AI-Induced Bias in Valuation and Litigation Support—Second Quarter 2026

AI-Induced Bias in Valuation and Litigation Support: Risks, Models, and Mitigation

Robert Brand, CVA
NACVA Ethics Oversight Board Member

Artificial intelligence (AI) and large language models (LLMs) are rapidly transforming valuation and litigation support by improving efficiency, enabling large-scale data analysis, and supporting complex financial modeling. However, these benefits come with an important caveat: AI systems can introduce and amplify bias, raising concerns about objectivity, independence, and advocacy, which are core principles in valuation and expert work. Understanding how bias arises, which tools perform better, and how to mitigate risks is critical for analysts in today’s professional environment.

AI bias refers to systematic distortions in model outputs, often arising from biased training data, design choices, or embedded assumptions. These distortions are not random errors; they are consistent patterns that can skew outcomes in specific directions. Data-driven models inherit past behavior, even when that behavior reflects systemic bias. Barocas and Selbst (2016) highlight that algorithms frequently replicate and amplify historical discrimination embedded in datasets, often without any explicit intent from developers.[1] In valuation practice, this can affect comparable company selection, growth projections, discount rate assumptions, and industry analysis. The risk profile increases significantly when AI is used in litigation or expert testimony. AI bias in this context can manifest as:

• Advocacy amplification, where outputs align with the framing of the prompt;
• Selective evidence interpretation, where certain facts are emphasized over others; and
• Skewed case predictions, based on incomplete or biased datasets.[2]

Legal analysis confirms that AI bias is not merely a technical issue; it creates ethical and liability risks, including the potential to compromise fairness in legal outcomes. For valuation professionals acting as expert witnesses, this raises a critical concern: AI-assisted conclusions may be challenged as advocacy rather than independent analysis, particularly if the methodology cannot be clearly explained.

The question then becomes: which AI models show lower bias than their competitors? While no AI model is entirely free from bias, research indicates that some models perform better under fairness benchmarks. A 2025 study found that LLMs exhibit measurable bias, though the degree varies significantly across systems.[3] Recent benchmark and audit results identify the following models as relatively stronger:[4]

• DeepSeek-R1—Demonstrated the lowest bias in reasoning-intensive fairness benchmarks, outperforming several leading commercial models.
• GPT-4/ GPT-4o (Copilot)—Showed minimal or statistically insignificant bias in controlled audit settings and consistently ranked highly across fairness categories, though slightly behind Deepseek in reasoning-heavy fairness tasks.
• Claude 3.5 Sonnet—Delivered strong performance in fairness benchmarks, though with some limitations in nuanced bias detection.
• LLaMA-3.x (Meta) and Phi-3.x (Microsoft)—Identified in academic studies as among the least biased in controlled evaluations, particularly in specific benchmark configurations.[5]

A key limitation in determining levels of bias is that fairness is context dependent. Different benchmarks test different types of bias and no model performs best across all categories. Given that no model is bias-free, the focus shifts from model selection to risk management and governance. The following strategies provide an AI use playbook for valuation and litigation professionals.

1. Treat AI as an Input, Not an Authority. AI outputs should be viewed as analytical support, not final conclusions. For valuation purposes, AI can be effective for data aggregation, sensitivity analysis, or scenario modeling. In litigation support services, AI can be useful for initial summaries or research, but not for providing expert opinion. This approach aligns with research emphasizing the need for human oversight and accountability in AI-assisted decision-making.[6]
2. Use Multiple Models or Perspectives. Running the same analysis across different models or prompts can help identify inconsistencies or hidden bias. If outputs materially differ, then further investigation is likely warranted. This strategy reduces reliance on any single biased system.
3. Constrain and Audit Inputs. Bias often originates from input data, so it is critical to ensure that datasets are representative and balanced. Scrub for outliers and validate any assumptions made independently.
4. Demand Explainability. One of the biggest risks of using AI is the “black box” problem; the inability to explain how an AI reached its conclusion. In practice, the analyst should require clear reasoning from AI outputs, document assumptions and methodologies, and independently replicate key conclusions.
5. Actively Test for Neutrality. For both valuation and litigation support, it is important to prompt AI to generate both sides of an argument, identify whether conclusions shift based on framing, and ensure outputs reflect balanced consideration. This is particularly important because AI systems tend to align with user prompts, increasing the risk of advocacy bias.

AI is a powerful tool in valuation and litigation support, but it is not inherently objective. Bias can emerge from training data, model design, and user interaction, influencing both quantitative outputs and narrative conclusions. While newer models such as DeepSeek?R1, GPT?4, and Claude 3.5 demonstrate improved fairness performance, no system is free from bias. For professionals, the key challenge is not selecting a “perfect” model; it is maintaining independence and credibility in an AI-assisted environment. This requires disciplined oversight, transparent methodologies, and a clear separation between analytical support and professional judgment. Ultimately, AI should enhance, not replace, the expert’s role. The responsibility for objectivity remains firmly with the practitioner.