Seeing Green: When AI Learns to Detect Corporate Illusions

Advertisement first, evidence later.

That is not a moral complaint. It is a business model. A company does not need to lie outright to reshape public perception. It can show a wind turbine, a smiling engineer, a school visit, a research lab, a family cooking dinner, a national flag, or a vague line about “the energy future.” The viewer receives a feeling before receiving a claim. Conveniently, feelings are harder to audit.

This is the territory mapped by Morio, Rowlands, Stammbach, Manning, and Henderson in A Multimodal Benchmark for Framing of Oil & Gas Advertising and Potential Greenwashing Detection.¹ The paper introduces a benchmark for detecting strategic framing in oil and gas video advertising across Facebook and YouTube. Its core contribution is not that AI can now convict companies of greenwashing. It cannot. Please put the courtroom robes away. The contribution is subtler and more useful: it shows how multimodal models can begin to classify the recurring visual and textual frames through which corporate environmental narratives are constructed.

That distinction matters. A greenwashing verdict requires comparison between communication and underlying performance. This benchmark mostly measures the communication side: what a video appears to emphasise, omit, soften, or wrap in pleasing imagery. The paper’s real value is therefore diagnostic. It gives researchers, journalists, ESG analysts, regulators, and corporate affairs teams a way to turn a previously squishy phenomenon — “this ad feels suspiciously green” — into structured evidence.

Not proof. Evidence. Already an improvement over vibes with a sustainability font.

The mechanism is selection, not fabrication

The paper starts from a classic idea in communication theory: framing is about selecting aspects of reality and making them more salient. In oil and gas advertising, that selection can do a great deal of work.

A campaign does not need to say, “We are solving climate change.” It can instead show children, hospitals, workers, domestic energy security, community investment, clean technology, and renewable-looking infrastructure. The implicit argument becomes: we are necessary, responsible, modern, and socially embedded. The fossil fuel base remains present, but politely out of focus.

This is why a purely text-based approach misses part of the story. Corporate advertising is not only a sequence of claims; it is a choreography of images, captions, voiceovers, music, and timing. In the paper’s dataset, many videos either lack usable transcripts or rely heavily on visual suggestion. That makes multimodal analysis not a technical luxury but a requirement. If the message lives in the footage, an auditor reading only the transcript is inspecting the wallpaper and missing the house.

The authors divide the benchmark across two platforms because the persuasion mechanisms differ.

Facebook ads in the dataset come from earlier work on climate obstruction advertising in the United States. Their labels are more directly tied to messages such as community benefits, job creation, pragmatic necessity, “clean” gas, emissions reductions, non-energy uses of oil and gas, and domestic energy security. These are fine-grained climate-obstruction frames, inherited from text-oriented prior datasets and then aligned with associated videos.

YouTube is different. The authors found that the Facebook schema did not fit official corporate YouTube videos well. Those videos tend to be longer, more polished, and more impressionistic. They do not always make blunt claims about jobs or energy independence. They suggest. So the YouTube side receives a different set of labels: Community and Life, Economy and Business, Work, Environment, Green Innovation, and Patriotism.

That split is important. The paper is not treating “greenwashing” as a single universal smell. It is building two related but distinct measurement tasks: one closer to climate obstruction framing, the other closer to corporate image construction through impressions.

This is the mechanism-first reason the benchmark matters. It is not just asking whether a model can label an ad. It is asking whether a model can recognise the structure of corporate self-presentation when that self-presentation is split across words, images, and implication.

The dataset turns corporate atmosphere into labelled evidence

The dataset contains 706 videos: 386 from YouTube and 320 from Facebook. Together, they cover 35,476 seconds of footage, 5,607 transcript segments, 1,183 labels, more than 50 entities, and 20 countries. The YouTube videos span 2010 to 2025; the Facebook videos come from a narrower U.S. political advertising context.

The numbers are not huge by internet-scale AI standards. They are large for expert-annotated, domain-specific, multimodal climate communication research. That distinction matters because the bottleneck here is not scraping; it is judgement. A generic web dataset can contain millions of clips and still be useless for measuring how environmental legitimacy is constructed.

The YouTube annotation process also reveals the difficulty of the task. The authors refined guidelines over multiple rounds and report a final Fleiss’ kappa of 0.61, while also noting an alternative calculation of 0.46. That is not a scandal. It is the point. These labels are not “cat” versus “not cat.” They are soft impressions: whether a video conveys environmental responsibility, work, community contribution, green innovation, economic development, or patriotism. Such categories overlap in the real world because corporate advertising is designed to make them overlap.

A reforestation clip, for example, can imply both Community and Life and Environment. A lab scene can imply Green Innovation, but also Environment, and perhaps Work. A national energy message can become Patriotism only if the cultural signal is legible. The model is being asked to classify ambiguity that humans themselves need discussion to stabilise.

The paper handles this honestly. It does not pretend that annotation is pure. It treats the ambiguity as part of the domain. That is refreshing, because nothing says “robust ESG analytics” quite like pretending soft political communication is a spreadsheet cell from heaven.

The benchmark asks models to read the whole performance

The experimental task is multi-label classification. Given a video, a model must output the relevant framing labels. The authors benchmark six vision-language models: DeepSeekVL2, InternVL2, Qwen2.5-VL at 7B and 32B sizes, GPT-4o-mini, and GPT-4.1.

The input pipeline is practical rather than exotic. Videos are split into sampled frames. Transcript segments are generated with Whisper-1 where speech is available. Frames are paired with transcript timing where possible. If no transcript exists, frames are sampled uniformly. The prompt gives annotation instructions, the sampled frames, transcript segments when available, and, in some settings, one labelled training example.

The study tests zero-shot and one-shot settings. The one-shot setup is especially interesting because the authors do not merely pick a random example. They propose an entity-aware retrieval method: restrict candidate examples to the same entity, then use CLIP embeddings over frames and transcript segments to find a similar training video. The logic is sensible. A company’s videos often have a recognisable communication style; an example from the same entity may teach the model which cues matter.

This is not a grand new architecture. It is a benchmarking and prompt-construction study. That makes the result more operationally relevant. Many organisations will not fine-tune a specialised video model for ESG communication analysis. They may, however, retrieve a similar prior example and give it to a VLM. The paper is testing a workflow that looks uncomfortably close to how real corporate monitoring tools will actually be built.

The main evidence: models recognise obvious frames before subtle ones

The headline result is mixed in the useful way. Models can do part of the task, but they are not ready to become autonomous greenwashing judges.

On YouTube, GPT-4.1 performs best overall in the zero-shot setting, with a micro-averaged F-score of 71.0%. GPT-4o-mini and Qwen2.5-VL 32B also perform reasonably well. The clearest categories are the ones with visible or textual anchors: Community and Life, Work, and Environment. Families, workers, renewable imagery, and direct environmental language give the model something to hold onto.

The difficult categories are the subtle ones: Green Innovation, Economy and Business, and Patriotism. These labels often require inference about intent, culture, and context. A lab scene is not automatically green innovation. A refinery worker is not automatically a jobs claim. A flag may be decoration, nationalism, or energy-security signalling depending on the surrounding message. Models can detect objects; interpreting corporate semiotics is less tidy.

The paper’s abstract gives the contrast neatly: GPT-4.1 detects environmental messages with about 79% F1, while the best model reaches only about 46% F1 on Green Innovation. That gap is the article hiding inside the benchmark. “Environment” often has recognisable icons. “Green Innovation” requires understanding the difference between a claim of future transformation and a stock shot of someone in safety goggles pretending to be the future.

Facebook shows a related pattern, but with stronger gains from one-shot prompting. In the one-shot setting, GPT-4.1 reaches 72.6% micro-F1 on Facebook, and Qwen2.5-VL 32B reaches 70.5%. The improvement suggests that fine-grained obstruction labels benefit from concrete examples. The model needs calibration: what exactly counts as “oil and gas are needed for non-power uses,” or “domestic production benefits the country”? Apparently, even machines appreciate being shown the corporate euphemism once before being examined on it.

This distinction between evidence types matters because it prevents a lazy reading of the paper. The benchmark table is the core model evaluation. The ablations explain why the pipeline behaves as it does. The error analyses diagnose fragility. The temporal and company-level studies show a possible downstream use, not a finished compliance product.

The ablations show that context is not decorative

The ablation results are modest but revealing. For Qwen2.5-VL 32B, removing transcript input lowers performance from 66.2% to 61.2% on YouTube and from 70.5% to 60.6% on Facebook. That suggests transcripts matter, especially in the Facebook domain where the original labels are tied to ad text and climate-obstruction rhetoric.

But the paper also notes that transcripts do not always help smaller models on YouTube. This is a useful complication. More context is not automatically better if the model cannot integrate it well. In multimodal PR analysis, the problem is not merely “feed everything into the model.” The problem is aligning the right pieces of evidence: frames, spoken language, captions, metadata, entity history, and the annotation standard being applied.

The entity-aware retrieval results are also not clean enough for a victory lap. For Qwen2.5-VL 32B, combining transcript input, embedding search, and entity restriction gives the strongest reported ablation result. Yet across all models, entity restriction has mixed effects. The reason is intuitive: if an entity has too few training examples, restricting the candidate pool may force the model to learn from a less useful example. Similarity helps only when the search space contains something worth retrieving. A barren archive does not become strategic because we gave it embeddings.

The few-shot appendix points in the expected direction: increasing the number of examples improves performance for Qwen2.5-VL 7B, but at greater inference cost. For business use, that is the familiar trade-off. Better contextual calibration means more tokens, more latency, more compute, and more budgetary frowning from whoever owns procurement.

The hardest cases are where business risk actually lives

One of the paper’s most useful observations is that difficult labels are prone to over-labelling. For higher-performing labels such as Community and Life, Environment, and Work, precision tends to exceed recall. For harder labels such as Economy and Business, Green Innovation, and Patriotism, recall tends to exceed precision. In plain English: when the concept is subtle, models may start seeing it too often.

This matters because the subtle labels are also the ones most relevant to reputational interpretation. A company visibly showing workers is easy to classify and often harmless by itself. A company repeatedly blending innovation imagery with vague environmental language is more analytically interesting. But it is also harder to label reliably.

The model behaviour creates a governance problem. If an automated tool over-detects Green Innovation or Patriotism, it may generate noisy accusations. If it under-detects them, it misses sophisticated persuasion. Either error can damage the tool’s credibility. ESG analytics does not fail only when it is wrong; it fails when users no longer know what kind of wrong it tends to be.

The paper also reports model-specific labelling habits. Some models under-label; others over-label. DeepSeekVL2, for example, produces co-occurrence patterns that differ sharply from the gold labels, pairing Community and Life with Economy and Business far more often than the annotated dataset does, while missing the more common Community and Life plus Work pairing. This is a reminder that VLMs do not merely classify. They bring their own priors about what belongs together.

In corporate monitoring, those priors would become hidden policy. A model that systematically associates community imagery with economic messaging would create a different risk profile from one that associates community imagery with labour or daily life. The dashboard would look objective. Naturally.

The greenwashing pilot is a triage tool, not a verdict machine

The paper’s discussion moves from benchmarking to potential greenwashing detection. This is where the reader needs discipline.

The authors argue that Facebook labels such as “clean gas,” pragmatic necessity, community benefits, and non-power uses can correspond to forms of greenwashing or climate obstruction through selective disclosure. On YouTube, Environment and Green Innovation can involve vague environmental imagery or language without concrete commitments. Community, Work, and Economy labels can also support selective narratives about necessity and social value.

Then the paper runs two pilot analyses.

First, it examines temporal trends in YouTube environmental framing. The authors calculate the annual ratio of videos labelled Environment and compare GPT-4.1 predictions with gold labels. The predicted trend follows the ground-truth trend reasonably well, including an increase after 2020. The paper notes that this timing overlaps with the Biden administration but explicitly does not verify causality.

That caveat is doing real work. A rise in environmental framing after 2020 could reflect policy context, investor pressure, public climate salience, internal transition strategy, marketing fashion, or genuine investment. The benchmark can show that environmental messaging increased. It cannot, by itself, explain why.

Second, the authors conduct company-level profiling using anonymised Company X and Company Y. The model’s predicted label distribution roughly replicates the gold distribution. Company X over-indexes on Environment, suggesting an especially strong environmental image. Company Y emphasises Community and Life, often combining it with Work and Environment. In some multi-labelled videos, environmental content appears vague and embedded in the general positive impression rather than tied to concrete performance.

This is exactly the right level of ambition. The benchmark can reduce manual search costs. It can identify candidates for deeper review. It can help analysts ask sharper questions: Which companies increased environmental framing? Which firms combine community narratives with vague green language? Which campaigns shifted from jobs to innovation? Which entities use patriotic energy framing rarely, frequently, or only in particular markets?

What it cannot do is close the case. Greenwashing requires comparing the message with external facts: capital expenditure, emissions trajectories, lobbying behaviour, transition plans, project portfolios, legal disclosures, and policy engagement. The model watches the advertisement. It does not audit the balance sheet.

The business value is cheaper diagnosis, not automated moral judgement

For businesses, the paper’s practical path is straightforward.

Collect public corporate videos. Extract frames and transcripts. Classify recurring frames. Compare patterns across company, market, time period, and campaign. Use the output to prioritise manual ESG, legal, investor-relations, regulatory, or communications review.

That path can serve several groups, though not in the same way.

For ESG analysts and investors, the benchmark points toward narrative-risk screening. A company whose environmental framing sharply rises while transition performance remains weak deserves closer inspection. The model does not prove the mismatch; it helps find where to look.

For regulators and watchdogs, the value is triage. Public communications are too numerous for purely manual review. A framing detector can highlight campaigns heavy in vague environmental and innovation messaging, especially where claims are visual rather than textual. That matters because legal and compliance systems often chase explicit wording while marketing departments enjoy the richer diet of implication.

For corporate communications teams, the tool is a mirror. It can reveal how the company’s own messaging is likely to be perceived: community-first, work-first, environment-first, innovation-heavy, patriotic, or economically framed. That is useful even for firms trying to avoid greenwashing. Sometimes the risk is not a false claim but a pattern of emphasis that looks more virtuous than the underlying strategy.

For AI vendors, the paper is less flattering. It shows that generic VLM competence is not enough. A model may identify wind turbines and smiling workers while still failing to interpret what those cues are doing rhetorically. The product opportunity is not “AI that detects greenwashing.” The serious product opportunity is an auditable workflow for multimodal narrative analysis, with uncertainty, examples, human review, and external evidence integration.

The lesson is not that companies should stop showing trees. Though, honestly, some restraint would not be tragic. The lesson is that corporate narrative has become measurable enough to be governed, but not clean enough to be automated end-to-end.

Boundaries that matter before the dashboard arrives

The paper’s limitations are not generic academic modesty. They directly affect business interpretation.

The dataset is limited to Facebook and YouTube. It does not cover TikTok, YouTube Shorts, Facebook Reels, television, investor presentations, event sponsorships, outdoor advertising, or local-language campaigns. That matters because short-form video is often where emotional framing becomes most compressed and least explicit.

The dataset is biased toward large multinational companies, economically dominant regions, oil-producing countries, and videos understandable to English or Japanese speakers. Smaller firms, emerging markets, and non-English communication may behave differently. A model calibrated on this dataset should not be casually exported to every energy market and called global intelligence. That would be the usual enterprise software comedy, but with more carbon.

The Facebook labels are distant labels. They were originally based mainly on ad text and then mapped onto videos. The authors manually checked 20 randomly selected videos and found an 83% F-score against video-based annotation, suggesting reasonable quality, but the limitation remains. For Facebook, the benchmark partly tests whether video-aligned inputs can recover labels born from text-oriented analysis.

The YouTube labels are subjective and overlapping by design. The reported agreement is moderate, not perfect. This does not invalidate the dataset; it defines the problem. Impressionistic framing is inherently less crisp than object detection. Any business tool built on this kind of data must show examples, confidence, disagreement, and review pathways rather than pretending the label is divine revelation.

The experiments are also limited. The authors report single runs, do not perform hyperparameter search or model selection, and face computational constraints on frame sampling. Similar videos may appear across train and test sets, possibly making benchmark results optimistic. Transcripts are machine-generated and may include errors. These boundaries matter if someone wants to procure a system tomorrow and call it “AI-powered greenwashing detection.” A more accurate label would be “AI-assisted framing analysis with several ways to embarrass you if unsupervised.”

The real advance is making soft power inspectable

The most important idea in the paper is not the best F-score. It is the shift from analysing isolated environmental claims to analysing multimodal corporate persuasion.

Greenwashing is often discussed as if it were a fact-checking problem: find the claim, verify the claim, punish the claim. That works for some cases. But modern corporate reputation management is frequently softer. It builds associations. It places environmental virtue beside industrial necessity, community welfare, economic contribution, and national pride. It makes the company feel like infrastructure for civilisation rather than a participant in climate risk.

This paper does not solve that problem. It makes the problem more inspectable.

That is a useful kind of progress. It gives analysts a vocabulary for recurring frames, a dataset for benchmarking models, evidence that current VLMs can recognise some patterns, and a sober view of where they fail. It also offers a practical route from video classification to temporal and company-level profiling, while leaving the final judgement where it belongs: with humans comparing communication against reality.

The misconception to avoid is simple. This is not an AI lie detector for oil and gas companies. It is closer to an instrument panel for corporate atmosphere. It measures which narrative gases are being released into the public sphere, in what combination, and how often. The emissions metaphor is almost too easy, but occasionally the universe writes the joke itself.

For business, the implication is equally simple. The next generation of ESG intelligence will not be built only from reports, filings, and numerical disclosures. It will also need to read images, video, tone, repetition, and omission. Public trust is shaped there. So is reputational risk.

The paper shows that AI is beginning to see those patterns. It also shows that seeing is not yet understanding, and understanding is not yet judgement. A sensible organisation will use this kind of system to narrow the search, sharpen the questions, and document the narrative. It will not outsource moral reasoning to a model that still confuses a lab coat with a transition plan.

Cognaptus: Automate the Present, Incubate the Future.