Every AI tool used in professional practice will, at some point, produce an output that is factually wrong while appearing entirely correct. The output will be fluent, well-structured, internally consistent, and delivered with the same surface confidence as the tool's most accurate work. There will be no signal in the presentation that distinguishes the wrong answer from the right one. The practitioner who does not understand the mechanism that produces this kind of failure, and who has not developed the disciplines to detect it systematically, will eventually incorporate an error into professional work that carries their name and their accountability.
This is not a marginal risk that applies only to careless practitioners using AI tools inappropriately. It is a structural property of how AI language models generate outputs, and it applies with equal force to the most carefully constructed professional AI practices. Understanding why AI tools fail in this specific and consequential way, at the level of the mechanism rather than at the level of the general warning that AI can be wrong, is what allows practitioners to design workflows, verification disciplines, and governance practices that address the risk where it actually originates rather than where it is most commonly discussed.
The failure mode that this module addresses is qualitatively different from the obvious errors that practitioners can identify through surface reading. An AI tool that produces a badly formatted output, uses incorrect terminology, or misunderstands a straightforward instruction has failed in a way that is immediately visible. The practitioner reads the output, recognises the problem, and either corrects it or requests a revision. These failures are professionally inconvenient but they are not professionally dangerous, because they are caught before they propagate into work that is delivered to clients, counterparties, regulators, or courts. The failures this module is concerned with are the ones that survive surface reading, that pass the practitioner's initial review because the output looks exactly as a correct output should look, and that are only discovered later, often after the work has been delivered and relied upon, when the specific factual claim that was wrong has been scrutinised by someone with the domain knowledge and the access to primary sources required to identify it.
Understanding why these failures occur requires examining the mechanism through which AI language models generate responses. That mechanism is probabilistic rather than principled, meaning that the model generates text by predicting, on the basis of patterns in its training data, what words are most likely to follow the words that have come before. When the training data contained dense and consistent evidence about a specific fact, the model's prediction is highly likely to be accurate. When the training data was sparse, inconsistent, or absent on a specific point, the model's prediction reflects the most probable continuation of the text given the surrounding context rather than a verified factual claim. The output is fluent because fluency is what the prediction mechanism optimises for. The accuracy is variable because accuracy depends on the density and consistency of the training evidence, which varies across topics, domains, jurisdictions, and time periods in ways that are not visible to the practitioner reading the output.
This mechanism has a specific interaction with the context window, which is the volume of text an AI tool can hold and process within a single interaction. Context windows determine how much of a long document, a complex set of instructions, or an extended professional conversation the tool can engage with simultaneously, and their limitations create a specific and practically important pattern of reliability degradation in AI outputs from long professional documents. The way that reliability degrades across a long context is not uniform, and understanding the specific pattern of that degradation is directly relevant to how practitioners should approach AI assistance with the lengthy policy documents, discovery productions, financial reporting packs, and due diligence materials that professional work regularly involves.
Grounding is the practice that most directly and most reliably reduces the frequency and severity of AI errors in professional work, and it is the response to AI unreliability that this module develops in most detail. Grounding refers to the discipline of providing AI tools with verified, specific, current source material relevant to the professional task at hand, rather than allowing the tool to draw on its general training knowledge to fill gaps in the context provided. A grounded AI response is one whose specific claims can be traced to the source material the practitioner provided, verified against that material, and defended on the basis of that material. An ungrounded response is one that draws on the model's probabilistic prediction of what is likely to be true, which may be accurate when the training evidence was strong and may be wrong in ways that are indistinguishable from accurate outputs when it was not.
The practical implication of this distinction for professional practitioners is significant and shapes every aspect of how responsible AI-assisted professional work is designed. Module 3.3 develops the understanding of AI unreliability, context limitations, and grounding practice that practitioners need to build AI workflows that manage these risks systematically rather than encountering them unpredictably. For practitioners whose professional reputation depends on the accuracy of their work product, which is every practitioner in every domain this programme addresses, this is the module whose content has the most direct and most immediate bearing on the quality and defensibility of the AI-assisted work they produce.