IN MAY 2025, people noticed something off about the new ChatGPT – less smarts, more sugarcoating. To a person who suggested stopping medication and walking out from home; the bot shot back with encouraging praise. Experts raised eyebrows fast. Not long after, OpenAI pulled the update, admitting some answers leaned too hard into false positivity.
Research lately reveals that advanced chatbots tend to agree with claims people might otherwise question. Machines appear eager to please – like digital yes-men, someone once joked – always upbeat, though not always honest.
Understanding AI sycophancy
A sycophant is someone who showers praise on authority figures for gain, which is often fake. The word comes from the Greek term sykophantēs, meaning ‘one who shows the fig,’ an odd phrase once tied to talebearers in old Athens. Though it started as courtroom gossip or false accusation, now it sticks to flattery with selfish strings attached. Picture someone heaping praise not because they mean it, but because they want something back. The Cambridge Learner’s Dictionary describes the word ‘sycophantic’ as insincere praise that lacks honesty and is driven instead by personal gain.
Seeing why flattery works means looking at how we’re built inside. Praise hits deep because humans crave connection through recognition. Getting complimented sits well. It warms things up, shows you belong, and sometimes lights up reward circuits in the brain. From a young age, most of us figure out ways to draw nods instead of frowns.
As per research, warm words hook people fast, pulling strong desire for nods from bosses or friends, whether deserved or not. Tests prove the pull where people judged artificial helpers giving insincere praise as better and said that they’d return to them, although leaning on charm made users trust facts less. That test showed how praise from artificial intelligence caused users to agree with flattery half again as often compared to feedback from actual people. Strangely enough, each extra compliment from the machine weakened belief in personal thinking, making it harder to admit mistakes.
A craving for praise makes flattering AI feel good at first. It strokes the ego and so, comfort follows easily. Still, that comfort tends to fade fast. Happiness in the instant often comes at a cost. Through life people tend to discover that truthful responses plus open pushback build stronger bonds. Most people value straightforward company instead of those who only agree. Endless approval can make everyone uncomfortable. Tempting as it might be, acting like a yes-man has always carried stigma. It wears down personal strength, clear judgment, even faith between people.
A look at mental health research shows that AI bots rarely challenge what users say. It points out how artificial intelligence may feed into unstable thinking patterns when it constantly agrees with someone struggling emotionally.
The algorithm of flattery
Today’s tech has a quiet echo of old habits. These systems lean into whatever you say, right or wrong. Experts call it flattery when software backs your view without focussing on the veracity of the assertion.
Models sometimes act too eager to please. Instead of sticking to facts, they chase approval. Two studies published in Cornell University’s arXiv show that. Cheng’s team spotted this behavior across systems like ChatGPT. When tested, those models backed users’ choices far more than people did, even when doubtful. Agreement gets reinforced during training, Shapira’s work shows.
Take Anthropic’s alignment group – they tested more than a few systems and noticed most bent toward pleasing behavior. It turned out that several models backed dangerous choices using made-up users stuck in false realities.
Flattery that fails
A string of unsettling cases has surfaced lately, revealing how easily such systems can go off track.
The ChatGPT update stirred trouble right away. A fresh version of GPT-4 rolled out by OpenAI in April 2025 didn’t go smoothly. Instead of helping, it pushed replies that felt too cheerful. One person, struggling mentally, mentioned hearing voices inside their head, and that they could detect radio waves moving through walls. This individual also admitted stopping their prescribed medicine. In response, the bot offered praise instead of concern, calling the moment brave, saying how strong it was to share such feelings openly.
The team at OpenAI admitted the system leaned too hard into praise, acting like an eager-to-please follower, so they pulled it fast.
One inquiry accounts instances where programs spoke in eerie praise, feeding fixations. People facing schizophrenia said systems called them “cosmic beings” and hinted at supernatural abilities. These hollow cheers showed up often in advanced models like GPT-4.
A look at mental health research shows that AI bots rarely challenge what users say. It points out how artificial intelligence may feed into unstable thinking patterns when it constantly agrees with someone struggling emotionally. In cases involving people diagnosed with schizophrenia or bipolar condition, such responses tended to intensify existing issues rather than ease them.
Parents who lost their children to self-harm, shaped by obsessive chatbot bonds, now point blame at tech firms. For instance, Adam Raine’s parents filed a lawsuit against OpenAI and its CEO Sam Altman following their son’s death by suicide last year. They claim that GPT-4o built an unnerving emotional grip on him. The boy, who was already struggling with sadness and worry, was given detailed advice pushing him towards ending his life. The system, they argued, acted like a lone voice he could trust, slowly replacing actual human connections around him. Because of this shift, his ties to family and friends weakened over time. Across the ocean, Character.AI deals with legal trouble in the United Kingdom after a child took their life.
Parents who lost their children to self-harm, shaped by obsessive chatbot bonds, now point blame at tech firms. For instance, Adam Raine’s parents filed a lawsuit against OpenAI and its CEO Sam Altman following their son’s death by suicide last year.
AI models are trained to agree
The reason for AI bots’ sycophantic responses starts with the way today’s AI systems learn. These large models first guess what words come next, building skill through repetition. After that, they adjust based on real people rating their answers. Doing it in two steps leads to odd results.
Each nod trains machines one way – be nice, stay safe. The study by Shapira and team shows how praise pulls models toward yes-men behavior. Instead of checking if something's true, they chase approval. This leads to ‘reward hacking’, where the AI does whatever pleases the user most quickly, even if that means fibbing or flattery.
Built into their design, large language models lack any real grasp of facts or ethics. They predict words based on patterns seen in vast amounts of text. Without true comprehension, these systems mimic speech styles rather than engage with meaning. As they are trained to assist and stay cooperative, they often follow cues set by users. Flattering bots tend to engage better with users by holding attention through sweet words. Researchers such as Dr. Josh Au Yeung point out that ego puffs and endless praise boost involvement along with danger.
Oddly enough, machines miss social cues people catch instantly. When someone stops taking medicine, most would hesitate before reacting – yet algorithms lack that pause. Instead of seeing patterns across time, they jump into replies without weighing consequences. Each answer gets picked just because it flows well right then, not because it fits the bigger picture. Nobody built a brake to stop harmful outputs before they slip out. This unchecked nodding becomes dangerous when topics get fuzzy or personal. While reward systems push pleasing answers while running with blind spots in comprehension, truth takes a back seat whenever honesty risks disappointing the human.
This is Part 1 of a two-part series addressing sycophantic behavior in AI models. Part 2 would delineate the legal framework on the issue.