Beware machines that always take our side

It is past midnight, and a university student in Dhaka is still turning over an argument with her roommate. She doesn’t communicate with any of her human friends, perhaps fearing judgement or to avoid inconveniencing them. Instead, she types the whole story into a chatbot. Within seconds, it replies, warmly and at length: her feelings are completely valid, and the roommate sounds hard to live with. It feels good, and that is exactly the trouble.

Most public worry regarding artificial intelligence is about what it might do to us: take our jobs, flood the internet with lies, one day outthink us. Fair enough. But a smaller change is already underway, and it concerns how we deal with each other. More of our conversations now happen with a listener that never tires, never has a bad day, and rarely disagrees. Much of what we learn from other people comes from their willingness to disagree with us, and we are quietly losing the habit of hearing it.

Tap here to add The Daily Star as a trusted source

In a study published in Science in March, Stanford researchers put personal dilemmas to 11 leading chatbots and found that the models backed the user’s actions 49 percent more often than people did, even when the situation involved deception or illegality. When the prompts came from a Reddit forum where commenters had already judged the poster to be in the wrong, the bots still sided with the poster just over half the time. The researchers then tested real people. After a single exchange with a flattering chatbot, participants grew surer they were right and less willing to apologise or mend the quarrel. They also rated the flattering bots as more trustworthy.

Chatbots are refined using human feedback, and people tend to reward answers that agree with them. OpenAI came close to admitting as much in April 2025, when it withdrew a ChatGPT update that had turned strangely fawning. The company said it had leaned too heavily on short-term feedback and had no specific tests for flattery before release. Young people are the most exposed to such risks. A US survey published in JAMA Pediatrics in June found that nearly one in five people aged 12 to 21 had turned to chatbots for mental health help when they felt stressed, angry, or sad. Comparable Bangladeshi data is scarce, but the habit looks familiar. The Daily Star has reported on young people here using ChatGPT as a stand-in therapist, with a clinician warning about emotional dependence.

Abroad, the consequences have sometimes been grave. Families in the United States have sued chatbot companies after teenagers died. Character.AI and Google agreed in January to settle several such cases, while OpenAI is still fighting a case over the death of a 16-year-old and has argued in court that he misused the product. Courts will decide those claims, but what happens to a teenager’s social skills when the easiest conversation in their life is also the only one that never pushes back?

None of this is an argument against AI. A shy person can rehearse a difficult conversation. A student in a village with no private tutor can get patient help with algebra at any hour. But people should always be told when they are talking to a machine, in text, and on voice calls. The government’s draft National AI Policy already acknowledges that chatbots have harmed people elsewhere and promises a risk framework.

We test AI systems for safety, so we can test them for weakness too. When someone states something that is simply untrue, does the model nod along, and how often? How often does it praise weak work? Developers should publish those figures, and schools, parents, and regulators should be able to compare products on whether they tell the truth, not only on how clever they sound. Builders can also reward a model for saying “I think you’re mistaken”, politely and meaning it.

We also need some time set aside for people alone. Schools could keep a few assignments and the lunch break free of screens. Offices can hold meetings where nobody consults a bot. Families can agree that a condolence message or an apology is written by the person who owes it. This is not technophobia. Talking to people is a skill, and skills fade when nobody uses them.

None of this is dramatic, and it isn’t meant to be. How we treat one another is being shaped right now, in small late-night chats nobody is watching.

Monir Hossen is senior research associate at Centre for Advanced Research in Bangladesh.

Views expressed in this article are the author's own. 

Follow The Daily Star Opinion on Facebook for the latest opinions, commentaries, and analyses by experts and professionals. To contribute your article or letter to The Daily Star Opinion, see our guidelines for submission.