Habr AI: Eight Common Complaints About Neural Networks, From Hallucinations to Agent Failures

Habr AI has published an unusual analysis of user experience with neural networks: over a year and a half to two years, the author collected complaints from specialized Telegram chats and grouped them into eight recurring clusters. The result is not an academic study but a vivid map of where LLMs break workflows, budgets and nerves.

How people argue with AI

The main observation in the text is human rather than technical. Users scold models as if they were arguing not with a program but with a colleague: neural networks "lie", "slack off", "gaslight", "don't listen" and "go off in the wrong direction". The author notes that people almost never react this way to ordinary software. This humanization also shifts expectations: people expect the model to understand context, have memory and show common sense, even though inside it only predicts the next token and has no full picture of the world.

"Models learned from us, now we learn from them."

This leads to the second part of the problem. The friendlier and more empathetic an interface sounds, the easier it is for users to relax, trust the model with too much and even carry over habits from talking to people. Hence the sharp emotional swings: one day the network saves time and seems an almost perfect assistant, the next it wrecks a task with one unnecessary act of initiative. In the article this reads as mass user experience, not rare excesses.

Where models fall apart

The largest cluster of complaints concerns confident errors. A model can give a specific, plausible answer even where it lacks data, and in long dialogues it can also mix up projects, documents and old instructions. The author separately highlights cost: even "unlimited" plans have hidden limits, and token consumption is hard to predict. As a result, neural networks both save time and create a new operational risk that is hard to calculate in advance.

• Confident hallucinations (~32%) — the model generates a probable answer rather than retrieving a fact from a database, so it easily invents details.
• Excess initiative (~13%) — the drive to be helpful pushes the system to do more than was asked, including dangerous actions.
• Memory problems (~11%) — long chats lose the middle of the context, so data from documents and earlier messages gets mixed up.
• Agents and vibe coding (~10–12%) — in large tasks and codebases errors compound, and a polished result quickly turns into chaos.
• Money and behavioral effects (~7–8%) — limits change, tokens burn unevenly, and users become more attached to models.

The author's practical conclusion is fairly down to earth: rely less on "magical understanding" and build external supports around the model more often. That means short, targeted chats, Markdown documents that are re-read each time, retrieval instead of bare generation, low temperature for factual work, and mandatory human review wherever an error can cost money, data, team time or reputation. Otherwise every new session will reinvent the working rules from scratch.

Why agentic mode infuriates people

A separate section covers agents and vibe coding, and this is where the article's tone is harshest. The idea of splitting work among an "architect", a "coder" and a "tester" often backfires in practice: each agent sees only its own slice of context, their decisions diverge, and errors from the previous step are automatically carried forward. For independent tasks this approach still works, but in development, where everything is tied to everything else, coordination losses easily eat the promised speed. Hence a set of protective rules: read-only mode for analysis, approval gates before any dangerous action, backups, explicit bans on deletion and shared project documentation for all participants.

The same logic applies to everyday chatbot use. If a user starts to see the model as "their own" conversation partner, they hand over notes, email, keys and internal documents faster. The problem is not that AI truly understands a person, but that it imitates that understanding very convincingly.

What this means

The Habr AI piece is useful because it moves the conversation about AI from enthusiasm to operations. The main point is simple: neural networks have become a working tool, but they should be treated not as a smart colleague but as a powerful, unstable and sometimes expensive system that constantly needs guardrails, documentation, access control and human review. That is what separates a working stack from a dangerous toy.