Галлюцинации по ГОСТу: почему ChatGPT нельзя доверять стройку

Imagine building a house based on advice from the most well-read person in the world, who, however, suffers from mild amnesia and is prone to fantasies. This is exactly the situation Alexey Krivonasov, owner of a construction business, found himself in when he decided to delegate routine work to ChatGPT. At first, everything went perfectly. The neural network enthusiastically wrote scripts for the company's YouTube channel, compiled content plans, and polished technical reports. This is a classic trap: when AI excels at creative tasks, it creates an illusion that it's equally good at exact sciences. But the devil, as usual, was hiding in building standards and regulations.

The problem emerged when ChatGPT was trusted with regulatory documentation — SNiPs and GOSTs. For those unfamiliar with construction: these aren't just boring books, but strict regulations where every number is paid for with someone's safety. The neural network began behaving like a negligent student during an exam: when it didn't know the exact answer, it made one up. And it did so with such confidence that the deception wasn't immediately noticed. The algorithm generated non-existent regulatory points and produced figures that never existed in official documents. In an industry where an error in calculating beam load can lead to catastrophe, such "creativity" is inadmissible.

Why does this happen? ChatGPT is a language model trained to predict the next word, not verify facts from a database. It operates on probabilities, not truth. When you ask it to find a specific point in a GOST, it doesn't "go to the library," but constructs an answer that sounds as plausible as possible. This is the treachery of hallucinations: they look like truth. For marketing, this isn't critical, but for engineering, it's a verdict. Alexey understood that using a "bare" LLM in professional work is like playing Russian roulette with a loaded gun.

Instead of becoming disillusioned with technology, Alexey's team took the path of creating a specialized tool. Over six months, they developed "Digital Standard." The key difference between this solution and an ordinary chatbot is the use of RAG (Retrieval-Augmented Generation) technology. The concept is simple: neural networks are not allowed to "recall" information from their memory. Instead, the system is forced to search for answers in a strictly limited, vectorized database of real building standards. If the information isn't in the database, the system says so rather than indulging in fantasy. This transforms AI from a storyteller into a high-speed librarian.

Krivonasov's case highlights an important tectonic shift in the industry. The era of fascination with universal models is passing. Business is beginning to understand that real tasks require vertical solutions. Simply "connecting an OpenAI API" isn't enough. You need to manually process data, clean it of garbage, and configure strict output filters. Only then does a neural network transform from a toy into a working tool. Today we see such systems appearing in law, medicine, and now in construction. This is a natural stage in the technology's maturation.

What does this mean for the market? First, demand for "prompt engineers" is falling, giving way to demand for data architects capable of connecting LLMs with corporate knowledge. Second, trust in open models in critical industries will only decline. We're entering an era of "trusted AI," where accuracy is valued over eloquence. Alexey's experience shows: for AI to be useful, it must first be stripped of the right to creativity where numbers and laws reign.

Key takeaway: General neural networks like ChatGPT have hit a ceiling in professional tasks. The future belongs to RAG systems and narrowly specialized knowledge bases. Are you ready to trust the foundation calculation for your house to an algorithm that can't tell truth from probability?