[EDRM Editor’s Note: The opinions and positions are those of John Tredennick and Dr. William Webber.]
Over the past few months, we’ve been exploring how generative AI can transform trial preparation by analyzing complex litigation materials and producing sophisticated closing arguments. Our series began with an exploration of how to overcome traditional length limitations in Large Language Models, then demonstrated these techniques by creating closing arguments for TransOcean, Halliburton, and the government plaintiffs in the landmark Deepwater Horizon (Gulf Oil Spill) trial.
Each article in this series has aimed to show even the most skeptical legal professionals that GenAI technology isn’t just coming—it’s here, and it’s capable of handling complex legal tasks with remarkable efficiency. But we deliberately saved the most challenging assignment for last: creating a closing argument for BP itself.
The Ultimate Challenge: Defending the Indefensible?
For those unfamiliar with the BP Deepwater Horizon disaster, it represents one of the most catastrophic environmental events in history. On April 20, 2010, an explosion on the Deepwater Horizon drilling rig killed 11 workers and triggered an oil spill that released approximately 4.9 million barrels of crude oil into the Gulf of Mexico over 87 days.
The Phase One trial focused on liability, with U.S. District Judge Carl Barbier ultimately ruling that BP was guilty of gross negligence and willful misconduct, describing their actions as “reckless.” The court apportioned 67% of blame to BP, 30% to TransOcean, and 3% to Halliburton.
Most legal observers viewed BP’s position as nearly indefensible. The evidence against them was overwhelming, with internal communications showing BP officials were aware of risks but proceeded recklessly. The company faced billions in fines and settlements.
Creating a closing argument for BP represented our ultimate challenge: Could GenAI analyze the same trial materials we used for the other parties and construct a legitimate defense strategy for the party most observers considered primarily responsible? And could it do so while acknowledging the gravity of the disaster without minimizing BP’s role in safety?
John Tredennick and Dr. William Webber, Merlin Search Technologies.
Creating a closing argument for BP represented our ultimate challenge: Could GenAI analyze the same trial materials we used for the other parties and construct a legitimate defense strategy for the party most observers considered primarily responsible? And could it do so while acknowledging the gravity of the disaster without minimizing BP’s role in safety?
Bringing in the Latest AI Technology
For this final challenge, I was particularly excited to leverage Claude 3.7 Sonnet, Anthropic’s most intelligent model to date, released just days before undertaking this project. As I noted in my recent article on this breakthrough technology, Claude 3.7 Sonnet represents a significant advancement as the market’s first hybrid reasoning model, offering unprecedented capabilities for complex legal analysis and synthesis.
Having extensively tested Claude 3.7 Sonnet, our team confirmed it’s the most powerful model available for legal work, particularly for analyzing complex legal documents, synthesizing information across multiple sources, and constructing sophisticated legal arguments. Its enhanced reasoning capabilities make it uniquely suited for tackling the nuanced challenge of defending BP in this high-stakes case.
With over 20 years as a trial lawyer under my belt, I approached this exercise with genuine curiosity—what would I do if BP were my client? How would I structure a closing argument that both acknowledged the tragedy while still presenting the strongest possible defense?
Using our Merlin Alchemy platform (formerly known as DiscoveryPartner), we deployed the same methodology used in previous articles, but this time leveraged both Claude Haiku 3.5 and Sonnet 3.7 to tackle different aspects of the challenge. This multi-model approach optimizes both performance and cost-efficiency, using Haiku for initial document processing and summary generation, while engaging Sonnet 3.7’s superior reasoning capabilities for the complex analytical work required to construct a compelling defense.
The system processed almost 10,000 pages of trial materials in minutes, identified the most relevant evidence and testimony, and constructed a defense strategy that would typically require a team of senior attorneys days or weeks to develop.
John Tredennick and Dr. William Webber, Merlin Search Technologies.
The system processed almost 10,000 pages of trial materials in minutes, identified the most relevant evidence and testimony, and constructed a defense strategy that would typically require a team of senior attorneys days or weeks to develop. What’s particularly impressive is how Claude 3.7 Sonnet was able to navigate the complexity of acknowledging BP’s responsibilities while still constructing legitimate legal arguments based on the shared responsibility model in deepwater drilling operations.
The BP Defense Strategy: An Outline
The first step was creating a comprehensive outline for BP’s closing argument. I asked Alchemy to create the outline using this prompt:
I want you to act as the lead trial lawyer for British Petroleum in the Deepwater Horizon Oil Spill Phase One Liability Trial. You have years of experience defending large corporations in complex disasters. I need you to review the trial testimony and trial exhibits and think hard on a strategy to defend BP as effectively as positive.
Read all of the materials and develop an outline for the most effective closing argument. Your goal will be to push responsibility for this terrible tragedy to the other defendants and to show that BP should not be held liable if at all possible. Tell the story that BP was not the operator of the well and wasn’t responsible for the key elements that failed here.
We will take your outline and then build our closing argument.
The AI produced a detailed framework that organized BP’s defense into seven major sections, starting with an acknowledgment of the tragedy while establishing the complex nature of deepwater drilling operations. The outline then methodically built a defense based on the shared responsibility model in the industry, technical analysis of key decisions, alternative theories of causation, and legal standards for liability. Here is a short view of the outline:
Outline Table of Contents
The Prompt
Outline of BP’s Closing Argument
- Introduction: The Tragedy in Context
- Acknowledge the gravity of the disaster and loss of life
- Preview key argument: BP was not the operator of the well
- BP Was Not the Direct Operator of the Well
- Contractual Delineation of Responsibilities
- Operational Control in Practice
- Technical Failures Beyond BP’s Direct Control
- BOP System Design and Maintenance Issues
- Transocean’s Operational Philosophy and Safety Culture
- Cementing Failures by Halliburton
- Critical Decision-Making and Operational Failures
- Negative Pressure Test Misinterpretation
- Transocean’s Failure to Monitor and Respond to Well Control Issues
- Simultaneous Non-Standard Operations
- BP’s Safety Culture and Risk Management
- BP’s Commitment to Safety
- BP’s Reliance on Contractor Expertise
- Shared Responsibility Among Multiple Parties
- Multiple Contributing Factor
- Industry Context and Inherent Risks
- Conclusion: A Fair Assessment of Responsibility
- Summary of key arguments
- Call for fair allocation of liability
1
3
3
3
3
3
3
3
4
4
5
6
6
6
7
8
8
8
9
9
9
10
10
10
10
The outline itself runs about nine pages, which you can view in its entirety here.
The Closing: A Remarkable First Draft
We then took the outline and asked Alchemy to generate the draft closing argument. We did this in three groupings: Arguments 1-3, Arguments 4-5 and Arguments 6-7. In each case we developed a prompt that included instructions and the relevant sections of the outline itself.
Here were the instructions:
You are the lead trial lawyer for British Petroleum (“BP”) in the Deepwater Horizon Oil Spill Phase One Liability Trial. You have years of experience defending large corporations in complex disasters and are known for eloquent and persuasive arguments to the court.
You are preparing your closing argument for the end of trial based on the testimony and exhibits presented in the trial. I need you to review the trial testimony and trial exhibits and develop the best closing argument possible to defend BP.
Here is the outline for a portion of your closing argument. Your argument should follow the outline which you prepared earlier. This will only be a portion of your entire closing. I will combine the different argument sections after you have finished all portions of your argument.
[Insert relevant sections of the outline]
Make your argument in narrative form as if you were giving it to the Judge in this case. Include citations to testimony and exhibits that support your arguments. Your goal will be to push responsibility for this terrible tragedy to the other defendants and to show that BP should not be held liable if at all possible.
We ran each section prompt against over 9,000 sections of testimony and exhibit text, using a trio of algorithmic search methods to find the most relevant sections of the Phase One trial for our argument:
- Natural Language: A semantic, vector-based search for contextual meaning;
- Keyword algorithmic: to ensure that appropriate keyword text is considered; and
- Machine Learning: A machine learning classifier used based on initial seed rankings.
Each of the engines ranked the candidates for relevance. We merged the rankings using heuristics in order to surface what our research shows are the most likely document sections to be helpful to our information needs.
For our experiment, we stopped the enquiry at the top 500 consolidated records found by our three search engines. We easily could have gone deeper but we have found that the analysis of the top 500 records for the different topic inquiries would provide a good overview of the power these LLMs have to create a persuasive closing argument for purposes of this article.
The BP Closing Argument: A Technological Breakthrough in Legal Work
In response, Alchemy developed a comprehensive defense that methodically addressed each element of BP’s position while acknowledging the tragedy. The argument synthesized information across multiple sources to create a cohesive, strategic defense that emphasized the shared responsibility model in deepwater drilling operations.
Here is a section of the argument to give you an idea of the depth and breadth of the underlying reasoning and the LLM’s ability to identify and synthesize a complex fact structure.
V. BP’s Safety Culture and Risk Management
A. BP’s Commitment to Safety
1. Safety Management Systems
Your Honor, despite the tragic outcome, BP had established robust safety management systems. BP’s Operating Management System (OMS) was intended to ensure safe and responsible drilling operations (BP_00000002 §12 ¶116). BP’s written practices emphasized safety as the top priority, with time to production as the lowest concern.
The Deepwater Horizon was characterized as having a strong safety culture that prioritized safety. Patrick O’Bryan, BP’s Vice President of Drilling, described the rig as the best-performing in BP’s fleet for drilling performance and safety. Rig crew members emphasized that “safety was paramount” with strictly enforced safety policies.
BP maintained extensive communication and teamwork between BP and Transocean personnel (BP_00000002 §12 ¶120-121). Comprehensive communication processes included pre-spud meetings, well adviser programs, and daily planners.
The 2008 safety pulse check showed 100% positive responses about job planning and safety concerns (BP_00000002 §12 ¶119). Most importantly, BP and Transocean had zero red zone kicks between 2005-2009 (BP_00000002 §12 ¶124) and had more precautionary shut-ins than other major oil companies like Shell and Chevron.
2. Risk Assessment and Management
BP conducted comprehensive risk assessments through gate meetings (BP_00000002 §13 ¶125-126). BP’s well design involved a comprehensive five-month planning process. Each gate meeting included drilling surprise assessments, risk register updates, and conceptual design reviews.
BP’s operations incorporated continuous data gathering via PWD and high-quality logging (BP_00000045 §26 ¶262-269). BP’s Drilling Work Order Procedure (DWOP) establishes clear safety priorities:
- People first
- Protect the environment
- Protect installation
- Economic considerations (last priority)
The DWOP mandated that risks should be managed “to a level which is as low as reasonably practical” (BP_00000083 §28 ¶290). BP performed all required operational tests on rig samples (BP_00000002 §9 ¶85-90). Industry practice requires using actual rig samples for operational testing, not laboratory samples.
B. BP’s Reliance on Contractor Expertise
1. Industry Standard Practices
Your Honor, deepwater drilling is described as a “team sport” involving multiple entities. Steve Newman, Transocean’s CEO, characterized well construction as a collaborative process involving the well operator, government officials, drilling contractor, mud contractor, casing contractor, and cement contractor.
It is industry standard to rely on cement contractors for expertise (BP_00000015 §18 ¶862-872). BP selected industry-leading partners for the operation (BP_00000002 §11 ¶110-112), including Transocean, described as “one of the top drilling contractors in the world,” and Halliburton, one of the “top suppliers of foam cement in the world”.
BP functioned as the well architect and coordinator, not the direct operator (BP_00000002 §11 ¶110-112). BP’s drilling engineers are considered “well architects and planning coordinators”. Like other operators, BP relied on cement contractors for designing slurry and testing.
Transocean was responsible for maintaining well control equipment and preventing fires/blowouts. The Blowout Preventer (BOP) was approved by MMS over 40 times.
2. Technical Oversight and Testing
BP had strict requirements for cement testing (BP_00000002 §9 ¶88-90). All required operational tests were performed according to standards. BP’s cement specialist confirmed strict requirements to test using rig samples.
The Deepwater Horizon underwent continuous inspections and audits since its 2001 delivery. BP conducted comprehensive audits in 2001, 2005, 2008, and 2009. The rig was current on all required Flag State surveys and possessed all necessary international documents.
BP’s decision to use a long string was based on robust modeling (BP_00000040 §9 ¶407-412). Long string production casing is a common industry practice, with 57% of deep water wells using this design. The barriers outside the long string were tested according to normal procedures.
Mark Hafle emphasized that cost was not the primary driver in choosing a long-string casing design. The team prioritized well integrity and efficiency over cost savings.
C. Cement Job and Well Design Issues
The cement job involved two components: a base cement slurry and a nitrified foam cement slurry. Both slurry types were designed and tested by Halliburton. BP hired an independent lab (CSI Technologies) to conduct limited testing.
BP concluded the nitrified foam cement slurry likely experienced nitrogen breakout and migration, incorrect cement density, and potential contamination of shoe track cement. Halliburton proposed a nitrified cement slurry with increased technical complexity.
Ambrose’s team found that Halliburton included an antifoaming agent (D-Air 3000) that was incompatible with foam cement. The team concluded the antifoaming agent should not have been included without extensive laboratory testing. Ambrose’s team identified cement failure as a “precipitating cause” of the Macondo incident.
D. Blowout Preventer (BOP) Issues
The test ram was converted and could no longer seal from below. Lower annular preventer was changed by BP to a “stripping annular” with a 5,000 psi working pressure limit. The lower annular preventer was no longer rated to contain the potential well pressure.
Upper annular was examined at Michoud and was “completely destroyed”. Two pieces of drill pipe passed through the upper annular, suggesting severe erosion of the upper annular before the rig sank.
BP calculated the Maximum Anticipated Surface Pressure (MASP) at 8,404 psi. Regulations require that each BOP component’s working pressure rating must exceed MASP. Upper annular was rated at 7,500 psi, which is below the MASP of 8,404 psi. Both annulars were below the MASP, creating a significant safety concern.
E. Conclusion
Your Honor, the evidence presented demonstrates that this tragic incident resulted from a complex series of failures involving multiple parties. BP had established safety systems and relied on industry-leading contractors with specialized expertise. The misinterpretation of the negative pressure test was a collective failure by both BP and Transocean personnel, and Transocean’s subsequent failure to properly monitor the well and respond to kick indicators played a critical role in the disaster.
The cement job, designed and implemented by Halliburton, failed catastrophically. The BOP, maintained and operated by Transocean, had significant issues that prevented it from functioning as a last line of defense. The improper diversion decisions made by Transocean personnel allowed hydrocarbons to reach ignition sources on the rig.
While BP accepts its responsibility as the well owner, it is clear that liability must be shared among all parties involved in this complex operation. BP was not the direct operator of the well and wasn’t responsible for the key elements that failed here – the cement job, well control monitoring, kick detection, BOP operation, and emergency response.
The full closing argument spans 60 pages of sophisticated legal reasoning and arguments, supported by specific citations to the trial record. You can see the entire closing here.
What’s remarkable is that this entire project—from initial document analysis to final draft—was completed in about two hours on a Sunday afternoon, at a cost less than what many would spend on a nice dinner for two.
To be clear, we’re not suggesting that senior trial lawyers would—or should—read these AI-generated arguments verbatim in court. But the quality of this first draft is remarkable, likely surpassing what many associates might produce even with weeks of preparation.
John Tredennick and Dr. William Webber, Merlin Search Technologies.
To be clear, we’re not suggesting that senior trial lawyers would—or should—read these AI-generated arguments verbatim in court. But the quality of this first draft is remarkable, likely surpassing what many associates might produce even with weeks of preparation. It provides a comprehensive starting point that a skilled attorney could refine and personalize, saving countless hours of initial research and drafting.
Transforming Legal Practice
Our goal in this series was to demonstrate that generative AI can analyze complex trial materials and produce sophisticated legal arguments from multiple perspectives. In each case we asked the LLM to adapt to the differing legal positions that would be taken by different parties, while maintaining rigorous analytical standards. The technology doesn’t replace legal expertise—it amplifies it, allowing lawyers to focus their talents, knowledge and experience on refining and perfecting arguments built on comprehensive analysis.
For legal professionals who remain skeptical about AI’s potential impact on their practice, this series offers compelling evidence that the technology is ready now to assist with even the most challenging aspects of trial preparation. The question isn’t whether AI will transform legal practice, but how quickly and completely legal professionals will adapt to and leverage these powerful new tools.
The future of legal practice increasingly depends on thoughtful integration of artificial and human intelligence. By embracing these technologies, forward-thinking attorneys can enhance both their efficiency and effectiveness, delivering better results for their clients while maintaining the high standards our profession demands.
Assisted by GAI and LLM Technologies per EDRM GAI and LLM Policy.
