training
Training is the process by which an AI model learns to perform its task through countless examples, trials, and corrections.
In plain terms
In a traditional program, a human writes the rules. For an AI model, they do not: it is shown examples, and it extracts the patterns itself. This process is called training.
It is like when a child learns to ride a bike. They do not memorize physics formulas; instead, they try again and again, adjusting their movements slightly after every wobble. The model guesses similarly, checks how far off it was, and fine-tunes its internal settings – only doing this billions of times in a row.
Training is expensive and time-consuming work, and largely takes place before anyone starts using the model. For large language models, it takes weeks or months on thousands, or even tens of thousands, of specialized chips. When you ask a chatbot a question, it is no longer training but use: the model works from knowledge acquired earlier.
This is why a model's knowledge stops at a specific point in time, and why the data it was trained on matters so much. It works from what it saw in the training material, including its errors and biases.
How it works
- Data collection: developers compile training materials, such as texts, images, or audio recordings, often together with the correct answer.
- Guessing: the model receives an example and attempts to solve it, for instance by predicting the next word in a sentence. Initially, its settings are random, so its guesses are practically made at random.
- Error measurement: a formula, the so-called loss function, calculates how much the guess deviated from the correct answer.
- Adjustment: the model's internal numbers, the parameters (weights), are nudged slightly so that the error is smaller next time. This is repeated on countless examples, over many rounds.
- Verification: the finished model is tested on examples it did not see during training. This reveals whether it truly learned the task or merely memorized the training material.
An everyday example
You ask the chatbot about last week's news, and it replies that its knowledge only extends to an earlier date. This is because its training ended then, and it only knows about events since then if it is also able to search the internet. The same kind of training is behind your spam filter, which learned from millions of labeled emails what counts as spam.
Who is affected
A common misconception
Limits and risks
- A model is only as good as its training data: if the data is flawed, incomplete, or biased, the model will be too.
- Training ends at a specific point, so the model does not know about later events on its own and can confidently state outdated facts.
- Training large models costs enormous amounts of money and energy, so only a few companies can afford it.
- Training does not guarantee understanding: the model learns patterns and may unexpectedly err in situations different from the training material.
Similar, but not the same
- inference – Training is the acquisition of knowledge, while inference is the use of a trained model to answer your questions.
- fine-tuning – Fine-tuning is the shorter, targeted retraining of an already trained model for a narrower task, not training from scratch.
- prompt – With a prompt, you only instruct the finished model without changing its internal settings, meaning you are not training it.
Related concepts
training data, parameter, inference, overfitting, Neural network
A brief history
The term 'machine learning' was popularized by Arthur Samuel in 1959 with his paper on a checkers-playing program. The fundamental method for training modern neural networks, backpropagation, was widely popularized by a 1986 paper by Rumelhart, Hinton, and Williams.
Where things stand
As of late 2026, every machine-learning-based AI model is built through training, and the largest are trained in huge data centers. After pre-training, further targeted training, such as reinforcement learning, plays an increasingly important role. The debates focus mainly on the legal background of training data, costs and energy demand. In the EU, providers of general-purpose models have had to publish a public summary of their training data since August 2025.
In the news
- Cainiao opens its global warehouse network for training embodied AINov 5, 2026 · Ipar
- New standard proposal would simplify how publishers opt out of AI trainingNov 4, 2026 · Marketing
- LG AI Research tool checks copyright risk in training data 45 times fasterNov 4, 2026 · Jog
- South Korea launches AI data factory with 12 bipedal humanoid robotsNov 3, 2026 · Ipar
- Fear, lack of training and limited decision-making power slow workplace AI adoption in HungaryNov 2, 2026 · Work & HR
- White-collar workers quit careers over AI fears or seek 'AI-proof' jobsNov 2, 2026 · Work & HR
- Getting employees to embrace AI takes more than training, reports sayNov 2, 2026 · Work & HR
- Samsung: 95% of students use AI, but few get formal training; parents struggle to adviseNov 2, 2026 · Education & Family
Also known as
training · training · model training · training · training · training · model training · AI training
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.