Greenwashing and bothsidesism in AI chatbot answers about fossil fuels' role in climate change

Generative AI tools are increasingly part of the way we search for information – from search engines providing “AI overviews” to powering chatbots in instant messenger services.  

As many people have adopted using generative AI to find information, we wanted to see how reliable current AI chatbots are at providing users with accurate answers relating to climate change.  

We investigated this around the COP29 summit, at a time when access to reliable climate information is crucial and people may be more likely to seek out information on climate policy and COP29 itself.  

We asked four chatbots – OpenAI's ChatGPT, Meta’s MetaAI, X’s Grok and Google’s Gemini – a series of questions about climate change, oil and gas companies, and COP to see what kind of information they would provide.  

Fossil fuel companies’ complicity in climate change is well established, and has been exposed in multiple Global Witness investigations in 2024.  

Oil, gas and coal are responsible for 90% of carbon dioxide emissions, which are heating the planet and super-charging extreme weather events. 

Just five of the world’s biggest fossil fuel firms, BP, Chevron, Exxon, Shell and TotalEnergies, are set to spend $15 million every single hour between now and 2030 producing more oil and gas. 

Big Oil has also lobbied against climate laws, and many firms have ignored their own scientists – all while the planet suffers. 

And fossil fuel companies that have sent lobbyists to COP summits have also produced enough oil and gas since the 2015 Paris Agreement to cause 17 million heat deaths this century. 

Given the overwhelming failure of the fossil fuel industry to take serious action to protect our climate, we should expect generative AI tools to be able to communicate their complicity.  

The good news is that all the chatbots we tested demonstrated an ability to answer questions in detail, and explicitly cite/link to a variety of sources.  

However, we also saw evidence of greenwashing – for example, uncritically positive narratives about oil and gas companies and false equivalence, which present narratives about oil and gas companies’ efforts to protect the climate with the same weight as criticisms of their contribution to climate change. 

If generative AI tools are to position themselves as useful for accessing information, it is essential that they are designed and safeguarded in ways that prevent their capture by fossil fuel companies’ own misleading narratives. 

When asked generally about oil and gas companies, Meta AI acknowledges the limitations of their climate action.  

However, when we asked questions that specifically referenced oil and gas companies by name (BP, ExxonMobil, Chevron, TotalEnergies and Shell), Meta AI produced answers that included greenwashing. 

When asked about the role of BP at COP, Meta AI  – despite mentioning “questions about the industry's commitment to meaningful climate action” – concluded that “their climate efforts suggest they'll be an active participant in the conference, promoting sustainable energy solutions and advocating for climate action.”  

When asked about the 2023 emissions of fossil fuel companies, Meta AI said of each company that their emissions figures showed their progress, efforts or their commitment to sustainability, including where the data wasn’t available or the chatbot itself reported that some kinds of emissions had increased. 

“BP's carbon emissions in 2023 were as follows: 

• Operational Emissions: 32.1 million tons of CO2 equivalent (MtCO2e), which is a 41% decrease from the 2019 baseline of 54.5 MtCO2e. 
• Upstream Oil and Gas Production Emissions: 315 MtCO2, a slight increase from 307 MtCO2 in 2022, mainly due to increased production from major projects and higher asset performance. 
• Methane Emissions: 31,000 tons, with a methane intensity of 0.05%. 
• Carbon Intensity of Sold Energy Products: 77gCO2e/MJ, representing a 3% decrease from the 2019 baseline. 

These figures indicate BP's progress towards its net-zero ambitions, including reducing operational emissions, methane intensity, and carbon intensity of sold energy products.” 

(Note: Global Witness has not verified the accuracy of these emissions statistics)

And when asked if specific fossil fuel companies were helping the world reach net zero, Meta AI shared criticisms of Shell, Exxon, Chevron and TotalEnergies, but did not have anything negative to say about BP’s role in helping the world reach net zero. 

MetaAI was far more sceptical about Shell’s net zero ambitions than BP’s pledge.  

These responses are despite the fact that big fossil fuel companies have played an outsized role in unleashing climate chaos, and these five companies in particular are set to use one-eighth of the world’s carbon budget by 2050.  

Our investigation found that where generative AI models are used as sources of reliable information on climate (even when multiple sources are relied upon to produce an answer), there is a risk of “bothsidesism", where both sides of an argument are presented as carrying equal weight with “no easy answer,” despite the overwhelming evidence of fossil fuel companies’ complicity in climate change. 

There was acknowledgment of fossil fuel complicity in climate chaos. For instance, Grok, which in “Fun" mode declared itself to be giving answers “wrapped in a cloak of humor and a bit of skepticism,” was upfront about its critiques of fossil fuel companies.  

It also offered the most critical answers about the role of these companies at COP, including that these companies engage in lobbying, PR, running events and networking, as opposed to generic answers about the companies engaging in climate discussions.  

Meta AI also acknowledged in some of its responses (such as to the question “Are oil and gas companies going to COP?”) that there were significant concerns over the sincerity of fossil fuel companies’ engagement in COP.  

“Key Concerns: 

• Fossil fuel lobby influence: The presence of oil and gas companies may hinder climate progress. 
• Conflict of interest: Fossil fuel lobbyists may prioritize business interests over climate action. 
• Transparency: The lack of clear guidelines on lobbyist participation raises concerns.” 

Similarly, ChatGPT and Gemini both offered answers that included references to controversies and criticisms of fossil fuel companies and their greenwashing. However, these responses were tempered.  

Although Gemini produced analysis of both the sustainability efforts and failures of fossil fuel companies, it concluded that whether each company was helping the world to reach net zero was “a complex question” without an easy answer. 

When asked about their role at COP, Gemini responded for each company that “[company], like other oil and gas companies, attends COP to discuss its own climate plans, lobby for policies that benefit its industry, and participate in discussions about the energy transition.” 

Similarly, ChatGPT described the extent to which these companies are helping the world to reach net zero as “a subject of debate.” 

When asked about the companies’ role at COP, ChatGPT gave the same response for every company we asked about, with the same positives and negatives for each company, and concluding that they continued to be the subject of “scrutiny and debate.” 

When asked about Shell’s emissions in 2023, ChatGPT also presented us with two possible answers, one of which contained more evaluation of how the figures showed “Shell's ongoing efforts to lower its carbon footprint across its operations and product offerings.” 

While the presentation of different perspectives on generative AI tools is valuable, there is a risk that in an effort to achieve "balance", the underlying complicity of fossil fuel companies in climate breakdown is being masked from readers. 

These chatbots do all issue warnings that they may produce responses that are inaccurate in some way.  

The problem is that as these tools become more ubiquitous, and are marketed to users as useful ways to locate information, their role is becoming more and more akin to search engines, even though their answers are generated in a technically different way.  

When asked questions about specific companies, some of the AI models drew directly from the company websites for data (as we saw evidence of, where models cited them), meaning that there is a direct route for companies – intentionally or not – to influence answers about them. 

The models we tested did try to compensate for this by visibly drawing from a variety of sources that would ensure perspectives beyond only the fossil fuel companies were represented (by adding labels or providing links to alternative sources used). 

Grok is also able to pull in live content from X, though this comes with its own information risks, as we saw in our investigation earlier this year that found Grok was sharing conspiracy theories.  

Generative AI tools are increasingly likely to fall within the scope of digital regulation, such as the Digital Services Act, the incoming EU AI Act or the UK Online Safety Act.  

But in the US, there is a risk that AI deregulation may increase, with Donald Trump having repealed Joe Biden's executive order on AI.

The next few years will be significant in determining the future of generative AI development and whether it can promote access to reliable information or risks undermining the health of our information environments.  

Although there have been significant improvements in transparency – with models sharing and citing sources and providing links to sites where users can check information and read for themselves how these models weigh and assess sources (not only for relevance but also credibility) – is still opaque.  

We call on companies that are producing generative AI tools to publish details of the steps they take in development to assess sources for credibility. This process must avoid spreading climate disinformation while ensuring access to a diversity of sources. 

We are also calling on companies to not integrate generative AI tools into their online services without conducting and publishing an information audit of how those tools will affect users’ access to reliable climate information – and mitigating any arising risks.  

We contacted Google, OpenAI, Meta and X for comment on this investigation. Google, Meta and X did not provide a comment. 

OpenAI told us that "being objective," where a model should inform not influence, is a key principle of how they train their models. However, they said that this principle is one of the hardest to implement and they are aiming to make continual improvements. 

They said that the intelligence of their models meant that a model can make determinations about things like which sources it’s looking at, how many and how confirmed it is or not. 

They told us that the aim is to provide the user the capability the user has asked for.  

• We took steps to reduce the risks that the models would be influenced by our activity histories, such as using clean browsers and clean accounts where possible. 
• We asked the models a series of questions including factual questions (asking for information), evaluative questions (asking for opinions or judgements) and imaginative questions (asking for help creating content). The topics included climate change, COP29, oil and gas companies and climate policies. 
• The models we tested were: 
• ChatGPT4o and Dall-E, accessible via subscription to ChatGPTPlus 
• Grok 2.0, Fun Mode, accessible via Premium X Subscription 
• Gemini, accessible with a Google account  
• MetaAI, accessed via Facebook in Australia (due to geographic restrictions on the model availability globally)