Artificial intelligence model
An artificial intelligence model is a computational representation that a system uses to produce outputs from inputs. In machine learning, training fits the model from data. Examples include linear models, decision trees and neural networks. A model does not have to be a neural network.
The model within a service
A document-classification service needs more than a trained model. Someone must collect inputs, prepare them in the expected format, interpret predictions and handle errors. Access controls, the user interface and monitoring belong to the surrounding system.
This distinction matters when evaluating a product. A model can score well on a test while the service fails because it receives incomplete documents or sends results to the wrong person. Model-level performance does not establish the quality of the whole workflow.
Specialised and general-purpose models
A specialised model targets a bounded task, such as classifying messages. A foundation model undergoes broad training and can support several downstream tasks, sometimes after adaptation. These descriptions concern training and intended use; they do not establish a universal size threshold or a guarantee of capability.
For a language-model service, check whether factual answers come from model outputs alone or whether retrieval supplies supporting documents. The model's parameters can encode information and sometimes memorise examples, but they do not provide a dependable, directly editable database of facts.
Reading model documentation
A model card records intended uses, evaluation methods and limitations. The original model-card proposal also recommends reporting performance under relevant conditions and across groups, rather than providing one aggregate score.[1]
Before adoption, identify the exact version and licence, then check the evidence for your task. A benchmark in one language or domain may say little about another. Record the deployment configuration so that a later update can be evaluated against the same examples.
Responsible selection
Set an acceptable error level and test a modest baseline. Compare memory requirements, response time and operating cost alongside quality. Include human review where mistakes have consequences. The NIST AI Risk Management Framework offers a voluntary process for considering risks across the system lifecycle.[2]
See also
References
- ↑ Mitchell et al. (2019), Model Cards for Model Reporting.
- ↑ NIST, AI Risk Management Framework.
