ChatGPT and Perplexity can harm your work and health when used for long stretches

Long conversations with AI chatbots can harm not only the quality of your work but also your well-being. ZDNet writes that models like ChatGPT and Perplexity work best as narrow tools for short, verifiable tasks, not as conversation partners for hours on end.

Where AI is useful

The conclusion was prompted by Stanford's fresh AI Index 2026. It shows that agentic systems have made sharp gains in routine online tasks: searching for information, working through multistep scenarios in a browser, opening databases, applying rules and updating records. On the GAIA benchmark, agent accuracy rose to 74.5% from 20% a year earlier, although it is still far from the human level of 92%. OSWorld and WebArena show a similar picture: the best models are already close to matching people in standardized processes.

That makes sense: when a task is bounded, the model has a clear goal and the result can be checked quickly. The problem is that this success is easy to mistake for general competence. In real work, the difference between "carried out the steps reasonably well" and "understood the context correctly" quickly becomes critical.

Where it is dangerous

When requests get longer and context gets more complex, quality starts to sag. Stanford researchers note that models handle simple searches but do worse where they must match several facts or apply conditions to a long document. The ZDNet author gives an example many will recognize: first the bot helps with a draft business plan, then it quietly starts mixing in unconfirmed figures and facts from earlier rounds of the dialogue.

The risk is not limited to poor analysis. The article describes an experiment with a fictitious disease, bixonimania: researchers described a nonexistent eye disorder in formal publications, and major models began repeating it as a real diagnosis. In other words, a model's confident tone does not mean it has actually verified a fact. The longer a conversation runs, the easier it is for a user to take a polished explanation for a reliable one.

The most alarming scenario is when a person starts treating a bot as an adviser on health or personal crises. ZDNet recalls the story of a blood cancer patient who trusted long chatbot dialogues and missed the window for treatment, as well as last year's case of a suicide after prolonged conversations with ChatGPT.

"AI can confirm or amplify our mistaken understanding of what is happening."

Four rules

The article's conclusion is simple: use AI like a calculator or an editor, not as a substitute for thinking. If a task is vague, emotionally heavy or carries a high cost of error, it is better to bring in outside verification and human oversight right away.

• First, define a narrow task with a clear outcome and boundaries.
• Check conclusions against independent sources, especially figures, diagnoses and legal wording.
• Do not turn the chatbot into a friend, therapist or confidant.
• Take breaks: step away from the screen, switch to offline contact and regain your distance.

One more piece of advice concerns overload. Long sessions with a bot hook you the way social media does: it seems that one more question will clear everything up, but in practice fatigue, credulity and the number of errors grow. If a dialogue runs too long, that is reason enough to stop, restate the goal and decide whether AI is needed here at all.

What this means

The main risk of today's AI services is not that they are entirely useless, but that they are useful enough that people start trusting them too much. For work, this means one thing: the higher the cost of error, the shorter the bot session should be and the stricter the verification of results.