Learn AI /How Machines Learn /Inference
Inference
- Starter
- How Machines Learn
- Technical readers
Also called: running a model, serving.
Back to the milestone you came from
In five seconds
What it is
Using a trained model on new input to produce an output such as a score, category, translation or generated text.
In plain words
Inference is using an already trained AI system on new data to get a result.
An everyday example
A city office receives hundreds of public comments about transport. A staff member has a printed guide with examples of what usually counts as a complaint, suggestion, question or praise. As each new message arrives, the staff member reads it and places it into the most suitable tray. The guide does not change during that shift; it is simply being applied to new messages.
A person may use broader context, policy knowledge and judgement, and may notice unusual cases in a flexible way. A model during inference usually applies numerical patterns from training and may produce an output even when the input is ambiguous, out of scope or poorly represented in its training data.
Remember thisTraining establishes the model's learned parameters. Inference applies them to new input to produce an output.
Understand it
Understand it
Why it matters
If you use AI tools, inference is the moment where a result is actually produced. Knowing this helps you ask practical questions: How fast is it, what input does it need, what kind of output will it give, and what human checking is still necessary before action is taken?
From new input to model output
After viewing this, the reader should be able to say that inference takes new input, passes it through a trained model whose settings stay fixed, and returns an output.
01
New input arrives
Start with fresh data the model did not use during training, such as a new sentence or image.
Draw a rounded rectangle on the left labelled 'new input' with example text inside such as 'Bus service is unreliable in the evenings'.
02
Prepare the input
The system may convert the input into a format the model can process, such as tokens or numbers.
Draw a second rectangle labelled 'prepare input' with small icons for text splitting or numeric conversion; connect from step 1 with an arrow.
03
Run the trained model
The prepared input passes through stored mathematical settings learned earlier. These settings are used, not updated.
Draw a larger central rectangle labelled 'trained model' with a small lock symbol and a note saying 'settings stay fixed during inference'; connect from step 2.
04
Return an output
The system produces a result such as a class label, score, translation or generated text.
Draw a right-side rectangle labelled 'output' showing examples: 'complaint: 0.82', 'translate', or 'next words'; connect from step 3.
A left-to-right flow diagram with four boxes connected by arrows. Box 1 shows a sample input. Box 2 shows pre-processing if needed. Box 3 shows the trained model with a lock icon on its settings to indicate no learning happens here. Box 4 shows one or more possible outputs. Small captions under the arrows explain the transition at each stage.
Hovering or tapping each box reveals one sentence: input type, preparation, fixed settings during inference, and output type. A toggle switches examples between classification, regression and text generation without changing the structure.
Describe this without the diagram
The diagram shows inference as a four-step flow. First, new input arrives, for example the sentence 'Bus service is unreliable in the evenings'. Second, the system prepares that input into a machine-usable form such as tokens or numbers. Third, the prepared input goes through a trained model. A lock icon and caption make clear that the model's learned settings are not being changed at this stage. Fourth, the system returns an output, such as the label 'complaint' with a confidence score, a number, a translation or generated text. The main message is that inference uses an existing model on new data to produce an output.
With reduced motion: A static left-to-right diagram with four labelled boxes and arrows: new input, prepare input, trained model with fixed settings, and output. Example text in each box replaces any animation or hover effects.
Inference is the stage where a model is used, not built. A model is a mathematical system whose internal settings were adjusted earlier during training. During inference, you give it new input and it returns an output.
What goes in depends on the task. The input could be text, an image, a table row, a sound clip or sensor readings. Inside the model, the input is transformed through many calculations using the settings learned earlier.
What comes out also depends on the task. The output might be a predicted class such as “complaint” or “praise”, a number such as a demand forecast, a translated sentence or the next generated words. This output is often called a prediction, but the two terms are not identical: inference is the process, while a prediction is one possible result.
People often confuse inference with training. Training changes the model’s internal settings by learning from many examples. Inference normally keeps those settings fixed and simply applies them to new data. In some systems the boundary can blur, but for a starter-level view, inference means using an already trained model.
How “Sorting consultation messages with a prepared guide” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the printed guide | the trained model with fixed learned settings |
| each new public comment | new input data given to the model |
| reading the message and choosing a tray | running calculations to produce an output |
| the chosen tray label | the model's prediction or generated result |
A worked example
A national officer uses an approved translation model on a routine, non-sensitive paragraph from an implementing partner's monitoring report.
Starting from: A paragraph awaiting a working translation for review.
- The system converts the source text into tokens that the trained model can process.
- In a forward pass, the model applies its learned parameters and computes probabilities for possible output tokens.
- A decoding procedure selects output tokens in sequence until the draft translation is complete.
- The system converts the tokens back into readable text and presents the draft for human review.
Result: This is inference: new text entered the already trained model, and the model produced a translation. No model training occurred during that run.
In a United Nations workplace
In an evaluation office, a national officer faces a translation backlog while preparing a progress report for clearance by close of business. The officer sends routine, non-sensitive excerpts from implementing partner monitoring reports through an approved translation model, then checks the drafts before placing them on the file.
Each excerpt is new input. The trained model performs inference by applying its learned parameters to that text and generating a draft translation. The model is being used, not retrained, even if it processes many excerpts in sequence.
Before you try thisOnly material permitted for that approved tool may be submitted. Personal, protection-sensitive, security-sensitive or otherwise confidential text must not be entered unless the tool is expressly approved for it, because the input may be retained, exposed or processed outside authorised controls. A qualified reviewer must check the translation before clearance.
Where you may hear it
- At a vendor demo during a procurement meeting
- In an ICT governance board discussion on hosting and processing capacity
- During clearance of a concept note for an internal AI tool
- At a briefing on an approved translation service
Avoid confusion
Avoid confusion
Often confused with
Training
Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.
Prediction
The result a model produces from input, such as a class, number, ranking or next token.
Model hosting
The systems and services that keep an AI model available so it can receive requests, run on computing hardware, and return outputs reliably and securely.
Side-by-side
Common misconceptions
Common misconception: Inference means the model is still learning from each new user input.
Correction: Usually it is not learning during this step. In standard use, the model's settings were set earlier in training and are simply applied to new input.
Common misconception: Inference and prediction are exactly the same thing.
Correction: Prediction is usually the output, such as a label or number. Inference is the whole process that takes input through the model to produce that output.
Common misconception: If inference returns an answer, the answer is probably factual.
Correction: Inference produces an output based on patterns learned earlier, not guaranteed truth. The result can be wrong, biased, outdated or inappropriate for the situation.
Quick check
Which statement best describes inference?
Inference is the use stage. The model has already been trained, and new input is passed through it to get an output such as a class, score, translation or generated text.
Go deeper
Go deeper
The simple version
Inference is using an already trained AI system on new data to get a result.
The technical version
Inference is the forward pass, and related processing, that applies a model's learned parameters to input to compute outputs such as class probabilities, numeric estimates, token probabilities or generated sequences.
Reviewed external resources
EncyclopaediaWikipedia search: Inference (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers find several possible articles on inference, and it suits readers who want a broad starting point rather than one fixed explanation.
search results, not one pagelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives structured introductory machine learning lessons, and it suits readers who want a guided path with practical examples.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is useful for step-by-step explanations in plain language, and it suits visual learners who like careful walkthroughs.
Starterlink checked 2026-07-31
VideoVideo explanations: Inference (opens in a new tab)YouTube search
This YouTube search results page helps readers compare different video explanations of inference, and it suits readers choosing a teaching style that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Inference (opens in a new tab)Reddit search
This Reddit search results page shows community discussions and questions about inference, and it suits readers who want to see common points of confusion and informal explanations.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Training establishes the model's learned parameters. Inference applies them to new input to produce an output.