As public concerns about climate change continue to grow, it’s clear that emissions of greenhouse gases must be reduced on a systemic and individual level. One of the most immediate ways for anyone to reduce their own environmental impact is with a change in diet – but for most people, how to do that isn’t necessarily obvious.

Researchers at E.Mission, a West Midlands climate advocacy agency, have coined the term “carbon forkprint” as part of their work educating the hospitality sector on climate action. Together with Riza Batista-Navarro, a lecturer in text mining at the University of Manchester, they derived a plan to automatically give people suggestions for how to reduce the carbon impact of their home cooking, powered by collective intelligence.

“There’s so much attention on ways individuals can reduce their carbon footprints, but it’s a minefield trying to understand it all,” says E.Mission director of partnerships James Bagshaw. “The two biggest chunks of emissions come from transport and energy, but public transport is an urban planning issue, while electric vehicles, heat pumps and insulation are expensive.” Food, however, is the third-biggest contributor to emissions and can be mitigated on an individual level immediately without big costs.

Their plan was to develop a browser extension that would recognise when people looked up a recipe online, review its ingredients and calculate the recipe’s total carbon emissions. It would then suggest a lower-carbon ingredient to swap in and reduce the original dish’s overall footprint, or the user could opt for a different lower-carbon dish altogether. “At its core this is about giving people agency over climate change,” says Tom Maidment, E.Mission’s technical director. “It’s about getting this data to people, because if you can’t apply it at the point of decision it’s irrelevant.”

The experiment was structured in three phases, to test the three following hypotheses:

1. “Carbon Forkprints” of online recipes can be automatically calculated by a machine learning algorithm
2. machine learning, powered by collective intelligence, can identify and suggest appropriate lower-carbon recipe ingredient
3. a browser extension will be a persuasive format for presenting those suggestions to the public

130 members of the public were recruited to participate in the experiment. While the cohort was skewed more towards women than men (reflecting the division of domestic labour in many homes), and had a higher-than-average proportion of graduates, it still represented a “relatively good” demographic and geographic sample of the overall UK population. Each participant was asked to provide five different online recipes they’d normally use, from sites such as BBC Good Food, food blogs, and more. After eliminating duplicate, video and non-English submissions (for being outside the scope), there was a starting corpus of 587 unique recipes.

Stage one: Automatic “forkprint” estimation
Batista-Navarro and her research team scraped each recipe from the web and built the training dataset for a machine learning algorithm—the first time she had developed one specifically for recipes. Names, quantities and units of measurement were collated for each ingredient, with the names automatically recognised using a model trained on an existing body of recipes with food names annotated. That food data had to then be linked with existing data from E.Mission’s own carbon footprint database.

“This is where we had all sorts of challenges,” says Batista-Navarro. “Maybe the database uses ‘aubergine’ and a recipe uses ‘eggplant’. If a recipe says ‘one apple’, what does that weigh? But through entity linking we were able to retrieve information from our own ‘footprint knowledge base’.”

Stage two: machine learning identification and suggestion of ingredient substitutions
The team went back to the original participants and gave them each 15 recipes – their own original five, plus ten more randomly picked from those submitted by others. In each recipe, the team highlighted the highest-carbon ingredient (usually meat or dairy) and asked each participant what they would do if they had to swap it out for something else, whether replacing it with something else or removing it from the meal altogether. Crucially, however, they weren’t told that the experiment was about reducing carbon emissions at this stage.

“It needs to be an open choice, even if it’s a higher-carbon swap,” says Bagshaw. “If people opt for beef because chicken isn’t available for a curry, that’s good to know. Just because a swap is low-carbon doesn’t make it suitable in terms of taste, nutrition or the format of the recipe. We wanted to collect the swaps from the crowd and then feed them back to the crowd, because they’re more likely to be accepted than us just spitballing. That’s the whole point of the collective intelligence in our experiment – it needs to not have an agenda when collecting the data.”

Stage three: Suggesting swaps via a browser extension
A bespoke browser extension, built for this experiment, was distributed to the original cohort. Whenever it detects a user browsing a recipe, it suggests up to two swaps per dish, targeting the highest-carbon ingredients. If users chose to cook a recipe, they hit a button to launch a questionnaire that asks why, and whether they also accepted a suggested swap. After five recipes, participants submitted a further, final questionnaire about their overall experiences with the study.

The participants suggested multiple substitutions for 348 of the recipes in stage two, generating a dataset of 1,277 substitutions containing both “positive” (ie, both lower-carbon and culinarily suitable) and “negative” substitutions. A neural network was trained on 80% of that dataset to understand what made a swap “suitable”, and the remaining 20% was used to test the accuracy of that training, defined as the AI successfully identifying the same substitutions that the crowd had suggested.

While that came to 85% accurate, the model became significantly worse in the final phase of real-world testing, with recipes that weren’t in the original dataset at all. It particularly struggled with contextualising some substitutions where one-for-one substitutions aren’t viable – with eggs, for example, which can act as a binding agent and can’t be replaced easily without changing the fundamental nature of a recipe. Batista-Navarro believes that five substitutions per recipe isn’t enough for the model to learn from. Having “perhaps as many as 50 substitutions per recipe” would be ideal, but as that seems an implausibly high number to expect for most recipes it may prove a problematic aspect of the process to scale up.

“However, even with the shakiness of the substitution model in the real world, we still supplied a swap in 80% of cases, and 25% of those were accepted,” says Maidment. “Our most common swaps were from dairy to plant dairy, but there were also swaps such as beef to pulses, which is massive—it can be as much as 10-20 kilos of carbon per meal, a more than 90% saving.”

“Forkprints” often follow a long-tail pattern, where a small proportion of meals represent the bulk of someone’s overall food emissions each week. In the initial cohort, 50% of emissions came from only 7.4% of meals, while 30% of meals generated 80%. After substitutions, though, only one meal from the entire cohort exceeded 50 kilos of carbon. “We saw an average footprint reduction of 80%, which was the real benefit,” says Maidment. “Not much changed for the majority of meals, but that long tail shrank, and high-carbon meals disappeared.”

“We were blown away by the savings from just this one cohort,” says Bagshaw. “There was an average of 3.9 kilos of CO2 saved per participant. If a thousand people made similar swaps, that’s equivalent to driving a petrol car 27,000km – two-thirds the way around the planet. That’s a lot of kilometres for not a lot of people.”

While this experiment is only a proof of concept, “I think we’ve proved that concept, even with this somewhat self-selecting cohort,” argues Maidment, “and in many ways these are quite rough substitutions. We could definitely improve, and get even more significant reductions.”

Those future improvements could include factoring in other qualities like nutrition and taste into the swap model, which is currently only trained to target emissions when listing suggestions. The model’s web page parsing also needs refinement, and the browser extension still only works on laptops and desktops, not phones.

The team is in very early discussions over partnering with “large food organisations” to roll out their tools in other contexts, with the potential to save up to a million tonnes of CO2 annually if their results scale successfully. They’re also exploring other functionality improvements to the extension, such as including a text box where users can suggest further ingredient substitutions.

“We’re never going to abandon the collective intelligence part of this,” says Bagshaw. “This is a tool that, with users’ permission, will be forever gathering data about the recipes they’re reviewing, choosing, and cooking, and the substitutions they’re accepting or rejecting, to make the model better.”

“I think users appreciate collective intelligence more when they can see why you’re collecting the data, and what you’re using it for,” says Batista-Navarro. “Collective intelligence works best when it ties back into the users, and that’s part of the success of this project.”

For more information about this experiment please contact [email protected] or [email protected] and [email protected]

The opinions expressed in this publication are those of the author. For more information, view our full statement on external contributors.