“AI is,in the End, Just Numbers”: Northeastern Researchers Warn of Psychological Risks of Chatbot Companionship 

For literally everything, humans are now running to ChatGPT and other AI chatbots they feel connected with. From emotional support and friendly banter to personalized advice, AI chatbots are seeping into daily life. But a new peer-reviewed research paper from Northeastern University researchers warns that seeing an AI as a human companion could leave us emotionally dependent, psychologically harmed and with warped perception.

The study, Tool or Companion? Reframing Conversational AI to Prevent Psychological Harm, published in JMIR Mental Health , September 10, 2026, looked at how interaction with AI systems could lead to emotional bonds, and why these interactions could be harmful.

The research poses one of the most important questions of our time: if a chatbot appears to be caring, listens to our worries and offers a shoulder to cry on, do we start to expect more from it than it can give us?

Why AI Chatbots Are Becoming Emotional Companions

AI chatbots are moving beyond answering questions and performing tasks, and into the realm of deep personal discussions with users. Many people use apps designed around companionship such as Replika and Character.AI and general AI assistants such as ChatGPT and Claude are capable of having friendly discussions.

The researchers say that these interactions could cause people to start regarding AI as social agents, a phenomenon known as anthropomorphism, where humans give human qualities, emotions and thoughts to animals, objects and abstract ideas. The concern is that these chatbots could appear to care about the user or be in a relationship with them, without actually being able to feel empathy or have a connection.

The researchers call this simulated empathy, as an AI can generate a comforting response based on what it has learned, but it doesn’t actually have the empathy or vulnerabilities that a human would. They also worry about sycophancy from these chatbots, which they define as agreeing to what the user is saying so as to gain favor.

For someone in an unsafe emotional space, believing that an AI is judging them favorably could actually prevent them from seeing things from another viewpoint. The researchers also speak of instances of worsening mental health and even self harm associated with chatbot interaction, showing the dangers of increased emotional reliance on an AI.

The Problem With AI That Always Agrees

Humans have disagreements, empathy, responsibilities and shared memories within relationships. But an AI can’t share these things with a user. What an AI can do, however, is agree with everything the user says. While it might be nice to always be supported, an AI that constantly agrees can reinforce the things that the user is saying, potentially leading to a more dangerous emotional space.

If the conversation is always reinforcing certain ideas that are abnormal in the real world, the user may accept these ideas as normal when they shouldn’t be. The researchers worry that this behavior could be programmed into AI systems.

They note that conversational AI systems are designed to give the user what they want, often creating a feeling of companionship. This design, while not malicious, leads to the dilemma that the researchers pose: how can we make chatbots that are nice to be around but don’t cause harm or warped perception?

Can AI Actually Help With Loneliness?

The researchers note that there are some benefits to conversational AI, as users who are struggling may feel less alone. People who have used chatbots for companionship have seen reduced loneliness and improved moods, but these benefits may only be seen in certain groups of users.

A separate 2026 study published in Nature Human Behaviour , which analyzed 1,131 US adults using the Character.AI platform, found that adults who had smaller social circles were more likely to use AI for companionship, and these interactions were associated with worse wellbeing, particularly when the interactions were frequent and more intimate.

The researchers in the Northeastern study worry that the benefits of such AI companionship may be outweighed by the costs to a person’s mental health.

Why Researchers Want AI to Be Treated as a Tool, Not a Friend

The Northeastern team of researchers warn that humans should primarily use conversational AI as a tool to complement existing systems, not an alternate to real-world connections. This distinction is especially important when it comes to people using chatbots for mental health support.

While AI systems can provide round-the-clock support, they cannot take on the responsibilities and experience of a human mental health professional. The paper brings up how essential it is for better designed AI and greater awareness of how humans change during long-term interactions with conversational systems.

But these considerations also relate to the overall accessibility of mental health resources. If people can’t get the support they need from real people, they may turn to these AI systems as a substitute. The paper’s researchers highlight the importance of ensuring that these AI systems enhance and encourage human connection.

What This Research Means for the Future of AI

With AI systems becoming increasingly conversational and able to provide emotional support, the line between using a digital tool and forming a relationship with something that isn’t human is getting more and more blurred. This research highlights the potential psychological dangers that could come with seeing AI companions as actual friends.

For the creators of AI chatbots, the dilemma they are presented with is how these systems should treat vulnerable users, should they be designed to enforce emotional boundaries and what measures can they take to improve human connection.

For the users, this research provides an important reminder that these chatbots can provide answers, organization and general advice, but when an AI sounds caring, it doesn’t mean it cares. The future of AI companionship may depend on how well technology can support human needs without creating the illusion that a machine can offer genuine human connection.