AI hallucination:
what it is and why it comes from missing context

AI hallucination is a plausible answer that does not match the facts

The most widely accepted definition describes hallucination as generated content that sounds fluent and coherent but does not match the facts or the source provided. Academic research has studied the subject for years. A central reference is the survey of hallucination in natural language generation published in ACM Computing Surveys, which organises metrics, causes and mitigation methods across different tasks.

What matters for a company lies in the word plausible. An absurd answer is thrown out on the spot. A plausible answer gets through a quick read, because it has the format, vocabulary and tone of a correct answer.

The model generates the most likely continuation, with or without the right data

An LLM produces text by predicting, at each step, the most likely next piece based on what it received and what it learned in training. The mechanism is the same whether the information is in the prompt or not. If the question is about the SLA a customer signed up for and the contract never reached the model, the most likely continuation is a typical market SLA, written with the same fluency.

The model has no reliable internal signal that separates what came from the source from what came from the statistical pattern. The gap in the input becomes text in the output. That is why AI hallucination is, above all, a context problem.

AI hallucination shows up in five recurring forms

• Invented fact: an event, clause or attribute that does not exist, such as a termination penalty that is not in the contract.
• Invented source: a citation of a document, regulation or ticket that was never written, with a believable number and date.
• Plausible but wrong number: a value within the expected range, such as a margin close to the industry average, that nobody questions because it looks reasonable.
• Generic rule applied to a specific case: the standard returns policy quoted for a customer who negotiated special terms.
• Two records mixed together: one account's history combined with another account's contract, usually when the two have similar names.

Inside a company, the wrong answer goes into proposals, systems and decisions

In personal use, a hallucination costs a fact check. In business use, the answer moves on. An invented deadline goes into the proposal sent to the customer. A wrongly attributed discount is written to the ERP by an agent with write permission. A plausible number goes into the committee pack and supports an investment decision.

Each step gives the wrong information a stronger appearance of validity, because it now sits in an official document. The further from its origin the error is discovered, the more it costs to undo.

A bigger model and an instruction not to make things up do not solve it alone

Newer models tend to hallucinate less on general knowledge questions, though not every new generation improves on this. They still have no access to that customer's contract or that branch's stock. For internal data, model size does not replace the data, and AI hallucination about the business keeps the same cause.

An instruction in the prompt to answer only from the sources helps, and it has limits. When the source sent is the wrong one, out of date or incomplete, the model follows the instruction and answers from it. The hallucination changes origin and is still there.

What reduces AI hallucination is specific, verifiable context

The measures that work act on the input and on checking the output:

• Correct, specific context: that customer's record, in the current version, instead of passages that look similar to the question.
• Permission-based scoping: only the data the person may see, which also avoids mixing accounts from different portfolios.
• Defined terms: what each metric and status means in the company, so the model does not apply the generic meaning.
• Cited, verifiable source: each statement linked to the source system and record, with a path to check it.
• Human review for the actions with the greatest consequences: writing to a system, sending to a customer and financial decisions go through a person.
• Continuous evaluation: a set of questions with known answers, run with every change of source, model or prompt.

The Company Brain delivers the missing context and points to the source

The Company Brain acts on the main cause. The knowledge graph links the customer to the contract, orders and tickets, which reduces records getting mixed up. The business ontology records what each term means in the company. The semantic layer translates the question into the right query and applies the permission inherited from the source system.

The resulting slice is served to the model over MCP with the origin of each piece of information. When the answer cites an SLA, the citation leads to the clause in the contract. When the data does not exist in the sources, the context arrives empty at that point, and the workflow can treat the absence as an absence. AI hallucination drops, and what remains is verifiable.

Measuring AI hallucination takes real questions with known answers

No architecture eliminates hallucination. The Company Brain reduces how often it happens and makes each answer auditable, and that depends on up-to-date sources, definitions maintained by the teams responsible and correct permissions at the source. Wrong data in the ERP is still wrong in the answer.

The practical starting point is to gather a few dozen questions the teams already ask, with the correct answer checked by someone who knows the process. That set measures AI hallucination before and after each change, and shows which sources and definitions need to enter the context first.