Note: We're sharing this evaluation as a cross-post, including most of the evaluation content (with light redactions), to test engagement with this format. All Unjournal evaluations can be found at unjournal.pubpub.org. 

Cultured meat (CM) has been seen as a potential solution to the animal welfare consequences of factory farming, as well as health, environmental, and animal disease vector issues. However, several studies and TEAs between 2020 and 2022 outlined a fairly pessimistic picture on CM’s potential to scale. 

In 2021-2022, Rethink Priorities reviewed some of these sources and used them as a basis for a forecasting exercise, where their commissioned panel of forecasters and experts predicted a low chance of cultured meat reaching significant production volumes. 

We got the impression that this forecasting paper was influential within the EA community and may have contributed to reduced support and funding for cultured meat. Now that we're hearing news of promising developments in CM (albeit from sources with possible conflicts of interest), we wanted to re-evaluate this study in light of the current state of the field. 

Overview

This evaluation package is for Rethink Priorities' 2022 study, authored by Neil Dullaghan and Linch Zhang, titled "Forecasts estimate limited cultured meat production through 2050." We commissioned two evaluators to review this study — David Manheim, an economist/forecaster, and Estere Seinkmane, a biologist working in cellular agriculture. 

According to both evaluators, the picture seems to have changed since 2022 —  they point to a range of newer TEAs and real-world shifts. But they also suggest that even in 2022 the authors made choices and errors that may have biased the results towards a more negative outlook. 

Manheim gives a detailed overview of the methodological and framing choices that may have contributed to this bias. Seinkmane brings expertise in cellular agriculture to this review, noting where the authors’ lack of industry knowledge may have led to some consequential misunderstandings. 

Each evaluator also provided a quantitative assessment:

Below is a slightly abridged version of both evaluations. The full evaluation package, including metrics and references, can be found here.

Manheim's Evaluation

There are different topics I will attempt to address in evaluating this work. The first is evaluating the research as performed and reported, including the soundness and the methodological limits of the exercise, and the resulting validity of the directly resulting outputs. The second is understanding and evaluating the generalizability, including the assumptions, specifically whether assumptions or limitations of the methods impacts the resulting conclusions, and what this means for the implications of the report.

To foreshadow the conclusions, I find several minor issues with the first set of topics (methodology) but very major issues with the second set (framing and conclusions) - though this partly reflects my inside view, formed in large part before the work was done. Critically, the presentation of the Techno-Economic Analyses did not clarify that these analyses are conditional estimates, not predictions, and that high price scenarios were all based on the present-day costs as of the publication time, (which have since dropped significantly) not predicted future cost.

Evaluating the research methods

Forecasting methods

Small sample limitations

The forecast exercise was well designed, albeit with a small sample. It is understandable but additionally unfortunate that this severely limits the conclusions which can be drawn from the difference in views between the single expert and the five forecasters. Given the small sample, there was a fairly clear split in views, where forecasters one and two were quite optimistic, whereas three through six were pessimistic.

Lack of discussion and updating

It is somewhat unfortunate that the forecasters and the expert were not given an opportunity to discuss and update, as this is an important part of what enables superforecaster success. Specifically, Gardner and Tetlock’s (2015) book points out that discussion dynamics and structured collaboration significantly amplify performance compared to individual experts, due to shared reasoning and cross-examination of assumptions.

Geometric means → lower estimates

Pooling using geometric means is very reasonable as a way to account for extreme probabilities, but also provides almost uniformly lower estimates in this study than a simple average. (The transformation into odds had the additional effect of generally reducing the aggregated estimates, though this was mixed.)

Magnitude of units error in the table for Delft overstating the cost of FGF2 1000x

A potentially critical issue was found with the table provided to forecasters and experts for question 24, which says that the Delft analysis had a cost of FGF2/kg of 1.3-2.3B/kg, while the report itself, in table 3, this was the cost per gram, making the numbers approximately 3x what was reported for Humbird, not 3,000x. (The Risner et al price was, in fact, 1000x higher - but was also focused on the promise of eliminating FGF2, i.e. setting the price to zero. Notably, this is a possibility that the forecasts do not seem to take into account.) I do not know enough about this to check it in depth.

Author’s response:  

Hi, Yeah that looks like a typo at first glance.
 Plausible it influenced things a little, though unsure it's critical to the end results. it's unclear to me who actually read the reference material and did not. I quickly looked and see forecaster 3 included some mentions of prices/kg of inputs but it was not immediately obvious they used the numbers we provided in the reference document, so possible they independently made their own errors too.

Results Reporting

Overstated conclusions, hidden disagreement

First, the small number of forecasters seems to lead to an overstatement of the strength of the arguments on which to base the conclusions, and lacking uncertainty ranges compounds this. The aggregate reported results, most notably, a 9% chance of greater than 50m metric tons of cultured meat sold in 2051, hide the disagreement, in this case, the fact that two forecasters reported a greater than 40% probability of this occurring. This is importantly misleading, especially given the conclusions.

Otherwise, the reporting of results seems clear, and the reported results generally make sense, with two exceptions; consumer acceptance, and investment.

Consumer acceptance, misinterpretation of conditional probabilities

For the evaluation of consumer approval, it seems strange to report the absence of correlations between predictions of wide scale protests and willingness to try, given that they are conditional, and the conditional effectively screens off correlation. This undermines the claim that “consumer approval did not appear as a major constraint,” especially because the direction of causation is unclear, and the relationships over time between consumer approval/acceptance and the volume of production will clearly have feedback loops.

Investment, misinterpretation of causality

Similarly, the correlation between investment and funding and cultured meat production is partly a function of similar feedback loops. Forecasters are not predicting a casual relationship, they are predicting outcomes. If investment or subsidies were high, this would impact the production volume - and given increasing returns to scale, this might be a virtuous loop. On the other hand, regardless of viability, if investments and subsidies are low, production will be low. It seems especially unfortunate as this may have become a self-fulfilling prediction; discussions since 2022 reveal that the report itself decreased philanthropic funding, making the field less likely to achieve its potential.

Cruxes

First, it is particularly valuable and commendable to have clearly reported reasoning and the structural relationship between the causes of cultured meat production being low or high and the different prior analyses. Doing so as a way to update from prior work in detail, rather than simply reviewing prior work and presenting a conclusion, is incredibly valuable and yet often unusual in research. (This type of synthesis is often left for systematic reviews, which then aggregate results without decomposing the issues, compounding the problem!) 

[...]

A minor comment is that I would have further confidence if the forecasters were explicitly shown the report and had a chance to comment - and again, it would have been even better if they had been given a chance to update their views. (However, I also note that if the explanation were materially misleading, there could have been responses on the Forum Post which I expect the forecasters would have seen - and no such comments have been made.)

Evaluating Framing and the Question of TEAs

The fundamental question which is being asked is whether cultured meat is a viable product or industry, and what would allow it to succeed. For this reason, the focus of the questions on specific product types, and the technology trajectory, are critical.

Doing so without the benefit of hindsight is functionally impossible, so this discussion will inevitably be at least somewhat unfair to the report’s discussion, which was produced close to four years ago.

Techno-Economic Analysis

To begin discussion of the report’s framing, we need to address TEAs, introductions to which can be found here; 1, 2, and a more detailed but narrow overview here. (These are focused on energy technologies.) While I am not an expert in the area, it seems important to note that fundamentally, TEAs are not predictions, they are conditional analyses. That is, they focus on an in-development technology, and consider what specific changes to a technology’s inputs would allow.

Prediction of changes to input prices are partially premised on no significant changes to the technology. They can, however, be used to address such questions. The Rethink report unfortunately presents the TEA results - from Delft (2021), Humbird (2021), and Risner et al (2020) without clarifying that the different price estimates were for different scenarios, and all of the highest price scenarios were present-day costs as of the publication time, not projections.

It is again unfair to the research to note that in retrospect, the situation has changed - but it is they were using current prices for inputs, and retrospective evaluation itself is valid. Given that, I will note that in the 5 years since the publication of the TEAs, costs for inputs like FGF-2 have already come down by a couple orders of magnitude, below the ranges of analysed costs for 2 of the 3 studies. And the question of lowered input prices was discussed with one author at the time the study was being performed.

To explain the further questions about framing, assumptions, and generalizability, I will first digress briefly to explain how I see adoption in general, and apply it to this case. Again, some of this is probably unfairly retrodictive, seeing advances in the past several years since the report was released.

Technology Adoption Pathways

Fundamentally, the way technology evolves in successful domains is by successive changes to the technology, often spurred by greater adoption and scaling, which reduce prices and increase quality or output, feeding a virtuous cycle.

In starting that cycle, technology adoption typically follows a process, where consumer “innovators” start using a new technology, and then early adopters, and so on. Three key barriers to this exist: consumer adoption, returns to scale, and timelines. If consumers in the successive categories: of innovators, early adopters, early majority, etc. do not want the product at the available price points, the scaling cannot continue. If there are limited or insufficient returns to scale, prices do not come down enough, and quality does not improve enough, to enable cheaper products acceptable to larger parts of the market. And if timelines are longer than expected, companies or industries can falter, for example, running out of investment capital and generating too-low returns to continue, and the market can collapse. (But timelines can also sometimes slip significantly without this occurring. Famously, for Musk’s companies, I’m unaware of a case where Tesla or SpaceX did anything as quickly as initially promised!)

But we need to be careful, because [while] the typical story follows the paper’s analysis and assumes a single technology product - [in contrast,] when looking at something like cultured meat, we are considering a class of products, more like “electric vehicle” than “solar panel.” And as noted, successful companies or industries often start by marketing a luxury product to prove viability, and expect production prices to come down with scale. As an obvious example, Tesla began with an electric sports car, then expanded to a luxury sedan, and at least initially had the idea that it would culminate in a low-cost vehicle.

The Cultured Meat Case [and luxury products]

Given the three barriers, I want to break down the question, and consider the path for luxury products. The discussion in the research was focused on adoption of meat in general, and didn’t explicitly consider this. In fact, this critical issue for the analysis is ignored; meat is not a single product, it is a very significant range of products with different relevant characteristics and price points. Much like “electric car” was a range of products, we expect that successful adoption curves go through luxury products first.

But all of the TEAs were about the case for mass production of meat, not luxury products. Because TEAs are used for cost-benchmarking, and require a comparator, this will be misleading. The Delft TEA used a comparator of “slurry or paste made out of meat cells,” which could be sold as something like ground beef, while the other two did not have an explicit product comparators. (Humbird (2021) and Risner et al (2020) also mention prices for hamburgers in passing. On the other hand, Risner et al also explicitly notes, “there may be opportunity for viable competition in the specialty foods markets, where ACBM costs compare more favorably to such items as almas beluga caviar (USD 10,000/kg), Atlantic bluefin tuna (USD 6500/kg), and foie gras (USD 1232/kg) [57]”)

Consumer adoption seems like a critical barrier when considering mass-market adoption, but is far less concerning for luxury products; as Business Insider put it, “Being ethical is the new luxury.” Capturing a fraction of the luxury market is enough for profitability, at least in the interim - and doesn’t require scaling up to nearly the same extent.

Neil said that he would change his mind if there was “Public demonstrations of large batch production (>1,000 metric tons [at any price]) in the next two years.” It has been three, and this has not happened. On the other hand, given regulatory and production delays, I think we’ll have significant updates over the coming year, with factories coming online.

Wildtype, Upside Foods, and Believer Meats all have projects which have capacity to produce over 10,000kg/year online sometime in 2024/2025. Upside food for chicken, and Wildtype’s salmon are targeting in the range of 25-250 tons each. A third company, Good meat, claimed in late 2023 that it could scale production to 1,300 tons at their new production facility, though there is no evidence they did so, and even then, it would have happened later than the suggested demonstration [of actual production] that Neil suggested.

Smaller production volumes for luxury product is also starting. Wildtype’s pitch starts into the slightly higher priced products, as traditional salmon sells for over $10/pound, approximately double the price of ground chuck beef, and easily quadruple the price of whole chicken. But Vow foods is now producing [lab-grown] pate and foie gras; both seem to address a luxury market segment at prices closer to $100/kg, with a successful single batch production of half a metric ton. They claim to be able to produce over a ton per week, which would be a scale similar to the lower cost meat production of other firms.

The question now is whether the adoption of small amounts of luxury goods really leads to drastic price decreases for mass market consumer goods. Tesla, again, is an interesting comparator; they still have not managed to do what was originally promised, and delivered an electric car cheaper than comparable gasoline cars - though recent analyses do show that total cost can be lower.

None of these are reaching the levels needed to replace meat. On the other hand, the production volumes are potentially finally reaching levels that could potentially spur the virtuous cycle of price reductions. Given that, I expect the forecast average correctly predicted a failure to reach 100,000 metric tons by 2031 - but see the 46% prediction of reaching that volume by 2051, in retrospect, as being far too low; Tesla may not have made electric cars affordable, but the technology had, and has, far more ability to drop in price.

Conclusions

The summary of the report suggests that “The main cruxes of disagreement appear to stem from [differences in] beliefs [over] how often technology can replicate and outperform biological systems, [over the] choice of reference classes, and [over] how much to anchor based on the estimates in the Humbird (2021) techno-economic analysis.”

I think this is correct, but betrays confusion on the part of the analysis. The question should not be just focused on how often technology can replicate and outperform biological systems, but how quickly it can occur. I agree with the conclusions of the report that the timelines and near-term predictions for cultivated meat do not seem to plausibly support the optimistic predictions that were made by both industry and consultancies for the coming decade or two.

If we want to ask the question, “Is Cultured Meat likely to be commercially viable in the near to medium term?” It seems the answer should be yes, despite not reaching very large scale volume. But the analysis doesn’t address this question; profitable luxury products are implicitly excluded. On the other hand, this [small-scale luxury production] does not directly address the question of the eventual impact on animal welfare.

That is where the final conclusions seem greatly overstated; a 50% chance that a material proportion of meat is from cultured sources is also necessarily envisioning a scenario where there is immense pressure to solve the technological problems. And the case for reaching price-parity does not need to be airtight to justify investment, for a few reasons. First, the cost of cultured meat isn’t the only relevant trend; as the continued critical need to stop and reverse climate change meets the growing demand in the developing world for luxury goods like meat, it seems likely that both the price, and legal, moral, and economic pressures [on] traditional meat production, increase. Second, the case for reaching price-parity does not need to be airtight to justify investment; a 1-in-11 chance that over 50M metric tons of cultured meat is produced in 2051 is plausibly a good investment just based on the return on investment; the philanthropic impact adds to that. Lastly, whether or not the world reaches net-zero by 2050, and whether or not ethical pressure against factory farmed meat consumption accelerates, there is reason to expect cellular agriculture to become increasingly prominent as a target for investment. It seems very likely that something will displace current factory farming. And unless the barriers to cultured meat are far stronger than the ones suggested in the analyses, it seems cultured meat still has a significant chance of being a large part of that change.

Of course, the promise of cultured meat is not realized by reaching 50m metric tons, but by displacing the current production of well over 500m metric tons of meat and seafood yearly - a number that is increasing rapidly as the developing world gets richer. But the trajectory to get there is important, and seems to route through current progress in artificial meat. A principle unaddressed question for the timeline, then, is whether humanity is willing to invest in longer-time-frame innovations that will pay off later. Philanthropic funding seems likely to play a key role - so for those who prioritize animal welfare, the partial cessation of funding based on this analysis seems very unfortunate.

Seinkmane

My evaluation focused on assessing the assumptions and providing up-to-date context and background for the forecasting paper. The paper (post) was published in 2022, and relied on 2020-2021 data. Since then, at least eight cultivated meat (CM) products have been approved, claimed costs are below $20/kg, CM researchers exceed 250, and total funding surpassed $3B, coming from VC, public research funding, and philanthropic organisations. The CM definition (cell types, species, % in hybrid products), the media cost components, and other relevant technologies could be clarified, especially in the context of more novel findings, to improve further CM forecasts.

Why D&Z post and forecasts are now out of date

My primary concern is that as of August 2025, forecasting by Linch and Neil has now become out of date: published more than 3 years ago, and relying on the analysis from sources – primarily the Humbird TEA – from 2020-2021. Considering that in turn Humbird relied on data available at the time of his analysis, presumably coming from work conducted within a couple of years before that (as there is an inevitable delay between the most current practices and the published data), the assumptions and context are now more than 5 years out of date. For a field that has just emerged and is developing rapidly, 5-7 years makes a huge difference.

Approved products

For instance, even in 2022 there was only one CM product approved and available, in one country in the world – Eat Just/GOOD Meat chicken nuggets in Singapore. In 2023, there have been a series of regulatory approvals and proof-of-concept restaurant launches in [the] US, and currently to my knowledge there are 8 companies/products approved in US, UK, and Asia-Pacific, with more companies having already filed for approval too.

Media and production costs

The production volume – the key metric in this forecast – is obviously linked to cost. The forecasts particularly rely on data and assumptions from the Humbird TEA. There have been multiple approaches taken to reduce media cost, both the cost of individual components (e.g. use of plant, algal and yeast hydrolysates instead of purified amino acids) as well as the development of other technologies that indirectly but significantly decrease the cost of required media – e.g. cell line engineering (used to reduce reliance of cells on growth factors) and media recycling. Some of these approaches are mentioned in the post, and some aren’t, but they are already being used by the companies mentioned above developing and gaining approval for their products, as well as in R&D by researchers. 

In their “what would change our probabilities”, Linch says prices below $100 per kg. The latest claims from the companies are <$20 per kg (Aleph Farms latest TEA; Patsika et al 2024, linked to Believer Meats). Admittedly, these are claims from companies with a clear COI and not all data is publicly released. However, many companies are reporting similar numbers, and it shows the progress has already possibly been quicker than the forecasters have estimated and [quicker than what] the data they relied on suggested.

Number of researchers

Beyond production volume and cost, in the summary the authors state:

“Engineering new types of bioreactors, building out supply chains of key ingredients, securing broad political support for public funding, and establishing a pipeline of researchers seem promising ways to nudge production trajectories upwards over a timeframe of a century, not decades.”

I think the developments in the last 5 years have already shown this statement is at least partially incorrect in its timeline assessment. I focus on the number of researchers and funding in particular here.

For the number of researchers, their cutoff was >250 researchers by 2036 but already in 2025 I think we have long surpassed it. Even in their post, they say GFI lists 60 researchers as of 2022 – however these are mostly principal investigators, so the actual number of active researchers  would have been higher even at the time, including PhD students and postdocs. [...] As of August/September 2025, GFI’s ecosystem map lists 220 researchers (mostly PIs-only), 169 companies and 6 research centres worldwide working specifically on CM.

Funding

In the last three years there has been a downturn of VC funding going to CM, linked both to the “hype cycle” and the overall biotech investment downturn and macroeconomic factors. Even taking that into account, GFI reports $54M and $46M rounds raised by CM companies in 2024, with total CM funding to date surpassing $3B– again, already providing some “social proof” evidence Linch was looking for in the “what would change our probabilities”.

Importantly, public funding, especially in the UK and Europe, has been increasing significantly over the last few years. The Netherlands invested $65M into cellular agriculture primarily focusing on CM (although cellular agriculture also includes precision fermentation). The main UK public funding body, UKRI, has awarded £12M to a cellular agriculture hub, plus about £30M more to two other hubs focusing more widely on alternative proteins including CM.

Finally, big philanthropic funding has also become available. Apart from GFI and New Harvest grants awarded to specific areas thanks to the non-profits’ donors, Bezos Earth Fund has committed major sums towards sustainable protein research hubs. Three have been established (in the US, Singapore and the UK), with total funding of $100M so far. 

Things that could have been done better [provided in order] to help make future synthesis and forecasting exercises better

Cell type and species

Neil & Linch have focused on beef (their questions refer to “cow cells”) and they don’t define cell type very specifically, I assume they mostly meant bovine muscle cells. This is understandable as this cell type has been the focus for much of the early CM research, including the first CM burger by Mark Post in 2013. However, there is now a lot of interest in various types of CM: cells from different species (incl. pigs, chicken, duck and various fish species), as well as different tissue types (muscle, fat, connective tissue). I think for future forecasts CM type should be better-defined, as both the species and the tissue type influence the media requirement – and therefore the cost – to a great extent (O’Neill et al 2022, Leber et al 2025)   

Proportion of CM in the product

Neil & Linch choose cultivated meat to be defined as 51%+ cultivated. Indeed it matters a lot as hybrid products are the obvious first step towards fully CM products, and in many cases considered to be a “final product” too, with evidence that even a smaller portion of CM in a plant-based product can have a big impact (e.g. Alam et al 2024). So I think the best approach for predictions would be to focus on the cost/total production volume of the CM portion of the final product, which can vary depending on the product.

Diversifying reference materials and including cell biology and bioprocessing expertise

The authors admit themselves that they relied primarily on one source – the Humbird TEA – with the addition of a couple other papers. I have compiled a list of literature that would hopefully provide other and more recent sources to facilitate future forecasts. Aside from peer-reviewed publications and regulatory dossiers + patents available that have already been submitted for product approval, GFI releases information regularly, including their state of industry reports and reports focusing on specific topics such as media ingredients. 

The authors also admit they do not have background in the field, neither CM-specific nor wider cell biology or other relevant technical expertise. Their chosen forecasters – apart from one – do not have that background either. For someone who has worked with mammalian cells and cell media, it appears that the authors do not have clear understanding of cell media components, which they also know and state “We incorrectly included recombinant proteins in the question wording and transferrin and insulin in the reference material for the amino acid question. As these are unrelated and more expensive, it is possible this biased forecasters’ estimates. None of the forecasters seemed to notice this error in their reasoning.” [...] Therefore for any further forecasts and synthesis I would strongly recommend further consultations with relevant experts, both in the CM field and in adjacent fields (wider cell biology and bioprocess engineering backgrounds).

Takeaways

Both evaluators discussed areas of improvement for future studies. Manheim suggests a larger sample, avoiding questions involving hard-to-interpret conditional probabilities, and an opportunity for forecasters to discuss among themselves and update their forecasts. Seinkmane recommends more collaboration with field experts and drawing from a greater range of sources. Even so, Manheim also praised the report for its reasoning transparency and overall design. 

Our goal for this evaluation, and for our related Pivotal Question, was to take steps towards understanding the impact potential for investments in cultured meat, predicting its potential with implications for other animal welfare interventions, and informing the design of future forecasting studies. We hope this evaluation is informative for funders, advocates, researchers, and anyone interested in the future of cultured meat. 

Credits

Original evaluations authored by Estere Seinkmane and David Manheim. David Reinstein was the evaluation manager and provided a summary and discussion. Adapted for the EA Forum by Ashley Meader.