Dacception

Dacception is a term used in AI safety and data governance to describe the misalignment between data inputs and the information produced by an artificial intelligence system. It denotes situations where outputs convey inaccuracies, misrepresentations, or misleading conclusions about the real world, arising from issues in data quality, model training, or interaction context. The word is a blend of data and deception and is used to discuss reliability challenges across language models, image generators, and other data-driven systems. The term is informal and not universally standardized, and its precise definition can vary among communities.

Common causes include data contamination or leakage during training, biased or unrepresentative datasets, and distribution shift between training and deployment. Adversarial prompts and prompt engineering can also influence outputs toward incorrect conclusions. Model hallucination—a tendency to generate plausible but false statements—is closely related but dacception emphasizes the data-to-output chain, including how inputs, labels, and datasets shape results.

Dacception can lead to erroneous summaries, incorrect factual statements, or misrepresented statistics in generated content. It poses challenges for evaluation, verification, and risk management, especially in high-stakes domains such as healthcare, law, or finance. Addressing it involves improving data quality, auditing training pipelines, validating outputs against trusted sources, and developing transparency and reproducibility practices.

See also: data poisoning, model hallucination, data integrity, AI alignment.