AI Terms Explanation
Concise explanations of AI concepts for litigation teams. Each term defines the concept, its relevance, and common pitfalls.
Anti-Regurgitation
A set of techniques AI companies use to try and stop their models from spitting out verbatim text from their training data. These are filters, not a fundamental fix, and they are often leaky.
Data Poisoning
The act of secretly corrupting an AI model's training data to plant hidden behaviors, biases, or backdoors. It's the AI equivalent of industrial sabotage.
Data Contamination
A critical flaw in AI evaluation where the 'test questions' are leaked into the 'study materials.' The model appears to be brilliant, but it's just memorizing answers it has already seen.
Data Scraping
The automated process of copying massive amounts of data from websites without permission. This is the primary method AI companies use to acquire the data needed to train their models.
Deepfake
A synthetic video or audio clip created by an AI that is so realistic it appears to be authentic. It is the digital equivalent of a perfect, undetectable forgery.
Diffusion Model
The core technology behind AI image generators like Midjourney and Stable Diffusion. It creates images by starting with random noise and gradually refining it, guided by patterns learned from a massive dataset of existing images.
Finetuning
Taking a general-purpose AI model and training it further on a small, specific dataset to make it an expert at a single task. It's the AI equivalent of on-the-job training.
Explainability
The attempt to understand why an AI model made a specific decision. In practice, it often involves using a second AI to generate a plausible-sounding, but not necessarily true, explanation for the first AI's output.
Embedding
A complex numerical fingerprint (a vector) that represents a piece of data like a word, sentence, or image. It's how AI models 'understand' and compare concepts.
Mixture of Experts (MoE)
An AI architecture where instead of one giant model, the system is composed of many smaller 'expert' sub-models. When a query comes in, the system routes it to the most relevant experts to handle.
Hallucination
The tendency of an AI model to generate confident, plausible-sounding falsehoods. It is not a bug, but a fundamental characteristic of how these models work.
Jailbreaking
The art of tricking an 'aligned' AI model into violating its own safety rules by using clever, deceptive prompts. It's the AI equivalent of getting past a security guard by telling them a confusing story.
Machine Unlearning
The technically immense challenge of forcing a trained AI model to forget specific information, as required by privacy laws like GDPR. It is not as simple as deleting a file.
Model Collapse
The process where AI models, trained on the AI-generated content that floods the internet, begin to degrade. Each new generation learns from the flawed output of the last, leading to a downward spiral of quality and a distorted understanding of reality.
Model Extraction
A form of digital espionage where an attacker steals a proprietary AI model by repeatedly querying it and analyzing the responses to create a functional replica.
Post-training
The process of taking a general-purpose AI model and molding it for a specific job. This is where the model is taught its specific rules, biases, and, most importantly, its legal liabilities.
Quantization
The process of shrinking a large AI model by reducing the precision of its internal numbers. This makes the model smaller and faster, but it also creates a legally ambiguous 'copy' of the original.
RAG (Retrieval-Augmented Generation)
An AI system that 'looks up' information from a specific set of documents before answering a question. It's an attempt to ground the AI in facts and reduce hallucinations.
Pre-training
The first and most important stage of AI training, where a model is fed a massive, unfiltered chunk of the internet. This is where the model develops its core knowledge and, more importantly, its inherent flaws.
Regurgitation
When an AI model outputs text or an image that is a verbatim or near-verbatim copy of something from its training data. It is the AI equivalent of plagiarism.
Reinforcement Learning from Human Feedback (RLHF)
The process of 'civilizing' a raw AI model by having humans rank its responses. It's an attempt to teach the model to be helpful and harmless by showing it what humans prefer to hear.
Synthetic Data
Data that is artificially created by an AI, rather than being collected from the real world. It's the AI equivalent of lab-grown meat.
Torrenting
A decentralized, peer-to-peer method for downloading files, most commonly used for the mass-scale distribution of pirated, copyrighted material like movies, books, and software.
Transformers
The underlying AI architecture that allows models like GPT-4 to understand context. It works by weighing the relationships between all words in a text simultaneously, making it powerful but also prone to unique flaws.
Training Data
The vast collection of text, images, and code that an AI model is exposed to during its development. It is the model's entire universe of 'experience.'
Vector Databases
A special type of database designed to store and search embeddings (the numerical fingerprints of data). It's the high-speed filing cabinet that powers modern AI search and RAG systems.