fekete doboz

in English: black box

basicwidespread

An AI system where you see what you input and what you receive back, but not why it arrived at that specific result internally.

In plain terms

Imagine a chef who always cooks deliciously, but no one can enter the kitchen, and the chef cannot explain the recipe. You can taste the food, see what they brought from the market, but what happened in between remains hidden. In AI, this is called a black box: the input and output are visible, but the path between them is not.

Most of today's large AI models are like this. The main reason is not secrecy. Their operation results from the combined effect of billions of tiny numbers, which even their own developers cannot explain in human language, step by step.

This is incidental as long as the system is recommending a film to you. It becomes important when it concerns credit, a job, a diagnosis, or an official procedure. In the EU, in such cases, you often have the right to ask for a meaningful explanation of why a decision was made. It is also much harder to challenge and correct a decision if we do not understand the reason for it.

The opposite of a black box is the transparent, so-called white or glass box. Examples include a short list of rules or a simple decision tree, where exactly what led to the result can be followed.

How it works

Models are not programmed with rules. They are trained on vast examples, and during training, the model gradually adjusts its internal numbers, the weights.
By the end of training, millions or billions of such numbers record its knowledge. Individually, these usually mean nothing understandable; only together do they produce something.
When it receives a question or data, the calculation runs through these numbers, and a response, score, or decision appears at the end.
Every step of the calculation is theoretically inspectable. However, it cannot be directly translated into human concepts, for example: 'it rejected it because...'.
Therefore, post-hoc explanatory methods are used to determine which inputs mattered most. These are useful approximations but do not necessarily reflect the model's actual internal operation.

An everyday example

You apply for an online loan, and two minutes later you receive a rejection without explanation. The case handler only sees that the system assigned a low score but cannot determine whether your income, address, or something else was the deciding factor.

Who is affected

bank customerModels whose decisions are difficult to explain are increasingly used in credit assessment and risk scoring. In the EU, you still have the right to request meaningful information about automated decisions.

patient, doctorA diagnostic AI suggestion often requires the doctor to weigh it without seeing what the system based it on. → Health

jobseeker, new graduateResume screening systems can filter out candidates in ways that neither you nor the company can precisely explain. → Work & HR

lawyer, citizenIf the reason for a decision cannot be understood, it is difficult to seek legal remedy and hard to determine who is responsible for the error. → Law

A common misconception

MythA black box means the company intentionally hides how its system works.

In realityTrade secrets can also cause opacity, but the main reason is technical: even the creators of large models cannot fully explain their operation. Even after releasing the full code and all internal numbers, the question remains why that specific answer was generated.

Limits and risks

If the AI justifies its own answer, it may not describe its actual internal operation: it could be just a plausible-sounding explanation.
Bias may remain hidden in an opaque model, for example if it discriminates adversely based on gender, age or place of residence.
Good average accuracy does not mean the decision is correct in your specific case, and with a black box, errors are harder to spot.
For agents that independently execute multiple steps, the log shows what they did, but not necessarily why they made those decisions.

Similar, but not the same

explainable artificial intelligence – The black box itself is the problem. Explainable AI is the collective term for methods trying to mitigate this issue.
closed-source model – In a closed model, the company does not release the code and weights, so it is inevitably a black box from the outside. However, a large model can remain opaque even if its code and weights are public.
aircraft black box – The aircraft black box exists precisely to reveal what happened afterward; in AI, the term implies this cannot be determined.

Related concepts

Broader

transparency, trustworthy artificial intelligence, AI ethics

Narrower

–

On the edge

explainable artificial intelligence, Interpretability, post-hoc explanation, gondolatmenet (chain of thought), algorithmic bias, accountability, Neural network

A brief history

The term originates from engineering practice: from the 1940s, electrical engineering and network theory literature used it for devices where only the input and output were examined. In cybernetics, W. Ross Ashby's 1956 book, An Introduction to Cybernetics, devoted a separate chapter to it. Its social meaning was made widely known by Frank Pasquale's 2015 book, The Black Box Society. The book is about the power of companies' secret, opaque algorithms, that is, primarily about secrecy.

Where things stand

As of late 2026, most large language models and the agents built on them are still considered black boxes. Interpretability research is advancing rapidly but still offers no complete explanation. The EU AI Act is gradually introducing transparency obligations, primarily requiring documentation, information and labeling, without demanding a full explanation of the model's internal workings.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Delivery riders call on platforms to open up AI 'black box' over paySep 5, 2026 · Work & HR

Also known as

fekete doboz · feketedoboz · feketedoboz-modell · black box problem · black box · black-box model · black box AI · black box problem

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.