ChatGPT and Perplexity can harm work and health with prolonged use

Long conversations with AI chatbots can harm not only work quality but also user wellbeing. ZDNet writes that models like ChatGPT and Perplexity work best as narrow tools for short, verifiable tasks rather than as conversation partners for hours.

Where AI is useful

The basis for this conclusion is Stanford's recent AI Index 2026. It shows that agentic systems have made rapid progress in routine online tasks: searching for information, executing multistep browser scenarios, accessing databases, applying rules, and updating records. On the GAIA benchmark, agent accuracy reached 74.

5% compared to 20% a year ago, though it still falls short of the human level of 92%. A similar picture emerges in OSWorld and WebArena: the best models are already nearly matching humans in standardized processes. This makes sense — when a task is bounded, the model has a clear objective, and the result can be quickly verified.

The problem is that this success is easily mistaken for universal competence. In real work, the difference between "executed the steps reasonably well" and "correctly understood the context" quickly becomes critical.

Where it's dangerous

When requests become longer and context becomes more complex, quality starts to decline. Stanford researchers note that models handle simple searches but perform worse when they need to cross-reference multiple facts or apply conditions to long documents. The ZDNet author provides a familiar example to many: initially, the bot helps draft a business plan, but then imperceptibly begins mixing in unverified figures and facts from earlier dialogue rounds.

The risk isn't limited to poor analysis. The article cites an experiment with a fictional disease called bixonimania: researchers described a nonexistent eye disorder in formal publications, and major models began recounting it as a real diagnosis. In other words, a model's confident tone doesn't mean it actually verified the fact.

The longer a conversation goes, the easier it is for the user to accept a well-articulated explanation as reliable. The most alarming scenario is when a person begins treating the bot as an advisor on health or personal crises. ZDNet recalls the case of a cancer patient who trusted lengthy chatbot conversations and missed the treatment window, as well as last year's suicide case following extended conversations with ChatGPT.

"AI can confirm or reinforce our mistaken understanding of what's happening."

Four rules

The article's conclusion is simple: use AI like you would a calculator or editor, not as a replacement for thinking. If the task is vague, emotionally taxing, or carries a high cost for error, it's better to immediately include external verification and human oversight.

• First, formulate a narrow task with a clear outcome and boundaries.
• Verify conclusions against independent sources, especially numbers, diagnoses, and legal formulations.
• Don't turn the chatbot into a friend, therapist, or confidant.
• Take breaks: step away from the screen, switch to offline communication, and restore distance.

Separate advice concerns overload. Long sessions with the bot are addictive like social media: it seems one more question will clarify everything, but in reality fatigue, credulity, and error rates grow. If a dialogue lasts too long, that's already a reason to stop, reformulate the goal, and decide whether AI is even needed here.

What this means

The main risk of today's AI services isn't that they're completely useless, but that they're useful enough to begin trusting them with things they shouldn't. For work, it means one thing: the higher the cost of an error, the shorter the session with the bot should be, and the stricter the verification of results.