Artificial intelligence (AI) is transforming industries at an unprecedented pace, creating new opportunities while introducing complex legal and technical challenges. As AI systems become increasingly embedded in products, services, and business operations, courts are beginning to confront questions surrounding liability, transparency, governance, and responsibility that have few established legal precedents. To better understand where AI litigation is headed, we asked members of DOAR’s AI Expert Team to share their perspectives on the disputes emerging today, the technical issues likely to shape future cases, and what organizations should be doing to prepare.
Q: As AI systems (including generative, predictive, and perception-based technologies) become more embedded in products and operations, what types of litigation are emerging—or likely to emerge—most quickly?
AI Expert Team: The litigation moving fastest involves consumer-facing AI systems because they interact directly with the public, and a single incorrect response can quickly become documented harm. Customer service chatbots are an early example. Courts have already held companies responsible for inaccurate information provided by their own AI systems, making clear that organizations cannot avoid liability by arguing that a chatbot functions as a separate entity. That pattern—a system confidently misstating company policy or providing incorrect information—is likely to repeat anywhere AI is placed between businesses and their customers.
Another major area of growth involves automated decision-making. AI systems used in hiring, lending, housing, and insurance are already being challenged under existing anti-discrimination laws. Because these models learn from historical data, they can reproduce or amplify existing human biases, meaning a system may function exactly as designed while still producing discriminatory outcomes. We also expect continued litigation involving hallucinated information, inaccurate references, and other misleading outputs generated by AI systems.
Intellectual property disputes will continue to expand as developers seek increasingly large volumes of training data. Questions surrounding copyrighted works, licensing, web scraping, and ownership of AI-generated content are becoming central to many disputes, while experts whose knowledge helped refine AI models may also challenge how their expertise has been incorporated into future systems.
Looking ahead, litigation involving emotionally persuasive AI, autonomous vehicles, medical AI, robotics, computer vision, and other safety-critical technologies is likely to grow significantly. Companies operating internationally should also expect increased scrutiny under jurisdiction-specific regulations governing privacy, data protection, and AI deployment.
Q: Recent lawsuits involving chatbots have raised questions about emotional influence, harmful outputs, and user reliance. From a technical perspective, how difficult is it to predict, test, or control these types of behaviors in real-world settings? How are courts beginning to evaluate whether an AI system functioned as intended versus failed in a foreseeable way?
AI Expert Team: Modern AI systems are inherently difficult to predict because they generate responses based on statistical patterns rather than following a fixed set of rules. Large language models are non-deterministic, meaning the same prompt can produce different answers at different times, and the range of possible user interactions is effectively limitless. While developers can tune systems to behave more consistently, they cannot predict with certainty how an AI model will respond to every possible input.
That does not mean testing is impossible. Responsible development includes benchmarking models against known questions, adversarial testing designed to expose unsafe behavior, and moderation systems that evaluate responses before they reach users. These safeguards are essential, particularly for systems capable of influencing important decisions or interacting with vulnerable users. However, current testing methodologies all have limitations. AI models remain something of a “black box,” and developers continue to discover new ways users can circumvent existing guardrails through creative prompting or unexpected interactions.
As a result, courts are increasingly focusing less on whether an AI system could fail and more on whether its failures were foreseeable and appropriately addressed. AI systems inherit the strengths and weaknesses of the data used to train them. If historical data contains bias or inaccurate information, the model may faithfully reproduce those patterns without experiencing a technical malfunction. Likewise, a model may perform exactly as intended while still generating harmful recommendations or unsafe outputs.
The critical question therefore becomes whether developers implemented reasonable safeguards proportional to the risks their systems presented. Courts and litigants are likely to examine the quality of training data, testing protocols, moderation systems, edge-case evaluations, and the ongoing monitoring of deployed systems to determine whether foreseeable risks were responsibly managed before harm occurred.
Q: What role could training data, model architecture, prompting methods, or retrieval systems play in future discovery disputes involving AI technology?
AI Expert Team: Training data is likely to become one of the most important areas of discovery because it reveals how information was collected, whether it was properly licensed, and whether known biases or inaccuracies were introduced during development. In many intellectual property and privacy disputes, understanding how data was sourced may become just as important as understanding how the model itself functions.
Prompting methods are also likely to receive significant attention. The instructions developers use to guide AI behavior often reflect the safeguards, limitations, and priorities intentionally built into a system. Those prompts may help demonstrate what risks a company anticipated, what protections were implemented, and whether appropriate guardrails existed before deployment.
By comparison, the underlying model architecture may prove less significant in many disputes. Many organizations begin with commercially available foundation models or established architectures, making their competitive advantage—and much of the relevant discovery—lie instead in proprietary fine-tuning, retrieval-augmented generation (RAG), custom data pipelines, and operational controls built around those models.
Taken together, these technical layers document what an organization knew, what it built, and what it chose to protect against. They also help distinguish between failures arising from the underlying model and those introduced through implementation, customization, or deployment decisions. As AI litigation matures, discovery will increasingly focus on the entire ecosystem surrounding an AI system rather than the model alone.
Q: How does experience in areas like infrastructure, data systems, or computer vision change the way you analyze AI disputes, compared with someone focused only on generative AI?
AI Expert Team: Generative AI often becomes the focus of litigation because it is the most visible part of a system, but it is not always where the underlying problem exists. Many incidents described as AI failures are actually security, infrastructure, or data failures—a broken security boundary, overly permissive access controls, poor data quality, or an issue elsewhere in the technology stack. Focusing exclusively on the model risks overlooking the true source of the problem.
A multidisciplinary background broadens the investigation. Understanding data systems, statistical analysis, cybersecurity, computer vision, and software architecture provides insight into how AI systems are built, deployed, and validated in real-world environments. Computer vision, for example, presents different risks than a chatbot. A system may perform well during testing but fail after deployment because it was validated using one population and deployed in another, creating foreseeable issues that may not be obvious to someone focused only on large language models.
Ultimately, broader technical experience changes where an expert begins the analysis. Rather than assuming the model itself is responsible, it encourages examination of the entire system to identify where the actual failure occurred.
Q: As AI systems become more conversational, human-like, or autonomous, how could that affect the way jurors perceive responsibility, reliability, and trust? What are the biggest misconceptions courts or juries may have about how AI systems actually work?
AI Expert Team: As AI systems become more conversational, people naturally begin to anthropomorphize them. Jurors may view these systems as objective, trustworthy, or even capable of independent judgment simply because they communicate with confidence and appear human-like. In reality, today’s AI systems do not reason or understand information the way people do. They generate statistically likely responses based on patterns learned during training, and they can present inaccurate or entirely fabricated information with complete confidence.
One of the biggest misconceptions is that AI “knows” what it is saying. It does not. A polished, authoritative response is not evidence that the information is accurate. Likewise, many people assume AI systems will produce the same answer every time they receive the same prompt, when in fact they are inherently non-deterministic and may generate different responses.
For juries, responsibility should not focus on the AI itself but on the human decisions behind its development and deployment. Important questions include whether users understood they were interacting with AI, whether the system disclosed its limitations, whether sources were identified when appropriate, and whether meaningful safeguards were in place. The more human a system appears, the more important transparency becomes in helping users understand what the technology can—and cannot—reliably do.
Q: How important are internal safeguards, moderation systems, and testing protocols when evaluating whether an AI company acted responsibly? Are companies moving quickly enough to establish governance, risk management, and accountability frameworks around AI deployment?
AI Expert Team: Safeguards, moderation systems, and testing protocols are central to evaluating whether an organization acted responsibly. No AI system can be guaranteed to perform perfectly in every circumstance, but companies can demonstrate that they took reasonable steps to identify foreseeable risks before deployment. Testing, red-teaming, moderation, and continuous evaluation help show that an organization actively worked to expose and reduce harmful behavior rather than simply hoping the underlying model would perform as expected.
The most challenging question is not whether failures can occur—they can—but whether companies identified known risks and implemented safeguards appropriate to the potential harm. Particularly in high-risk applications, organizations should be able to demonstrate not only that testing occurred before deployment but that monitoring continued once systems were released into real-world environments.
Many companies, however, are struggling to keep governance frameworks aligned with the pace of technological development. Competitive pressure encourages rapid deployment, while widely accepted standards for AI testing, governance, and accountability are still evolving. As AI systems become increasingly autonomous, one of the largest remaining gaps is ongoing oversight after deployment. Organizations often devote significant resources to launching AI systems but comparatively less attention to monitoring how those systems perform over time.
Q: Beyond chatbot-related harm cases, what other categories of AI-related litigation do you expect to grow significantly over the next few years? Looking ahead, which developments in AI regulation, litigation strategy, and expert testimony do you believe will have the greatest impact on companies deploying AI?
AI Expert Team: Beyond conversational AI, litigation is expected to expand across a wide range of industries and technologies. Automated employment decisions, lending, insurance, healthcare, biometric privacy, facial recognition, autonomous vehicles, robotics, and other safety-critical systems all present opportunities for disputes involving discrimination, product liability, privacy, and bodily injury. Because many of these claims fit within existing legal frameworks, courts do not necessarily need entirely new legal theories to address them.
Regulation is also likely to place increasing emphasis on transparency. Companies will face growing expectations to explain how AI systems use data, how important decisions are made, and whether training data was obtained appropriately. As governments continue developing AI-specific regulations, organizations should expect greater scrutiny of both their technical practices and the representations they make about AI capabilities.
At the same time, expert testimony is becoming more comprehensive. Rather than simply explaining how an AI model reached a particular output, experts are increasingly being asked to evaluate the broader ecosystem surrounding the technology—including governance structures, testing protocols, data management practices, operational safeguards, and organizational decision-making. As AI systems become more deeply integrated into consequential decisions, the central question is likely to remain whether organizations acted reasonably in designing, deploying, and overseeing the technology.
