these patrons selected a meat dish elsewhere—a worst-case scenario for the intervention—we still would have an estimated 21.4% reduction in meat dishes sold as a result of the intervention.” But notice that this provides a lower bound for the number of meat dishes served rather than the amount of meat eaten. It’s a different estimand.

Meat is not automatic

The two nudges I remember best from the eponymous book are putting images of flies in urinals, which Thaler once called this his favorite nudge, and making organ donation upon death the default. The first targets a low-stakes interaction but plausibly produces large benefits. The second has high stakes but apparently low salience. Per Beraldo & Karpus (2021), “most people do not express a decision—either in favour or against donation—regardless of the system in place... In general, most people in Europe do not register a decision… do not hold a donor registration card…and do not discuss the issue of organ donation with their family.”¹⁰ In both cases, I buy the automaticity assumption: that this is a sphere of human behavior where people are using what Kahneman (2011) calls “System 1” thinking, employing the “brain’s fast, automatic, intuitive approach.”

I do not think this is a good model for eating meat. “Since Eve ate apples,” Byron said, “much depends on dinner.” Sure, a change in spoon size can plausibly reduce the total portion of meat served (Voşki et al., 2025), and placing a hole in a serving spoon for fish sauce might well reduce sodium intake (Kanchanachitra et al., 2020). People adjust portion sizes all the time. But the idea that you’re going to take meat away from someone’s plate, and they’re not going to mind/notice/compensate later, flies in the face of a lifetime of eating with and talking to people. Someone close to me recently said that she just doesn’t feel good after eating unless there’s animal protein there. Another person I know very well once said that food was how he knew his mother loved him, and she, a Jewish woman born in 1924, of course cooked a lot of meat. Meat means a lot to people. I’m sure you can provide your own examples.

Advocates of defaults say they aren’t taking away meat, just steering people so that it takes a little work to get it. But my read of the plant-based default literature is that they are indeed willing to put that work in. When they see a no-meat area, they put in the work to go somewhere else. Not everyone, but enough people that I expect net effects of defaults on meat consumption to be around 1-2 pp.

I’m also dismayed that the canonical cite for food’s being automatic still seems to be a Wansink paper, e.g. “The default option requires less effort and is effective due to consumers’ passive decision-making and/or cognitive or attentional limitations, which is particularly the case for daily food choices (Wansink & Sobal, 2007)” and “nudges are particularly well suited to the area of appetite, where food choices are often automatic, habitual and emotionally driven (Meier et al., 2022; Wansink & Sobal, 2007).” These papers were published in December 2025. If there were a there there, I would have expected a more credible foundational text to have emerged. And if eating meat were governed by unconscious processes, why do meat-free days prompt such vehement backlash?

Defaults lay the groundwork for transformational change

But defaults do not have to be on their own. We can layer other things in. Vasile Stănescu, Ph.D. points to an article interviewing former vegans about their reasons for reverting:

One woman told me she stopped getting invited to her book club because she’d politely declined their potlucks too many times. Another guy said his college friends just stopped texting him about hangouts. Here’s what hit me: these weren’t people being difficult about their veganism. They were just existing as vegans in spaces that weren’t built for them.

The author frames this as a story about isolation. But another read would focus on convenience, a perennially underrated social explanation. Accommodating vegans was inconvenient to their former friends, and they prioritized that over inclusion.

Plant-based defaults make vegan food convenient. After something changes your ideas about food (say, watching Dominion), and especially your self-conception as an eater, convenience either facilitates an identity shift or gets in the way. Per Warren Belasco’s 2008 book “Food: The Key Concepts:”

For the most part, people decide what to eat based on a rough negotiation - a pushing and tugging-between the dictates of identity and convenience, with somewhat lesser guidance from the considerations of responsibility. (The triangle is thus not quite equilateral, though the moralist might wish it were so.) “Identity” involves considerations of personal preference, pleasure, creativity, the sense of who and where you are. Identity includes factors such as taste, family and ethnic background, personal memories (the association between particular foods and past events, both good and bad). The cultural aspects of identity include widely shared values and ideas, extravagant notions about the good life, as well as a community’s special food preferences and practices… “Convenience” encompasses variables such as price, availability, and ease of preparation… In other words, convenience involves concerns such as “Can I get it?,” “Can I afford it?,” “Can I make it?”

Belasco (2008), p. 7. h/t Jennifer Channin.

Identity + convenience is the synergy I’m most excited about. Instead of putting a default in place and hoping for the best, let’s combine it with concerted, concurrent efforts to change ideas about food and forge new identities. Better yet, let’s tap into pre-existing identities, explain how plant-based diets are consistent with them, and then take folks somewhere to enjoy convenient, delicious plant-based food together.

That, I think, is how a new way of seeing the world might take hold.

References

Bibliography here.

1. Instead of writing substack posts and meta-analyses I could just repeat these over and over and get a lot of my kvetching across. Rossi later writes: “The major sources of program design failures are: (a) incorrect understanding of the social problem being addressed, (b) interventions that are inappropriate, and (c) faulty implementation of the intervention.” Just hook it to my veins Peter!

2. McLean et al. (2020) make a related point about the importance of context in international development work.

3. Me too, Joe. This is not bloodless for me. Our duty to see the world straight on is sacred. HPJEV agrees.

4. Bjorn’s Faunaltytics summary was a big help, as were Meier et al. (2022) and Hummel & Maedche (2019). See here for a complete list of reviews we looked at.

5. This study changes the default portion size of meat rather than making the default plant-based, FWIW.

6. These studies were cluster-assigned and had fewer than ten clusters in total.

7. Bullock & Green (2021) make a related point concerning mediation analysis: that omitted variables correlated with mediators (e1) and omitted variables correlated with outcomes (e3) are a problem even if there’s no big obvious omitted variable that comes to mind. “The threat of omitted variables implies that e1 and e3 may be highly correlated even if no single omitted variable is strongly correlated with M [mediators] or Y [outcome].”

8. Our meta chose a cutoff of ten clusters because of a remark in Paluck et al. (2021), on which I worked as an RA: “Reliably estimating standard errors requires at least 10 clusters, but many intervention studies do not approach even this minimal number.” For more on random assignment, see Gerber & Green (2012), chapter 2.

9. The term for this other estimator is Intent to Treat (ITT). See Gerber & Green (2012), chapter 5 for discussion.

10. Unfortunately, when organ donation by default is scaled up — I swear I didn’t know this before starting this essay — the impact on actual organ donations is small. Comparing countries that have opt-in vs opt-out policies, Arshad et al. (2019) find that “no significant difference was observed in rates of kidney (35.2 versus 42.3 respectively), non-renal (28.7 versus 20.9, respectively), or total solid organ transplantation (63.6 versus 61.7, respectively).” That last number — that the scaled up net difference is 2.1 pp.— sticks out to me. This 1-2 pp. thing is turning into Seth’s Iron Bucket of Cold Water.