Generative artificial intelligence (AI) creates new content by learning patterns from existing data. It can write a paragraph, produce a realistic image, compose music, imitate aspects of human speech, or generate a video from a written description. Although these abilities can seem fundamentally different, they share a central principle: a model learns statistical relationships in data and uses those relationships to generate new outputs.
Generative AI does not work by simply retrieving a finished answer from a database. During training, it adjusts a large set of numerical parameters to capture patterns in text, images, sounds, or other forms of information. When given a prompt, it uses those learned patterns to construct an output one piece at a time or through a sequence of progressively refined representations.
Understanding how this process works requires looking at how AI models learn, how they represent different kinds of information, and how they turn those representations into content. It also helps explain why generative AI can produce remarkably convincing results while sometimes making mistakes, inventing details, or creating material that appears plausible but is not accurate.
What generative AI is and how it differs from other AI
Artificial intelligence is a broad field concerned with building computer systems that perform tasks associated with human intelligence, such as recognizing objects, understanding language, making predictions, and solving problems. Different AI systems accomplish these tasks in different ways.
A traditional classification model, for example, might examine an image and predict whether it contains a dog or a cat. A forecasting model might use historical measurements to estimate tomorrow’s electricity demand. These systems primarily assign labels, estimate values, or make decisions based on patterns in their input.
A generative model instead learns patterns that allow it to produce new data. It might generate an image of a dog in a setting that was not present in any single training example, write an original explanation of how electricity works, or create a melody that follows familiar musical conventions without copying a particular song.
The distinction is not absolute. Some models can both analyze and generate information, and many modern AI systems combine several capabilities. The defining feature of generative AI is that producing new content is a central part of its function.
The word new also needs qualification. A generated result can be novel as a particular arrangement of words, pixels, or sounds without representing a wholly new idea or an unprecedented pattern. Models learn from existing material, so their outputs reflect the regularities, conventions, and limitations of that material. They may recombine familiar elements in unexpected ways, but they can also reproduce memorized material or imitate existing styles.
Generative AI is therefore best understood as a system for learning and applying patterns, not as a machine that creates from nothing.
How generative AI learns from data
Most modern generative AI systems are built using machine learning, a branch of AI in which computers improve their performance by adjusting internal parameters based on data. Many of the most capable systems use deep learning, which employs artificial neural networks containing multiple layers of interconnected mathematical operations.
Despite the name, an artificial neural network is not a digital copy of a human brain. It is a computational system whose parameters determine how information is transformed as it passes through the network. During training, an optimization algorithm adjusts these parameters to reduce errors on a learning objective.
Training begins with a collection of examples. Depending on the model, those examples may include books, articles, code, photographs, illustrations, audio recordings, video clips, or combinations of these materials. The data are processed into representations that the model can handle mathematically.
The model then performs a learning task. A language model might predict a missing or next token in a sequence. An image model might learn to reconstruct an image from a corrupted version or predict the noise added to an image. An audio model might learn relationships between sound representations and their surrounding context.
A token is a unit of information used by a model. In text, it may be a whole word, part of a word, punctuation, or another sequence of characters. Images and audio can be represented using other kinds of units, including numerical codes, patches, or values describing a signal.
During training, the model compares its prediction with a target or otherwise measures how well it performs its assigned task. The difference is used to calculate an error signal. An algorithm called backpropagation determines how the model’s parameters contributed to that error, while an optimization method adjusts them to improve future predictions.
This process is repeated across many examples. Gradually, the model’s parameters encode useful statistical relationships in the training data. A language model may learn grammatical structure, common factual associations, patterns of reasoning found in its examples, and relationships between writing styles. An image model may learn how shapes, textures, lighting, and objects tend to appear together.
These relationships are not necessarily stored as explicit rules. A model generally does not learn a separate instruction for every possible sentence or image. Instead, information is distributed across its parameters, which collectively influence its responses to new inputs.
Training can require substantial computing power because large models may contain billions of parameters and process enormous quantities of data. However, size alone does not guarantee quality. Data selection, model architecture, training objectives, optimization, and later evaluation all influence the final result.
How a prompt becomes generated content
After training, a model can be used for inference: the process of applying a trained model to new inputs. This is the stage most people encounter when they type a question into a chatbot, request an illustration, or describe a scene for a video generator.
The prompt supplies information that influences what the model generates. Depending on the system, it may include text, an uploaded image, an audio recording, a video clip, or a combination of these inputs.
The model converts the prompt into an internal numerical representation. It then uses its learned parameters and the available context to calculate what content should come next or how an existing representation should be transformed.
For a text-generating model, this often involves repeatedly predicting a probability distribution over possible next tokens. The system selects a token according to its generation procedure, adds it to the sequence, and calculates the next distribution using the expanded context. Repeating this process produces a passage of text.
Other generative systems follow different procedures. A diffusion model may begin with random noise and repeatedly refine it into an image. An audio model may predict sound tokens or waveform components, while a video model may generate a sequence of visual representations designed to remain coherent over time.
In all these cases, the prompt constrains the output without necessarily determining it completely. A request for a picture of a red bicycle beside a tree narrows the range of likely results, but it does not specify every pixel, the precise shape of the bicycle, or the position of every leaf.
The generation process also involves choices that affect the result. Some systems use sampling, which introduces controlled randomness when selecting among possible outputs. Others use more constrained decoding methods. Even when the prompt remains unchanged, repeated generations may differ because the model can follow different plausible paths.
The result is not necessarily a direct reflection of a single training example. It is produced by applying learned patterns to the current context. Whether that result is accurate, coherent, or useful depends on the model’s capabilities, the task, the quality of the prompt, and the way the output is generated.
How generative AI creates text
Text generation is commonly associated with large language models, or LLMs. These models are trained to process sequences of language and learn relationships among words, phrases, sentences, and larger passages.
Many modern LLMs use a neural network architecture called a transformer. A key component of a transformer is attention, a mechanism that helps the model determine which parts of its input are relevant when processing a particular token.
Consider the sentence, “The child put the glass on the table because it was unstable.” Interpreting the word “it” requires information from elsewhere in the sentence. Attention allows a model to weigh relationships between tokens rather than treating each word as an isolated item. Across many layers, the model builds representations that reflect context, syntax, and semantic relationships.
Transformers are particularly effective at handling relationships among elements in a sequence, although their exact capabilities depend on their training and design. Some language models also use architectures other than transformers.
During pretraining, a common objective is next-token prediction. The model receives a sequence and learns to predict the token that follows it. Repeating this task across diverse text teaches the model many regularities of language, including how explanations are structured, how questions relate to answers, and how different subjects are discussed.
Once trained, the model generates text by predicting one token at a time. Suppose a user asks for an explanation of photosynthesis. The prompt establishes a topic and a communicative goal. The model calculates likely continuations based on the prompt and its learned patterns, then continues generating until it reaches a stopping condition or output limit.
The process is more sophisticated than choosing the most common word after each word in isolation. Each prediction can depend on a large amount of preceding context, and the model’s internal representations capture relationships across that context.
Additional training can make a language model more useful as an assistant. Supervised fine-tuning exposes it to examples of desirable responses, while preference-based methods can encourage responses that human evaluators or other training procedures judge more helpful, accurate, or appropriate. One family of methods, reinforcement learning from human feedback, uses human preferences to help shape model behavior. Not every model uses the same combination of techniques.
These methods influence how a model responds, but they do not guarantee that its statements are true. A model is trained to perform particular computational tasks, not automatically equipped with a reliable mechanism for checking every claim against reality. It may produce a fluent explanation that contains an incorrect date, a fabricated reference, or a mistaken inference.
Language models also differ in how they handle factual information. Some rely primarily on patterns encoded during training, while others can use search tools, databases, calculators, or other external systems. Access to such tools can improve reliability for suitable tasks, but it introduces its own limitations, including incorrect retrieved information and errors in interpreting results.
The essential point is that text generation combines learned linguistic patterns with the context of the current interaction. Fluency is evidence of a model’s ability to generate plausible language, not proof that every statement is correct.
How generative AI creates images
Image generation presents a different challenge because an image contains spatially organized information. A model must produce relationships among shapes, objects, textures, colors, lighting, and perspective rather than simply arrange words in a sequence.
One important approach is diffusion modeling. Diffusion models learn to generate data by reversing a process that gradually adds noise to an image or another representation.
During training, an image is progressively corrupted with noise according to a defined procedure. At different stages, the model learns to estimate the noise or predict an equivalent quantity that helps recover the underlying signal. Across many examples and noise levels, it learns how image structure changes as noise is introduced.
During generation, the process starts from a noisy representation, often random noise. The trained model repeatedly estimates how to remove part of the noise, producing a progressively more structured result. After many refinement steps, the system converts the final representation into an image.
This is not simply a matter of sharpening a photograph. The initial random pattern does not contain a hidden, fully formed picture waiting to be uncovered. The learned model guides the process toward image structures that are plausible under its training and conditioning information.
Text-to-image systems commonly use text conditioning to connect a written prompt with visual generation. The system first converts the text into a numerical representation that captures relevant relationships among its words. The image-generation process uses this representation to favor visual results associated with the description.
A prompt such as “a ceramic bowl on a wooden table in soft morning light” specifies several related features. The model can generate a scene in which the bowl, table, and lighting appear together in a coherent arrangement, even though the exact composition may not have appeared in its training data.
Many systems use a compressed representation of an image rather than repeatedly processing every pixel directly. An encoder transforms an image into a lower-dimensional representation, often called a latent representation or latent space. A decoder can turn that representation back into an image. Performing diffusion in this compressed space can reduce computational demands while retaining important visual structure.
Other image-generation approaches exist, including autoregressive models that predict sequences of image tokens. These systems differ in architecture and generation procedure, but they share the broader principle of learning a representation of image patterns and using it to produce new visual data.
Image models can struggle with spatial relationships, intricate details, or objects whose parts must obey precise constraints. A generated picture may contain an inconsistent reflection, an incorrectly shaped hand, or an object that does not quite match the requested description. These errors occur because the model optimizes a learned generation objective rather than enforcing every physical and geometric rule of the real world.
A convincing image therefore need not be a physically accurate one. The model can produce a plausible visual pattern without understanding the scene in the same way a person who observes and interacts with the physical world might.
How generative AI creates audio
Audio generation involves producing sound as a waveform or as a representation from which a waveform can be reconstructed. A waveform describes how a sound signal varies over time. Human speech, music, environmental sounds, and other audio differ in their timing, frequency content, rhythm, and structure.
Directly generating a high-quality waveform is computationally demanding because sound contains many samples per second. For that reason, some systems first transform audio into a more compact representation, such as a sequence of discrete audio tokens or a continuous latent representation. A neural decoder can then convert the generated representation into sound.
One approach is autoregressive audio generation, in which the model predicts the next audio token or unit based on the preceding sequence and any supplied context. Another approach uses diffusion, progressively refining a noisy audio representation toward a plausible sound. Some systems combine multiple methods, using one model to generate a compact representation and another to reconstruct the waveform.
Text-to-speech systems illustrate how these principles can be applied to spoken language. Such systems learn relationships between text and speech, including pronunciation, rhythm, timing, and vocal characteristics. Depending on their design, they may generate intermediate representations such as phonetic units, acoustic features, or audio tokens before producing the final waveform.
Generating natural speech requires more than assigning one sound to each written word. The pronunciation of a word can depend on context, while sentence structure and meaning influence emphasis, pauses, and intonation. A capable model learns patterns that help coordinate these features across an utterance.
Voice cloning involves a related but distinct task. A model may be conditioned on recordings of a particular speaker to reproduce aspects of that speaker’s vocal characteristics. The quality of the result depends on the training or reference data, the model’s design, and how well it handles the requested speech.
Music generation requires models to capture relationships involving melody, harmony, rhythm, instrumentation, and musical form. A system may generate music from a text description, a short musical prompt, or an existing audio segment. It can produce a sequence of musical events, an audio representation, or a finished waveform, depending on its architecture.
Audio models can also generate sounds beyond speech and music, such as rain, footsteps, or machinery. These tasks require learning the temporal and spectral patterns associated with different sound sources.
Yet generated audio is not necessarily acoustically or semantically reliable. A spoken voice may sound natural while mispronouncing a word, placing stress awkwardly, or conveying an unintended emotion. Music can have a convincing texture while losing rhythmic consistency or structural coherence over a longer passage.
As with images, realism is a property of the generated signal, not a guarantee that the sound accurately represents a real event or a particular person’s genuine speech.
How generative AI creates video
Video generation extends generative modeling into both space and time. A video consists of a sequence of frames, each containing visual information, but its meaning also depends on how those frames change. A moving object must occupy different positions over time, lighting may change, and people or animals must maintain reasonably consistent appearances as they move.
A system that generates each frame independently could produce images that look convincing in isolation but change unpredictably from one frame to the next. Video generation must therefore learn temporal relationships as well as visual ones.
Modern approaches often represent video as a sequence of frames, patches, or compressed latent units. A model can learn patterns in these representations, including how visual features relate across neighboring frames and how motion develops over time. Depending on the architecture, it may generate video tokens autoregressively, use diffusion to refine noisy video representations, or combine these and other techniques.
In a diffusion-based video model, the system may begin with a noisy representation of a sequence of frames and progressively refine it into a coherent clip. Text or image conditioning guides the generation toward the requested scene. The model must account for both the content of individual frames and relationships among frames so that the result depicts a reasonably continuous event.
Video systems may also generate motion from an initial image, extend an existing clip, or transform a source video according to instructions. These tasks vary in difficulty. Generating a short scene with simple movement is different from maintaining the identity of several characters during complex interactions or reproducing a precise physical action over a long sequence.
Some systems generate audio and video together, while others create them separately and synchronize the results. Joint generation can help align speech, sound effects, and visible actions, but synchronizing multiple kinds of content remains a demanding problem.
Temporal consistency is one of the central challenges. An object may change shape unexpectedly, a character’s clothing may shift between frames, or an action may violate ordinary physical expectations. A person might appear to walk without properly transferring weight, or an object might move in a way that does not follow the forces acting on it.
These failures reflect limitations in the model’s learned representations and generation process. Video models learn statistical patterns associated with motion, but those patterns do not necessarily amount to a complete internal simulation of physics. They can produce motion that looks plausible in familiar situations without reliably predicting how every object would behave in the real world.
Longer clips and complicated scenes intensify the challenge because small inconsistencies can accumulate. Maintaining identity, spatial relationships, object permanence, and causal continuity requires the model to coordinate information over extended periods.
Video generation is thus not merely image generation repeated many times. It requires a system to represent how a scene changes, not just how it looks.
Why many generative AI systems use more than one model
Text, images, audio, and video have different mathematical structures, so a single generation method is not automatically suitable for all of them. Language is commonly represented as a sequence of discrete tokens, while images and audio may be represented as continuous signals, discrete codes, or compressed latent values. Video adds a temporal dimension to visual structure.
A multimodal model is designed to work with more than one type of data. It may learn shared representations that connect words with images, speech with text, or video with sound. These connections allow a system to respond to questions about an image, summarize spoken language, or use a written description to guide visual generation.
For example, a model trained on aligned image-and-text data can learn that certain phrases correspond to recurring visual features. A system trained on audio and transcripts can learn relationships between spoken sounds and written words. Such connections can support tasks that involve translating information from one modality to another.
Not all multimodal systems use a single unified architecture. Some combine specialized models, each responsible for a particular task, with components that pass information between them. A language model might interpret a prompt, an image model might generate a visual scene, and an audio model might produce its soundtrack.
Other systems are trained more jointly, allowing their components to learn coordinated representations. The appropriate design depends on the task, computational constraints, desired quality, and need for consistency across modalities.
Multimodal capability also does not guarantee that a system will correctly connect all forms of information. It might describe an image inaccurately, misunderstand a spoken instruction, or produce a video whose motion conflicts with its accompanying narration. The model must learn reliable relationships among modalities, and those relationships can be incomplete or ambiguous.
Combining modalities expands what generative AI can do, but it also increases the number of relationships the system must manage.
What generative AI learns—and what it does not necessarily understand
Generative AI can produce content that demonstrates substantial pattern recognition. Language models can explain technical subjects, image models can create intricate scenes, and audio models can generate speech with natural timing and intonation. These abilities emerge from training systems to capture relationships in complex data.
However, the meaning of understanding depends on what is being claimed. A model can learn useful representations of concepts and apply them to new combinations without necessarily possessing the grounded knowledge, intentions, or conscious experience associated with human understanding.
For instance, a model may learn that ice melts when heated and use that relationship correctly in an explanation. Whether it can apply the same relationship to an unfamiliar physical situation depends on the quality of its learned representations and reasoning procedures. It should not be assumed to possess a general-purpose physical model merely because it can describe physical laws.
This distinction matters because generative systems are often evaluated through their outputs. A convincing answer can hide a weak underlying inference, just as a plausible image can hide an impossible physical arrangement. Conversely, a model may make an isolated mistake despite having learned many useful relationships.
Researchers continue to investigate how models represent concepts, generalize beyond their training data, perform multistep reasoning, and connect language to the physical world. There is no single answer that applies to every architecture or task. Different models display different capabilities, and performance on one kind of problem does not establish equivalent competence elsewhere.
It is also important to distinguish generating a response from verifying it. A model can produce a statement because it fits learned patterns without having independently confirmed the statement. Unless the system uses a reliable verification process or suitable external evidence, the apparent confidence or fluency of its output should not be treated as proof.
The practical lesson is not that generative AI lacks all forms of understanding, nor that its outputs should be dismissed as meaningless. It is that the system’s capabilities must be judged by what it can reliably do, under what conditions, and with what evidence of correctness.
Why generative AI sometimes produces errors
Generative AI systems can make mistakes for several related reasons. First, training data contain gaps, contradictions, biases, and errors. A model learns from the material it receives, and its internal representations may reflect weaknesses in that material.
Second, the training objective may not directly measure the quality that users care about. Predicting a likely next token can encourage fluent text, but fluency is not the same as factual accuracy. Generating a visually plausible image does not require strict physical consistency. Producing natural-sounding speech does not ensure that the spoken words match the intended message.
Third, prompts may be ambiguous or underspecified. If a request permits several interpretations, a model may select one that does not match the user’s intent. Even a clear prompt may ask for information beyond the model’s capabilities or available context.
Fourth, generation often involves uncertainty. When several continuations are plausible, the model must choose among them according to its decoding method. Small differences in earlier choices can influence later content, especially in long sequences.
A well-known language-model failure is sometimes called a hallucination: the production of information that is false, unsupported, or invented but presented as if it were reliable. A model might provide a nonexistent book title, attribute a statement to the wrong person, or invent details to fill a gap in its response.
Hallucinations are not exclusive to language. Image generators can create visual details that do not match the prompt, and video systems can depict impossible motion or inconsistent objects. The same general issue appears in different forms: a model generates a plausible continuation of learned patterns without reliably enforcing the relevant constraints.
Additional techniques can reduce these failures. Retrieval systems can supply external documents, calculators can handle exact arithmetic, structured tools can validate certain outputs, and task-specific tests can expose recurring weaknesses. Human review remains valuable when errors have significant consequences.
No single method eliminates all mistakes. The most reliable approach is to match the model to the task, test its performance, and independently verify important claims or outputs.
How training choices shape generated content
A generative model’s behavior depends not only on its architecture but also on the data and objectives used to train it. These choices affect the range of material it can produce, the associations it learns, and the errors it tends to make.
The composition of training data matters. A model trained on a broad range of writing styles may handle varied language more effectively than one trained on a narrow collection. A visual model with limited examples of a particular object or setting may struggle to generate it consistently. Data quality also matters: inaccurate labels, duplicated material, and inconsistent examples can affect what a model learns.
Coverage is not the same as fairness or accuracy. A model may learn common patterns from its data while underrepresenting less common experiences, languages, or cultural contexts. It may reproduce stereotypes or reflect historical inequalities embedded in its training material. Filtering and evaluation can address some problems, but they cannot automatically remove every source of bias.
Training objectives shape behavior in a different way. A model optimized to predict language may become highly capable at producing coherent text without being optimized to check factual claims. An image model trained to match visual patterns may create attractive compositions without reliably obeying physical constraints. Models intended for specialized applications may need additional training and evaluation targeted to those tasks.
The prompt and decoding settings also influence the result. More constrained generation can improve consistency for some tasks, while more varied sampling can produce a wider range of outputs. Neither approach is universally best. The appropriate balance depends on whether the goal is factual precision, creative exploration, stylistic diversity, or another property.
Finally, a model’s performance in one setting may not transfer to another. A system that generates effective marketing copy may not be reliable at interpreting a medical scan. A model that produces attractive illustrations may not be suitable for engineering diagrams requiring exact geometry. Capability is task-dependent, and meaningful evaluation must reflect the intended use.
How generative AI differs from human creativity
Generative AI complicates familiar ideas about originality because it can produce work that appears novel without following the same process as a human creator.
Human creativity draws on perception, memory, emotion, goals, physical experience, cultural participation, and deliberate experimentation. People can revise their work in light of personal intentions and consequences. Generative models operate through computational processes shaped by training, prompts, internal representations, and generation procedures.
The distinction does not mean that generated material cannot be useful or aesthetically compelling. A model can produce a surprising combination of ideas, help explore alternative designs, or create a draft that a person develops into a finished work. The result can have value regardless of whether the model’s process resembles human thought.
Nor does the fact that models learn from existing material mean that every output is a direct copy. Learning general patterns can support combinations that were not present as complete examples in the training data. At the same time, some outputs may closely resemble training material, and statistical novelty does not establish that an output is entirely independent of prior works.
The relationship between training data and generated content depends on the model, its training process, and the particular output. Questions about copying, attribution, ownership, and permissible use involve technical, legal, and ethical issues that cannot be settled by the mere fact that a model generates new sequences or images.
Human involvement also varies. A person may use AI to brainstorm possibilities, generate an initial draft, or automate a narrowly defined task. In other cases, a system may produce content with limited review. Responsibility for checking accuracy, respecting rights, and considering the consequences of publication remains an important part of using these tools.
A productive way to understand AI-assisted creativity is to distinguish the ability to generate content from the broader human activities of setting goals, making judgments, interpreting meaning, and deciding what deserves to be shared.
The practical consequences of generative AI
Generative AI can reduce the effort required to produce first drafts, illustrations, synthetic voices, and video sequences. In education, it can help explain difficult concepts in different ways or create practice material. In research and engineering, it can assist with coding, explore design alternatives, and support the analysis or communication of complex information. In entertainment and the arts, it can expand the range of tools available to creators.
These benefits depend on context. A generated draft may save time while still requiring extensive revision. A synthetic voice may be useful for accessibility but inappropriate if it misrepresents a real person’s consent. A generated technical illustration may communicate a general concept but be unsuitable for a safety-critical specification unless its details are verified.
The technology also changes the economics of producing persuasive media. Content that once required specialized equipment, substantial training, or a production team can sometimes be generated with comparatively accessible tools. This can broaden participation, but it can also make impersonation, misleading imagery, fraudulent audio, and fabricated video easier to produce.
A realistic-looking or realistic-sounding result is not necessarily evidence that the depicted event occurred. As synthetic content becomes more capable, evaluating provenance—the origin and history of a piece of content—becomes increasingly important. Digital signatures, content credentials, records of creation, and independent corroboration can help in some settings, although no single technique establishes authenticity in every case.
Privacy and consent also matter. Training data may include personal information, and generated outputs may imitate identifiable individuals. Whether a particular system retains, exposes, or reproduces sensitive information depends on its design and operating practices. Users should not assume that every system handles private data in the same way.
Environmental and resource costs are another consideration. Training and operating large models require computing infrastructure and electricity, while the total impact depends on factors such as model size, hardware efficiency, workload, and how often the system is used. Smaller or specialized models may be more efficient for some tasks, although efficiency must be considered alongside performance and the full costs of deployment.
The consequences of generative AI therefore extend beyond the quality of its output. They depend on how the technology is integrated into institutions, how its results are checked, who can access it, and what safeguards govern its use.
What to expect as generative AI develops
The underlying scientific principles of generative AI are well established: models learn statistical structure from data, encode that structure in numerical parameters, and use it to generate outputs conditioned on prompts or other inputs. Neural networks, optimization, attention, latent representations, and probabilistic generation provide much of the technical foundation.
What remains uncertain is how far particular approaches can generalize, how reliably they can handle unfamiliar situations, and which combinations of architecture, training data, and external tools will work best for different tasks. Progress in one area does not guarantee equivalent progress in all others. Improvements in image realism, for example, do not automatically solve factual errors in language or continuity problems in video.
Future systems may become more capable at coordinating text, images, audio, and video, maintaining context across longer tasks, and using tools to check or act on their outputs. Their usefulness will depend not only on more powerful models but also on better evaluation, more reliable interaction with external information, and clearer ways to identify and correct errors.
The central principle will remain the same: generative AI creates content by learning patterns and applying them in new contexts. Its outputs can be original, useful, and sophisticated, but they are not automatically accurate, physically possible, unbiased, or trustworthy. Understanding both the mechanism and its limitations is the key to using the technology intelligently.