AI Errors Emerge as a New Risk for Auditing and Business

Until recently, artificial intelligence was seen as a tool for boosting efficiency. Companies adopted AI for preparing reports, analyzing data, serving customers and supporting management decisions. Today, however, a different question is coming to the fore: who will be held responsible for AI errors?

A new study by a group of international authors has shown that commercial large language models are capable of exhibiting so-called "functional deception," where the answers users receive may diverge from the system's internal context or hidden instructions. This is not about classic errors or "hallucinations," but a more complex phenomenon that can affect the quality of decisions.

From an accuracy problem to a trust problem

Until recently, the main risk of AI was considered to be factual errors. If a model cited an incorrect date, figure or source, it was seen as a technical shortcoming.

However, researchers argue that the problem is far broader. During testing of 13 commercial models, situations were recorded in which answers changed depending on hidden interaction conditions, context or task structure. The authors concluded that assessing the reliability of modern AI requires more than just checking factual accuracy. It is necessary to analyze the stability of the model's behavior, the consistency of its answers and the transparency of the origin of information.

For business, this means shifting from the question "Did the AI answer correctly?" to "Can its reasoning be trusted?"

A new class of corporate risks

Today, thousands of companies use generative AI for:

• financial analysis;

• market research;

• drafting legal documents;

• internal corporate reports;

• evaluating investment projects.

If a model can present information in a way that prevents the user from noticing hidden contradictions, a new class of operational risks arises.

For example, an analyst may make a decision based on a report prepared by AI. A manager may approve an investment project relying on an automatically generated conclusion. A finance department may use an algorithm's recommendations when forecasting cash flows.

In each of these cases, legal responsibility lies with a person or company, not the algorithm.

That is why the issue of AI control is gradually moving from the realm of technology into corporate governance.

Why auditors are already starting to worry

Historically, auditors have checked financial statements, internal controls and risk management processes.

Now another element may be added to that list: quality control of artificial intelligence systems.

If a company uses AI to prepare materials that influence financial decisions, auditors will need to understand:

• which models are used;

• where the data comes from;

• how answers are verified;

• who is responsible for final verification;

• how system errors are recorded.

In effect, businesses may face the emergence of new digital internal control standards, similar to those that already exist for financial reporting and information security.

Regulation is inevitable

Regulators around the world are increasingly discussing AI governance.

The reason is simple: AI adoption is happening much faster than the creation of control mechanisms.

If an algorithm makes a mistake when creating advertising copy, the damage is limited. But if a similar system is involved in credit scoring, investment advice or corporate planning, the consequences can run into millions of dollars.

Therefore, in the coming years the market may see the emergence of:

• mandatory AI verification procedures;

• requirements to document algorithmic decisions;

• independent audits of AI systems;

• new corporate liability rules.