Prolonged AI use can be dangerous to your health and work: 4 ways to stay safe

Follow ZDNET: Add us as a favorite source on Google.

---

ZDNET’s key points

• AI is getting better at small tasks, but still lags behind in long-form parsing.
• The consequences of prolonged interactions with AI can be disastrous.
• Use AI as a tool for well-defined tasks and avoid falling down a rabbit hole.

---

Better to do a little right than a lot wrong. So said the great philosopher Socrates, and his advice can be applied to your use of artificial intelligence, including chatbots such as OpenAI’s ChatGPT or Perplexity, as well as agency AI programs that are increasingly being tested in companies.

AI research increasingly shows that the safest and most productive course with AI is to use it for small, limited tasks, where outcomes can be well defined and verifiable, rather than pursuing extensive interactions with the technology for hours, days, and weeks.

Also: Asking AI for medical advice? There is a right way and a wrong way, explains a doctor

Extended interactions with chatbots such as ChatGPT and Perplexity can result in at least misinformation and, in some cases, illusion i dead. Technology is not yet ready to take on the more sophisticated kinds of demands for reasoning, logic, common sense, and deep analysis, areas where the human mind dominates.

(Disclosure: Ziff Davis, the parent company of ZDNET, filed a lawsuit in April 2025 against OpenAI, alleging that it infringed Ziff Davis’ copyright in the training and operation of its AI systems.)

We’re not yet at AGI (Artificial General Intelligence), the supposed human-level capabilities of AI, so we’d do well to keep the technology’s limitations in mind when using it.

Simply put, use AI as a tool instead of being sucked down a rabbit hole and lost in endless rounds of AI conversations.

What AI does well, and not so well

AI tends to do well at simple tasks, but poorly at complex and deep types of analysis.

The latest examples of this are the main takeaways from this week’s release Annual AI Index 2026 from Stanford University’s Human-Centered AI Scholars Group.

For one thing, editor-in-chief Sha Sajadieh and her collaborators make it clear that agent AI is increasingly successful at tasks like searching for information on the web. In fact, agents are close to the human level in routine online processes.

Also: 10 Ways AI Can Cause Unprecedented Damage

Through three benchmark tests: GAIA, OSWorldi WebArena — Sajadieh and team found that agents approach human-level performance on multi-step tasks such as opening a database, applying a policy rule, and then updating a customer record. In the GAIA test, agents have an accuracy rate of 74.5%, still below the 92% human performance, but well above the 20% a year ago.

In the OSWorld test, “computer science students complete about 72% of these tasks in an average time of about two minutes,” while Anthropic’s Claude Opus 4.5, until recently its most powerful model, achieves 66.3%. This means that “the best model (is) within 6 percentage points of human performance.”

WebArena shows AI models “now within 4 percentage points of the human baseline of 78.2%” accuracy.

While Claude Opus and other LLMs are not perfect, they show rapid progress to at least reach benchmark levels that are closer to human-level performance.

This makes sense, since manipulating a web browser or looking something up in a database should be one of the easiest scenarios where the natural language message can connect to external APIs and resources. In other words, the AI ​​should have most of the equipment needed to interface with applications in limited ways and perform tasks.

Also: 40 million people worldwide use ChatGPT for healthcare, but is it safe?

Note that even with limited and well-defined tasks, it helps to check what you’re getting from a bot, as the average score on these benchmarks still falls short of human capability, and that’s in benchmark tests, a kind of simulated performance. In real-world settings, results may vary and not upward.

AI can’t handle the hard stuff

When they delved into deeper types of work, the Stanford scholars found much less encouraging results.

The research found, they noted, that “the models handle simple searches well, but struggle when asked to find multiple matching pieces of information or apply conditions across a very long document, tasks that would be simple for a human scanning the same text.”

This finding aligns with my own anecdotal experience with ChatGPT draw up a business plan. The answers were fine in the first rounds of requests, but then degraded as the model was thrown into facts and figures that it hadn’t specified, or that might have been relevant earlier in the process, but didn’t include any business in the current context.

The lesson, I concluded, was that the longer your ChatGPT sessions are, the more bugs are placed on them. It makes the experience infuriating.

Also: I created a business plan with ChatGPT and it became a cautionary tale

The results of uncontrolled bot crafting can be more severe. An article last week in nature The magazine describes how scientist Almira Osmanovic Thunström, a medical researcher at the University of Gothenburg, and her team invented a disease, “bixonimania”, which they described as an eye condition resulting from excessive exposure to blue light from computer screens.

They wrote formal research papers about the invented condition and then published them online. Papers were collected in bot-based searches. Most of the major language models, including Google’s Gemini, began to faithfully relate the bixonimania condition to chats, pointing to Thunström and team’s bogus research papers.

The fact that bots confidently claim the existence of fake bixonimania speaks to a lack of oversight of technology’s access to information. Without proper checking, you can’t tell if a model will verify what it’s spitting out. As one scholar who was not involved in the research noted, “We should evaluate (the AI ​​model) and have a pipeline for ongoing evaluation.”

The consequences can be serious

A more serious variant, where a user appears to have gone down a rabbit hole of trusting a bot, is described in a recent New York Times article by Teddy Rosenbluth about the case of an elderly man battling white blood cell cancer.

Instead of following his oncologist’s advice, the patient, Joe Riley, relied on extensive interaction with chatbots, particularly Perplexity, to refute the doctor’s diagnosis. He insisted that his AI research revealed that he had what is called Richter’s transformation, a complication of cancer that would become more adverse with the recommended treatment.

Also: Using Google AI Overview for health tips? It’s ‘really dangerous’, research finds

Despite emails from experts questioning Richter’s material in Perplexity’s summaries of the disease, Riley maintained his belief in his AI-generated reports and resisted pleas from his doctor and family. He missed the window for proper treatment, and by the time he relented and agreed to try treatment, it was too late.

Rosenbluth makes the connection between the story of Joe Riley and the case of Adam Raine last yearwho killed himself after long talks with ChatGPT about his inclination to end his life.

Riley’s son Ben Riley wrote his own account of his father’s journey with AI. While young Riley doesn’t blame technology per se, he points out that immersing yourself in chats and losing perspective can have consequences.

“The fact is that A.I ago they exist in our world,” Riley writes, “and just as it can fuel those suffering from manic psychosis, it can also affirm or amplify our misunderstanding of what is happening to us physically and medically.”

Be careful with unreliable AI

The inclination to engage in lengthy discussions about depression, suicide, and serious health conditions is understandable. People have become accustomed to long engagements of hours at a time on social media. Some people feel lonely and a natural language conversation with a bot is better than no conversation at all.

Also: Your chatbot is playing a character, why Anthropic says this is dangerous

Robots tend to be sympathetic, research has shownwhich can make hours of engagement with a bot more satisfying than the usual stuff with a person.

And the companies that make the technology, while warning users to verify the bot’s output, have tended to place less emphasis on negative reports from individuals like Riley and Raine.

4 rules to avoid the rabbit hole

A few rules can help mitigate the worst effects of overemphasizing technology.

1. Define why you are going to a chatbot. Is there a well-defined task that is limited in scope and for which the bot’s predictions can be checked against other sources?
2. Have a healthy skepticism. It is well known that chatbots are prone to this confabulationconfidently stating falsehoods. It doesn’t matter how many chatbots you use to try to balance the good and the bad; all of which should be treated with healthy skepticism as containing only part of the truth, if any.
3. Think of chatbots not as friends or confidants. They are digital tools, such as Word or Excel. You’re not trying to have a relationship with a bot, you’re trying to complete a task.
4. Use proven digital overload skills. Take stretching breaks. Step away from the computer for some non-digital human interaction, like playing cards with a friend or going for a walk.

Also: Stop saying AI hallucinates, it doesn’t. And mischaracterization is dangerous

Falling down the rabbit hole happens in part as a result of being parked in front of a screen with no idle time.