From narratives to arguments and actors

In the Narrative framing approach, we looked at how an issue is being discussed. Here, we go a step further and ask what specific positions are being taken, by whom, and who aligns with whom? Rather than tracking broad themes, we extract concrete claims from the debate, identify which actors support or oppose them, and map the resulting coalitions.

To illustrate the approach, I applied this approach to a live policy debate: alternative proteins in Europe.

Alternative proteins in European media

Alternative proteins — plant-based meat, cultivated meat, and fermentation-derived products — are a deeply contested food policy question in Europe. The debate sits at the intersection of EU climate policy, Common Agricultural Policy reform, food labelling disputes, public health concerns and industrial policy. This makes it a relevant test case for arguments and actors mapping, one where the stakeholders are relatively diverse and the regulatory stakes are real.

To understand how the debate is structured and how it evolved, I analysed articles referring to alternative proteins published between 2015 and 2025 across 21 major European media outlets¹ in seven languages. This represents a corpus of 3,500 documents. The analysis identifies statements, attributes them to named actors, induces a set of recurring claims from those statements, scores each statement’s relationship to each claim (supports, opposes, or neutral), and then maps actor relationships based on their respective positions.

The arguments: how is the debate structured?

In this approach, debates on a given topic are structured around claims: specific, debatable propositions that actors may support, oppose, or simply not address. Taken together, claims define an n-dimensional “stance space” that we can use to represent and compare actors.

Choosing the right set of claims is a critical design step: it defines both what we track and how we map actors & coalitions. Claims could be left to automation when the goal is to surface new and unexpected arguments, or manually defined as a narrow set to inform very targeted advocacy. Here, we adopted a hybrid approach: claims were first automatically induced from actor statements in the corpus, then manually refined into two families: product claims (what alternative proteins and conventional agriculture are and do) and policy claims (what governments and regulators should do).

Coalitions: who aligns with whom?

We then estimate to what extent each actor supports or opposes these claims based on their collected statements. This allows us to derive coalitions of actors i.e. groups whose positions align, whether or not they coordinate or even know of each other² (see Method overview below). The scatter plot below visualizes these induced structure: each point is an actor with colours indicating detected coalitions.

Three broad coalitions emerge. The heatmap below shows each coalition’s average position on each of the 16 claims.

To interpret what each coalition represents, I fed the stance centroids and actor lists to Claude Opus 4.6 and asked it draft their positioning profiles. Here is what the model produced:

• Coalition A (160 actors) — “Skeptics & food quality critics.” Nutrition researchers, food-culture defenders, and political opponents — Chris Van Tulleken, Carlos Monteiro, Marco Springmann on the UPF research side; Ron DeSantis, Giorgia Meloni, Viktor Orbán, Jim Pillen in politics; Piers Morgan, Joanna Blythman, Marion Nestle in food media and commentary. Strongest on the ultra-processed critique (0.36), stricter UPF regulation (0.36), transparent labeling (0.32), and traditional food culture (0.34). Skeptical of health claims (-0.28), food technology (-0.30), and market potential (-0.21). Mildly protective of farming livelihoods (0.09) but the identity is more about food quality and naturalness than economic protectionism.

• Coalition B (133 actors) — “Industry & market optimists.” Startup founders, investors, food-industry executives, and some scientists — Mark Post, Sandhya Sriram, Decker Walker, Benjamina Bollag, Tomer Halevy, Justin Kolbeck, Miyoko Schinner, Emmanuel Macron, Boris Johnson. By far the strongest signal on market potential (0.91) and food technology (0.67). Mildly positive on traditional food culture (0.15) and essentially neutral on meat reduction (0.10), farming protection (0.03), and the ultra-processed debate (-0.14). The commercial and pragmatic voice: enthusiastic about the opportunity without taking strong positions on the cultural or dietary politics.

• Coalition C (164 actors) — “Systemic transformation advocates.” Mission-driven founders, advocacy figures, public intellectuals, and policy voices — Patrick Brown, Bruce Friedrich, Josh Tetrick, Uma Valeti, Seren Kell, Ethan Brown, Bill Gates, George Monbiot, Paul Shapiro, Chris Bryant, Molly Scott Cato. Strong across product and policy claims: food technology (0.59), meat reduction (0.55), environment (0.47), market potential (0.43), health (0.32), animal welfare (0.23). Firmly rejects the ultra-processed framing (-0.67), traditional food culture (-0.32), and farming protections (-0.18). The most ideologically coherent coalition — pairs product enthusiasm with an explicit push for dietary and systemic change.

To browse coalition actors by coalition and explore their individual positions, you can use the interactive browser below.

A first observation is that the coalition A bring together people with very different perspectives: researchers like Marco Springmann may be critical of certain forms of processed alternative proteins, but are otherwise strong proponents of plant-based diets centered on unprocessed legumes. This contrasts with other actors in this coalition whose opposition to alternative proteins comes from a defense of the livestock and meat industry. One reason this coalition appears so broad is that the 16 claims lack the granularity to separate these actors’ positions: they don’t distinguish between types of alternative proteins (which atttract different criticisms) or clarify what APs are being compared to (meat, or plant-based diets). I am currently working on a methodological refinement to better capture such differences.

Another insight lies in the B–C split. Both coalitions are pro-alternative-proteins, but they seem to (at least in appearance) disagree on whether the goal is to add new products to the market or to replace the existing food system. For an advocate, this could mean that Coalition B actors are potential allies on technology and investment, but unlikely partners for campaigns framed around meat reduction or challenging farming interests. Or the other way around, it could help identify which Coalition C actors could be moved on meat-reduction messaging.

Notably, all of the 457 actors were identified automatically from the corpus — no names were pre-listed or manually selected. This means that the method can surface voices an analyst might not have thought of, and do so in virtually any language. To see each individual actor’s positioning across all 16 claims, open the full actor-level stance heatmap.

The same analysis can also be run at the organisation level rather than individual actors — applied to the same corpus, it yields a broadly similar coalition structure (see Appendix: Organisation-level coalitions below).

How arguments evolve over time

Beside mapping the coalition, we may want to track whether claims become more or less prominent over time. This is done in the chart below, showing the evolution of supporting/opposing statements per claim.

The charts show a debate that has shifted over the past decade. Early coverage was dominated by product-level claims: the environmental case, food technology, market potential. These peaked around 2020 — roughly the alt-protein investment boom — and have since declined.

Since then, a different set of arguments has been gaining ground: the ultra-processed critique, the dispute around labeling and the protection of traditional food culture. This is essentially Coalition A’s vocabulary.

One may want to normalise these trends (e.g. by means of a separate topic acting as a “control-group”) or weighing them (e.g. by the media readership or importance) and see if the findings hold.

Looking forward

The alternative proteins case above is one potential application, but the method itself is topic-agnostic and can be applied to any policy debate where actors take public positions. The corpus can also be extended to broader types of documents, including TV and radio broadcasts, podcasts, parliamentary debates, policy consultations, political manifestos or social media³, in virtually any language LLMs are sufficiently trained on.

Below are some of the ways this could support advocacy organisations or their funders:

Early detection of emerging arguments: By running the pipeline regularly on a rolling corpus, new claims can be detected as they first appear, before they gain traction. This could give advocates lead time to prepare responses or preemptive messaging, rather than reacting once a narrative is already established.

Surfacing unexpected actors: Because the pipeline identifies actors automatically from the text, it can flag voices that an analyst might not have thought to look for — new entrants in a debate, unusual alliances, or actors whose influence is growing but who aren’t yet on anyone’s radar. For an illustration, see the 457 actors identified in this demonstration.

Regular landscape assessments: Running the analysis on a quarterly or yearly cycle would produce a structured, comparable snapshot of the argument landscape and coalition structure over time. This could serve as an intelligence product for advocacy organisations or funders seeking to understand how a debate is evolving.

Guiding media outreach: The cross-outlet analysis (which arguments appear where, and with what stance) can inform where to pitch stories, which outlets are more receptive to certain framings, and where there are gaps in coverage that could be filled.

Measuring campaign impact: Mapping the argument landscape before and after an intervention — a campaign, a policy event, a controversy — could help detect whether anything shifted: new arguments emerging, actors moving, coalitions reconfiguring. This is more speculative and would require careful methodological work to distinguish signal from noise, but the structured nature of the data makes it a plausible direction.

Testing positioning: An advocate could feed a message or position and be indicated which actors and coalitions would likely support or oppose it, and how to adjust the framing to appeal to a broader coalition. This turns the landscape map into a design tool.

Following actors trajectories: Rather than a single snapshot, tracking how individual actors move through stance space over time reveals who is shifting, who is entrenching, and in which direction.

Assessing actors’ influence: The current analysis maps who says what, but treats all actors equally. A natural extension is to estimate how important actors are in shaping the debate. This could be done by analysing actors’ reach, their network position, and tracking how their arguments are picked up by others over time.

Disclaimer

This analysis is a prototype built to illustrate the method, not a finished intelligence product. The analysis should not be relied upon to provide accurate estimates of actors positioning: the results shown here have undergone limited manual sanity checks and have not been systematically validated against hand-coded datasets. Further validation and methodological refinement are needed before these results can be used for strategy purposes.

Acknowledgements

I am very grateful to Mo Putera and Thomas Manandhar-Richardson for their insightful feedback. Thanks also to Carl Frederick Luthin for suggesting to work on this topic.

Changelog

• 2026-04-01 — Added comments on Coalition A heterogeneity.