By Alexander Rink and Kat McNeill

Those of us who work in climate have a complicated relationship with artificial intelligence (AI). On the one hand, it can accelerate scientific research, improve environmental monitoring, reduce administrative work, and help organizations interpret increasingly large and complex datasets. On the other hand, the rapid growth of AI is dramatically increasing demand for electricity, data centres, infrastructure, and cooling, placing greater pressure on power systems and corporate climate commitments.

Paradoxically, AI may therefore both increase the need for carbon dioxide removal (CDR) and improve and accelerate our ability to deliver it.

That tension shaped two campfire sessions hosted by CDR.fyi at Carbon Unbound East Coast on May 19-20 in New York City. The sessions brought together carbon removal suppliers, researchers, technology companies, buyers, market participants, and students to explore a practical question: How can we use AI to accelerate CDR while managing the risks it creates?

AI is neither inherently positive nor an unavoidable threat. However, it is becoming increasingly integrated into how organizations work. General-purpose tools such as Claude, ChatGPT, Gemini, and Copilot are being used alongside specialized systems for scientific analysis, environmental monitoring, project siting, and operational optimization.

AI is not going away. As such, the more useful question is whether the CDR sector can apply it in ways that improve speed, quality, transparency, and trust, and ultimately, accelerate CDR.

Highlights

• AI is currently most valuable as an accelerator of human work, rather than as a replacement for scientific, technical, or commercial judgment, making domain expertise even more important in AI-enabled workflows.
• Specialized systems based on well-defined datasets, standards, and processes may offer greater value than general-purpose chatbots, since the latter can produce overly generic or factually incorrect outputs.
• Registry, validation, and verification workflows are among the clearest near-term opportunities for AI to reduce CDR delivery risk and accelerate cash flow cycles.
• AI can shorten the time required to generate insights, but physical testing, source verification, and expert review remain essential, as was done for this blog post.
• CDR organizations can benefit from adopting AI deliberately, with clear accountability for its outputs and a focus on measurable outcomes.

How AI Is Accelerating Both CDR Demand and CDR Delivery

The climate implications of AI are not one-sided. Growing adoption is contributing to rising electricity demand, particularly from data-centres. This raises questions about whether technology companies can meet their climate commitments and how the additional load will affect emissions, grids, and energy markets.

Participants at the campfire sessions also noted that data-centre growth may increase demand for carbon credits and durable carbon removal. More electricity consumption, greater pressure on corporate climate targets, and closer public scrutiny could all lead companies to seek additional ways to address residual emissions. In some compliance markets, data-centre growth may also contribute to greater demand for allowances and credits.

This does not mean that increased emissions from applying AI are acceptable, just because they may stimulate CDR demand. Avoiding and reducing emissions remain the priority, and carbon removal cannot become a justification for unchecked energy use, as increasing emissions - even from renewable energy sources - would likely render CDR’s benefits pointless. Nonetheless, the CDR sector would be advised to prepare for a world in which AI expands rapidly, electricity demand rises, and the need for credible climate solutions becomes even more important.

One participant summarized the overall dynamic with a single word: acceleration. AI is accelerating demand, competition, scientific capabilities, and operational change simultaneously. The opportunity for climate and CDR organizations is to use that same force to move credible solutions forward more quickly.

AI Is Most Useful as a Personal Assistant

Across both sessions, participants consistently described AI as most effective when it supports capable people rather than replaces them. Some of the use cases raised included summarizing scientific literature, preparing first drafts of contracts and technical documents, identifying potential project sites, supporting project finance analysis, synthesizing remote sensing and LIDAR data, analyzing commodity pricing, and assisting with experimental design.

These applications remove a significant cognitive burden. Work that once required several hours of searching, organizing, or drafting can often be completed much more quickly, allowing teams to spend more time on interpretation, judgment, relationships, and execution.

One CDR supplier, for example, has provided organization-wide access to Claude and is encouraging cross-functional experimentation. Its team is exploring applications in engineering, procurement, construction, project siting, and project finance. Another is using experienced team members to guide internal AI tools in processing geographically variable CO₂ pricing data and supporting sales activities.

Other participants discussed AI for legal review, market research, environmental monitoring, and communications. In many of these cases, AI is useful as a first pass. It can organize information, identify likely issues, and create a starting point for an experienced person to assess and improve.

The limitations are equally important. Several participants described senior team members spending time correcting inaccurate, generic, or poorly reasoned outputs. General-purpose models may misunderstand technical requirements, leave out important context, invent information, or present uncertain conclusions with too much confidence. One participant observed that AI often appears inclined toward optimistic conclusions, even when the evidence is unclear.

The discussion therefore surfaced an important distinction: AI can reduce time-to-insight, but it does not automatically reduce time-to-truth. Faster analysis is valuable only when users can distinguish a plausible answer from a reliable one.

Why Domain Expertise Matters More When You Use AI

AI tools make sophisticated capabilities available to more people. They help small companies perform research, analysis, drafting, and data processing that might previously have required a much larger team. That can be especially valuable in CDR, where many organizations are capital-constrained startups that are working to accomplish more with limited resources.

At the same time, AI makes it easier to produce work that appears authoritative while containing subtle technical errors. Experienced scientists, engineers, operators, lawyers, and carbon-market practitioners can often recognize where an answer is incomplete or incorrect. Junior staff may not yet have enough context to identify those weaknesses.

A project developer described how general-purpose models may lack the specific context needed to interpret carbon market standards and documentation. Experienced users may compensate by writing detailed prompts and checking each conclusion against their own knowledge. Less experienced users, by contrast, may accept the same output without recognizing what has been omitted.

Effective AI use, therefore, requires two forms of fluency. The first is the ability to work with the tool, including asking precise questions, providing relevant context, and iterating on the result. The second is the subject-matter expertise required to assess whether the answer is accurate and useful.

Broad access to AI does not reduce the value of experienced review. In technical settings, it may increase it, particularly where an output affects scientific claims, credit issuance, contracts, financing, safety, or project performance.

A practical operating principle emerged from the sessions: Everyone may use AI, but someone remains accountable for the result.

Purpose-Built Systems Create More Value Than General Models

While general-purpose tools are useful for brainstorming, drafting, and synthesis, many participants believed that some of the greatest value for CDR comes from more specialized or constrained systems.

Carbon removal projects operate within specific scientific, technical, contractual, and regulatory contexts. A model that understands language broadly may still fail to interpret a registry methodology, identify a missing piece of MRV evidence, or distinguish between similar but materially different carbon-accounting concepts. In these situations, the value often comes from narrowing the context rather than expanding it.

Participants described potential systems grounded in registry methodologies, validation and verification standards, project records, environmental datasets, scientific literature, remote sensing information, and project finance assumptions. Rather than allowing a model to answer from the full range of its general training, these systems can be limited to approved materials and required to identify the sources supporting their conclusions.

This is the thinking behind CDR.fyi’s AI Data Scientist, for example. The objective is not to create a general chatbot for carbon removal. It is to help users ask better questions of CDR.fyi’s market data, explore trends more efficiently, and reduce the friction involved in turning structured data into useful insight. As with other AI applications in CDR, the system is most valuable when it combines accessible analysis with clear source data, defined boundaries, and human judgment.

This closed-context approach can reduce hallucinations, make outputs more traceable, and provide greater consistency. The most useful system may not be the one that appears to know everything; rather, it may be the one that knows exactly which information is permitted for use, clearly identifies uncertainty, and declines to answer when the evidence is insufficient.

How AI Can Reduce Validation and Verification Delays in CDR

One of the clearest near-term opportunities raised during the sessions was the use of AI to improve registry, validation, and verification workflows.

Suppliers frequently experience substantial back-and-forth when submitting project documentation. Issues that could potentially have been identified before submission may emerge only after formal review begins. Project developers then revise documents, respond to issue logs, gather further evidence, and wait through additional review cycles.

These delays create more than administrative inconvenience. Suppliers may have delivery obligations and contractual penalties, while the timing of registry or Validation and Verification Body (VVB) approvals remains partly outside their control. Delays in the approval process pose a direct risk to supplier delivery and, hence, cash flow.

Participants proposed an AI-assisted pre-review system grounded in the relevant methodology, registry rules, and VVB requirements. A project developer could upload its submission materials and receive structured feedback before beginning formal review. The system could flag missing documents, inconsistent information, methodology gaps, unclear evidence, or issues likely to generate questions from reviewers.

Such a tool could also prepare a preliminary issues log or structured review package for the VVB. It would not replace human validation or verification. Instead, it could improve the quality of initial submissions and allow expert reviewers to focus their attention on questions that genuinely require scientific or professional judgment.

The goal is not to reduce rigour in pursuit of speed. It is to preserve or improve rigour while reducing avoidable repetition and delay. For suppliers, this could shorten approval cycles and reduce financing and delivery risk. For registries and VVBs, it could improve reviewer productivity. For buyers, it could contribute to greater confidence in expected issuance and delivery timelines.

Among the opportunities discussed, this may be one of the strongest candidates for shared market infrastructure rather than a tool that each supplier develops independently.

AI for Science and the Physical World

The campfire discussions also made clear that AI in CDR extends far beyond language models.

Participants described applications in bioacoustics, wastewater treatment, methane detection, algal-bloom prediction, materials discovery, remote sensing, and industrial optimization. These systems may use machine learning to detect patterns in sound, images, sensor readings, or operational data rather than to generate text.

One sustainability student discussed using bioacoustics to assess whether biochar contributes to ecological improvements in agricultural fields. Audio devices record changes in the presence of birds, insects, and other organisms, while AI models help interpret the resulting soundscapes. This offers a novel source of information about biodiversity and ecosystem health.

A participant from a CDR purchaser described using AI to identify potential carbon-negative building materials. AI can help generate candidates, explore combinations, and narrow the field of possibilities. The materials still need to be created and tested, however, and the same principle applies across biology, medicine, architecture, and climate technology.

AI can help design an experiment, identify a pattern, or generate a hypothesis, but it cannot remove the need to test what happens in the physical world. A model may produce a compelling prediction. Nature still determines whether it is correct.

This distinction is especially important in CDR, where measured physical outcomes ultimately matter more than the sophistication of the model used to predict them.

Data Quality Remains the Foundation

Every AI system depends on the quality of the information available to it. Incomplete, inconsistent, poorly structured, or biased data will constrain even the most capable model.

This is particularly relevant in carbon removal, where information is distributed across suppliers, registries, marketplaces, buyers, researchers, standards bodies, and public institutions. Terminology is not always consistent, documentation is often designed for human rather than machine use, and important context may be contained in PDFs, spreadsheets, emails, or internal systems.

Many promising AI applications will therefore depend on better data infrastructure. Methodologies and standards can become more machine-readable. Project and transaction information can use clearer definitions and structures. Source records, version histories, and confidentiality controls can be maintained more consistently. Systems can also become more interoperable.

This does not mean all information belongs in the public domain. Commercial confidentiality, contractual rights, data ownership, and security remain legitimate considerations. It does mean that organizations can distinguish between information that genuinely needs to remain confidential and information that is difficult to access simply because it has not been standardized or structured.

Better underlying data will improve more than just AI. It will also strengthen reporting, benchmarking, market analysis, and trust across the CDR sector.

Information Access, Curation, and Bias

Participants also discussed the role of AI in summarizing news, research, and market developments. One student noted the value of having more consolidated CDR news and media coverage, while also raising the question of what biases an AI-curated system might introduce.

The benefit of AI-supported synthesis is clear. The volume of climate and CDR information is growing, and few people can read every new paper, policy document, transaction announcement, or project update. AI can help users navigate that information and make technical material more accessible to broader audiences.

Curation, however, is never entirely neutral. An AI system makes choices about what to include, exclude, prioritize, and summarize. Those choices may reflect gaps in source coverage, the structure of the underlying data, commercial incentives, or the wording of the prompt.

Organizations using AI for market intelligence or external communications can preserve source traceability, enabling users to distinguish among original reporting, human interpretation, and AI-generated synthesis. Trust will depend not only on whether an answer sounds credible, but also on whether its evidentiary basis can be examined.

Climate Impact Is Not Yet Driving AI Procurement

The sessions also explored whether climate-focused organizations choose AI tools based on their environmental impact. Most participants indicated that functionality, security, ease of integration, and compatibility with existing systems currently matter more.

Part of the issue is that organizations do not yet have enough comparable information. It remains difficult to assess the energy use, emissions profile, or climate performance of one AI product against another. As a result, even companies that would like to include climate considerations in procurement decisions may be unable to do so meaningfully.

Participants also noted a broader issue around transparency and accountability. Technology companies that disclose their emissions and acknowledge the tension between AI growth and climate goals often attract more scrutiny than companies that disclose little or make no meaningful commitments.

The climate community can continue to hold companies accountable while also recognizing the value of transparency. Credible disclosure and measurable progress deserve acknowledgement, even when performance remains imperfect. Better environmental information about AI products would allow companies to consider climate impact alongside functionality, security, and cost.

What CDR Organizations Can Do Now

The campfire sessions did not produce a single blueprint for AI adoption, nor would one be useful across such a varied sector. Different organizations face different risks, data environments, and operational requirements. Even so, several practical principles emerged.

1. Begin with a defined workflow rather than with the technology itself. Organizations can identify processes that are repetitive, slow, information-heavy, or constrained by limited staff capacity, and then assess whether AI can improve outcomes. The objective is not simply to adopt AI. It is to solve a real problem.
2. Begin where errors are reversible. Research summaries, first drafts, internal analyses, document checks, and preliminary modelling are generally safer starting points than autonomous scientific, operational, legal, or crediting decisions.
3. Match oversight to consequence. A draft social media post does not require the same level of control as an MRV calculation or a contractual interpretation. The greater the scientific, financial, legal, safety, or reputational consequence, the stronger the review.
4. Invest in a domain-specific context. Providing models with approved documents, internal data, methodologies, and standards is likely to produce more useful results than relying entirely on general-purpose knowledge.
5. Train teams in both tool use and professional judgment. Learning how to prompt an AI system is valuable, but learning when not to trust its answer is even more important.
6. Measure whether the technology is creating value. Time saved, errors identified, cycle times shortened, review burden reduced, and commercial outcomes improved are more meaningful than the number of employees using an AI tool.

Accelerating with Care

AI will not solve the fundamental challenges facing carbon removal. Projects still require financing, permits, infrastructure, energy, biomass, minerals, and customers. Facilities still need to be constructed and operated. Scientific claims still require evidence, and carbon still needs to be measured, verified, delivered, and stored.

AI can, however, help the sector work through many of these challenges more quickly. It can reduce administrative burden, improve access to knowledge, analyze environmental data, support scientific discovery, strengthen project development, and shorten market processes.

It can also produce convincing errors, obscure accountability, amplify bias, increase electricity demand, and create misplaced confidence.

The most constructive response is neither unquestioning adoption nor avoidance. It is responsible acceleration: using AI where it improves outcomes, placing stronger controls around higher-risk applications, and maintaining human accountability throughout.

AI is already accelerating the world around us. The opportunity for the CDR sector is to ensure that it also accelerates credible climate progress.

FAQS

How is AI currently being used in carbon removal?

Primarily as an accelerator of human work: summarizing literature, drafting documents, analyzing remote-sensing data, supporting project siting and project finance, and streamlining validation workflows.

Are general-purpose AI models like ChatGPT or Claude useful for CDR workflows?

For drafting, synthesis, and research they can reduce time significantly, but they may misinterpret registry methodologies or carbon accounting standards. Purpose-built systems grounded in specific datasets and methodologies are likely to produce more reliable outputs for technical CDR use cases.

Can AI replace human judgment in MRV and verification?

No. AI can flag missing documents, inconsistencies, and likely issues before formal review, but scientific, professional, and legal judgment remains essential — especially where outputs affect credit issuance, contracts, or financing.

What is the biggest near-term AI opportunity in carbon removal?

Participants at CDR.fyi's Carbon Unbound East Coast sessions identified AI-assisted pre-review of registry and VVB submissions as one of the clearest near-term opportunities to reduce delivery risk for suppliers.

Does AI increase carbon emissions?

Growing AI adoption is increasing data-centre electricity demand and placing pressure on corporate climate commitments, which may simultaneously increase the need for durable carbon removal and create new incentives for CDR procurement.