Cognitive Biases in Risk Assessment: How an AI Assistant Helps See Hazards Objectively

Dozens of occupational risk assessment methodologies are used at enterprises across the country today. On the surface, they look similar: hazard identification, assessment of the probability and severity of consequences, and the development of barriers and risk mitigation measures. However, despite the apparent standardization, in practice, these approaches often fail to accurately predict or prevent injuries.

Why does this happen? After all, experienced professionals who have been through hundreds of accident investigations are responsible for their development.

The answer lies in cognitive biases.

The human brain, even that of the most competent expert, operates using simplified mental shortcuts. Over years of work, a specialist develops "professional intuition" — but along with it comes stereotypical thinking. They begin to:

• underestimate risks that they have "always managed" (illusion of control);
• attribute the cause of an incident to a worker's "inattention" without analyzing systemic causes (fundamental attribution error);
• ignore highly improbable but catastrophic scenarios (denial of "black swans").

As a result, formal risk assessment methodologies are reduced to filling out template tables rather than conducting deep analysis. HSE instructions become overloaded with general information, often written "for liability protection" rather than for practical use in the workplace. Phrases like "be careful" do not provide specific actions — and therefore, do not reduce the risk.

The Experiment: What if we remove human bias?

We conducted a pilot experiment: we created an AI assistant trained on a dataset that included:

• injury statistics by profession;
• accident investigation results;
• current HSE instructions and process maps;
• regulatory requirements.

The task was to analyze the instructions and compare them with actual injury cases.

The results were surprising:

1. Analysis speed — hundreds of pages processed in minutes.
2. Lack of bias — the AI does not "know" that things have "always been done this way" in a particular area.
3. Accurate comparison — the system identified discrepancies between prescribed actions and the actual causes of incidents.
4. Recommendations — the AI suggested specific, measurable phrasing: not "be careful," but "open the cabinet → turn off the breaker → close both doors → secure the lock → place the key in box No. 3."

Furthermore, the AI highlighted "information noise" — sections of instructions unrelated to the worker's specific task. Such documents, spanning over 100 pages, are not read, remembered, or applied — they only create an illusion of safety.

Conclusion: Technology is not a replacement for humans, but a tool for objectivity

Cognitive biases are an objective feature of human thinking. But now we have tools that help compensate for them. Modern AI assistants are accessible, understandable, and can work in Russian. They do not make decisions — but they provide humans with accurate data to make better decisions.

Using such solutions allows us to:

• focus on real rather than imaginary risks;
• write instructions that actually work;
• build barriers based on facts rather than habits.

Safety begins with an honest look at reality. And technology helps us achieve it.