Beyond the Hype
Why AI may Undermine Your Credibility as an Expert Witness
In a recent New York Surrogate’s Court decision (Matter of Weber, 2024 NY Slip Op 24258), an expert witness’s reliance on Microsoft Copilot for financial calculations led to a notable exclusion of their testimony. The court highlighted the risks of using generative AI for evidence-based tasks, marking a significant moment for experts in accounting and other fields that demand high standards of accuracy and verifiability. This article discusses the case, as well as the risks and limitations of using AI.
In a recent New York Surrogate’s Court decision (Matter of Weber, 2024 NY Slip Op 24258), an expert witness’s reliance on Microsoft Copilot for financial calculations led to a notable exclusion of their testimony. The court highlighted the risks of using generative AI for evidence-based tasks, marking a significant moment for experts in accounting and other fields that demand high standards of accuracy and verifiability.
The case involved the accounting of a trust managed by the trustee, Susan F. Weber, for the benefit of a family member. The objecting party, questioning the prudence of certain investment decisions, presented expert testimony from Charles Ranson, who used Microsoft Copilot, a generative AI tool, to verify his calculations in a Supplemental Damages Report. However, during the proceedings, Ranson could not specify the prompts or inputs used in Copilot to arrive at his conclusions, nor could he explain the tool’s methods or its accuracy for such tasks.
This lack of clarity raised red flags.
Not only were Copilot’s calculations inconsistent across different devices, but Ranson’s inability to detail Copilot’s data sources and processing methods further eroded confidence in the evidence. The court ultimately found his reliance on AI unreliable, noting that even Copilot itself advises users to confirm its outputs with expert oversight. Given these concerns, the court excluded Ranson’s AI-supported calculations from consideration, emphasizing that AI must meet well-established legal standards of reliability and transparency before it can be admitted as evidence.
The Limitations of AI in Calculations and Data Analysis
While AI has become a powerful tool across many industries, it is important to understand that most AI models, including Microsoft Copilot, are not inherently designed for precise financial or mathematical calculations. Unlike traditional analytical software, which follows strict rules and formulas to produce definitive answers, generative AI relies on probabilistic data models to generate outputs. This distinction means that AI tools often provide approximations based on patterns from vast data sources rather than calculating results from strict algorithms or concrete financial principles.
Generative AI like Copilot operates by processing user input and generating responses based on probability rather than definitive fact. This allows the AI to offer a wide range of possible answers based on the context it was trained on. For example, if asked to estimate an investment’s growth, Copilot may return results that vary slightly each time, reflecting the tool’s dependence on generalized patterns rather than specific calculations. This approach can lead to inconsistencies, particularly when different devices or systems run identical prompts, as occurred in this case. The court found Copilot’s results varied across several machines—albeit within a narrow range—which again raised questions about the tool’s reliability for expert testimony.
Furthermore, AI’s lack of transparency regarding its data sources and calculation methods presents a major limitation. In the courtroom, evidence must be traceable and verifiable. However, because Copilot does not disclose the databases or formulas it relies upon, it is challenging for users to validate its calculations rigorously. This opacity means experts cannot substantiate AI-generated results, especially if they are prompted by unverifiable or incomplete data. As demonstrated in this case, the expert could not explain what Copilot drew from to arrive at specific figures, undermining his own credibility and the admissibility of his evidence.
This judgment reinforces that AI tools like Copilot may be useful for general estimates or generating suggestions but lack the structure needed for precise calculations. Without the ability to verify sources, trace methodology, or ensure consistency, relying on generative AI for financial or mathematical evidence can pose significant risks—particularly in fields where accuracy and accountability are paramount.
About Charles W. Ranson
Charles W. Ranson brings over 40 years of expertise in corporate fiduciary services, with a career beginning in 1990 at Bank of America. In 1991, he joined Chase Manhattan Private Bank in Palm Beach, Florida, eventually serving as President and CEO of Chase Manhattan Bank’s Florida subsidiary (now JP Morgan Chase). During his tenure, Ranson led teams managing fiduciary, investment advisory, credit, and derivative services for high-net-worth clients throughout the Southeast U.S., developing client relationships and providing advisory support to some of the nation’s wealthiest beneficiaries and investors. Following his work with Chase, Ranson applied his skills in trust and estate management at US Trust and Atlantic Trust before founding Integritas Advisors, a registered investment advisory firm with the SEC. Additionally, as Founding Chair and Managing Director of Tiger 21 Florida, Ranson gained unique insights into the complex relationships that ultra-high-net-worth individuals and entrepreneurs maintain with wealth advisors and the challenges of multigenerational wealth transfer.
Since 2011, Ranson has offered litigation consulting and expert witness services in trust and estate litigation, assisting both defense and plaintiff counsel. He has provided expert opinions in reports and has testified in depositions and trials on matters of fiduciary duty and the standard of care for trustees, both corporate and individual, based on the specific circumstances of each case. His assignments include cases involving breaches of impartiality, accounting failures, diversification issues, administration of trusts according to their terms, and disputes over trustee compensation.
Other Issues with the Expert Witness Testimony
Aside from his reliance on AI for calculations, Charles Ranson’s testimony faced multiple other significant issues that ultimately undermined his credibility. One major problem was his inconsistent use of financial analysis standards. For example, Ranson calculated damages based on a “lost profit” model, assuming that the trust’s retained asset (the Cat Island Property) would have produced greater value had it been reinvested. However, he failed to provide substantial evidence supporting when or at what value the property could have realistically been sold. His damages model did not account for practical real estate market conditions, such as the 2008 financial crisis or the COVID-19 pandemic, both of which had substantial impacts on property values and sale feasibility. The absence of an expert basis in real estate, combined with speculative assumptions, severely weakened his assertions regarding damages.
Further, Ranson’s reports overlooked essential details, such as expenses associated with maintaining the Cat Island Property and distributions made to beneficiaries over time, which would naturally influence any fair assessment of trust performance. Additionally, his supplemental report failed to incorporate the actual sale proceeds of the property in 2022, despite this being a critical factor in evaluating the asset’s retained value. This omission indicated a lack of thoroughness in his analysis, which the court noted in its evaluation of his credibility.
Another issue with Ranson’s testimony was his inconsistent timeline for calculating damages. His calculations began from 2004, several years before the trust actually acquired the property in 2008. This arbitrary start date introduced inaccuracies into his assessment, as he failed to use dates that aligned with documented events. The court found this approach speculative and unsupported by the record, further weakening his analysis.
Finally, Ranson’s reliance on assumptions rather than verifiable data sources reflected poorly on his expertise in fiduciary matters. His testimony included vague references to economic conditions on Cat Island based on anecdotal evidence rather than reliable market data, and he did not cite industry standards or precedents for his valuation methods. Overall, these methodological flaws, speculative assumptions, and gaps in real estate expertise collectively diminished the reliability of Ranson’s testimony and underscored the importance of substantiating expert opinions with concrete, defensible data.
Implications for Expert Witnesses Relying on AI Tools
Of late, there has been a proliferation of all kinds of AI tools and “overnight AI gurus” who promote AI as the new shiny tool that can magically solve everything for everyone, including attorneys and expert witnesses. And while some of these tools are great at certain things, they are not great at everything.
As an expert witness, you need to, at the very least, understand how a particular tool works and what its strengths and weaknesses are before you should even consider using it for your work.
Large language models, like ChatGPT or Copilot, are great at understanding and creating language because they are trained on tons of text. But they are not very good at math, and here is why:
Language models learn by noticing patterns in words, sentences, and ideas; sort of like a huge game of “fill in the blanks.” If they see “2 + 2 = 4” many times, they will remember that specific fact. However, they do not actually understand the math behind it. They do not know why 2 + 2 equals 4; they just remember it as something that comes up in text.
Math is different because it is not just about patterns in words but rules and logic that apply universally, even when numbers change. When you ask a model to solve a math problem, it does not truly “calculate” the way a calculator or a math-savvy person would. It just tries to guess the answer based on similar math problems it has seen before. For simple math, it can sometimes guess right, but for complex or less common problems, it often struggles.
In short, language models are trained to recognize patterns in words, not to follow math rules or do calculations accurately.
In this instance, the expert’s use of Microsoft Copilot to support his financial analyses raised serious concerns when he could not substantiate the AI-derived figures. Without a clear understanding of Copilot’s methods, data sources, or the specific prompts used, the expert was unable to explain the calculations or confirm their accuracy. This lack of transparency eroded his credibility and highlighted a significant limitation of using AI as a substitute for established analytical methods.
The court was clear that, without proper validation, AI-generated calculations could not meet the standards of accuracy and reliability expected in legal proceedings. When an expert cannot explain or verify the methods behind AI-assisted results, it may suggest a lack of thoroughness and potentially weaken the weight of their testimony. Additionally, courts expect experts to use tools and methodologies that are widely accepted in their field; introducing AI without clear standards or evidence of reliability may lead to a failure to meet admissibility criteria.
Warnings for Accounting Experts: Avoiding Pitfalls in AI Usage
While AI may offer speed and assist in data processing, it is not yet suitable for the rigorous demands of financial and legal accuracy required in court settings.
In contexts where precision and reliability are paramount, using well-established and widely accepted analytical methods—such as standardized financial models, manual calculations, and documented data sources—is essential. Traditional methods offer transparency and allow experts to explain each step of their analysis; a requirement when testifying under oath.
AI tools often rely on probabilistic models rather than exact algorithms, which can lead to unpredictable results. Moreover, AI-generated outputs are highly sensitive to input quality and the way prompts are phrased.
Finally, even when using AI tools as part of the analytical process, manual verification of results is crucial. AI-generated data should be cross-checked with conventional calculations and validated against trusted sources.
Accounting experts should avoid relying solely on AI-generated figures in their reports or testimony; instead, they should treat AI as a supplementary tool and always confirm its outputs with proven methodologies. By doing so, experts maintain control over their analysis and uphold the standards of accuracy and accountability that the legal field demands.
Best Practices for Integrating AI in Expert Witness Work
For expert witnesses considering AI as part of their analytical toolkit, responsible usage is essential to maintaining credibility and accuracy. While AI can be helpful for initial data processing or providing general insights, experts should treat it as a supplementary tool rather than a primary one for evidence-based calculations. AI-generated data should serve as an initial step, to be further refined and validated by traditional methods. This approach ensures that the core of any testimony or expert report remains grounded in verified, reliable calculations that withstand legal scrutiny.
Experts using AI tools should document their processes carefully, noting each step where AI contributed to their analysis. By keeping a detailed record of the inputs used, prompts given, and outputs generated, experts can provide transparency if questioned about their methods. This documentation also allows experts to retrace their steps, address any inconsistencies, and refine outputs with greater precision. Additionally, maintaining a clear log of AI-generated data alongside verified calculations helps experts demonstrate the accuracy and reliability of their final findings.
Verifying AI-generated results with established methods is another critical best practice. Relying solely on AI calculations can introduce errors, as AI tools may not be optimized for the exacting standards of financial or legal analysis. Experts should manually cross-check AI results using established analytical frameworks and verify them against trusted data sources. This cross-verification process ensures that AI tools enhance, rather than replace, the thoroughness of expert analysis.
Finally, transparency about AI usage during testimony or report preparation is crucial. Experts should proactively disclose any AI reliance to attorneys, courts, and other stakeholders, clarifying how AI contributed to the findings and noting any potential limitations. By disclosing this information, experts build trust and avoid appearing evasive or over-reliant on technology, reinforcing their role as a credible source. Adopting these best practices allows expert witnesses to harness the benefits of AI responsibly while upholding the integrity required in high-stakes legal environments.
Ashish Arun is the founder and CEO of Exlitem, an AI-powered search engine helping attorneys find and engage expert witnesses. He also leads Expert Witness Profiler, a due diligence firm enabling attorneys to vet both their own and opposing expert witnesses.
Mr. Arun can be contacted at (866) 955-4836 or by e-mail to sales@exlitem.com.
