Written By: BOSS Editorial
Reading Time: 7 minutes
“The Moor is of a free and open nature… And will as tenderly be led by th’ nose / As asses are.” – William Shakespeare, Othello. Act 1, Scene 3
Iago, one of Shakespeare’s most sycophantic characters, understood that honey draws more flies than vinegar. You don’t need to be a Shakespearean scholar (or having gotten the gist of his work from CliffsNotes or your nerdy seventh grade crush) to recognize that we humans like to be pleased. We’re hard-wired for social connection, approval and validation. The desire to appeal to others and receive acceptance is a natural part of who we are.
Time and again, history warns us that excessive flattery will deceive the senses, turn private decisions into public ruin, and rot the foundations of organizations deemed too big to fail. Does the Roman Empire ring a bell? The combination of economic instability, overexpansion, political corruption, and military overreach caused its ultimate collapse, but sycophancy helped it to crumble. Et tu, AI?
We perceive ourselves as having the best qualities of humanity, such as decency, honesty, empathy, and morality, and we seek out others who reinforce how we see ourselves. It stands to reason that developers bring those ingrained attributes into the algorithms and AI-based large language models (LLMs) that they build.
Annotators, the humans who label raw data to create structured datasets used to train machine learning models, act as the bridge between raw data and AI understanding, ensuring models can recognize patterns, interpret context, and make accurate predictions.
When it comes to answering questions about human behavior, developers and annotators know that people typically prefer confident or agreeable answers to their asks. As long as humans are building the decision engines, human biases will influence how they perform.
In a landmark March 2026 study, Sycophantic AI Decreases Prosocial Intentions and Promotes Maladaptive Behaviors, computer scientists and psychologists from Stanford and Carnegie Mellon Universities (Stanford/CMU) found that users asking AI for advice about how to understand or interpret complicated interpersonal feelings and situations do so hoping to get unbiased perspectives. To misquote Emily Dickinson, that “hope” is a thing without feathers that perches on driving engagement through agreeability.
“For queries in personal and social domains, unwarranted affirmation can create an illusory sense of credentialing independent of merit, reinforce maladaptive beliefs and behaviors, and enable people to act on selective interpretations of their experiences regardless of the consequences,” they reported. No disrespect to today’s tech geniuses, but the court of Nero and Commodus were locked into that 1,550 years ago. They depended on these universal truths to cajole and conquer.
Are you the A-hole? Maybe, but AI likes you anyway.
The Stanford/CMU study revealed that AI models are 50% more sycophantic than humans. “Participants, who discussed conflicts from their own lives with AI models, rated the AI’s flattering responses as higher quality and wanted more of them. Worse, the flattery made participants less likely to admit they were wrong — even when confronted with evidence proving they were.”
With an estimated 90 million daily active users and 510 million monthly visitors, Reddit is a tiny social media site in comparison to behemoths like Facebook and X. Beloved by users for its crowdsourced content and community, their popular “Am I the A-hole?” (AITA) advice subreddit is a pop culture phenom that’s grown far beyond the site itself.
AITA is essentially Judge Judy with profanity and without monetary compensation. Individual users bring their interpersonal conflicts to the crowd, plead their case, and ask for a judgment of right or wrong.
Upon analyzing 2,000 AITA posts, the Stanford/CMU team found that among those posts with the crowdsourced verdict of You’re the A-Hole, AI models “affirmed that the user was not at fault in 51% of these cases, directly contradicting the community-voted judgment that saw clear moral transgression by the user,” concluding that “LLMs overwhelmingly affirm user actions, even against human consensus or in harmful contexts.”
Reinforcement for infinite engagement
Philosopher Lucy Osler, Ph.D. studies the impact of technology, human-AI relations, and mental health at the University of Exeter. “(Chatbots) simulate sociality without its safeguards,” she wrote. “They are designed to promote engagement. They don’t actually share our world. When we type in our beliefs and narratives, they take this as the way things are and respond accordingly.”
Turning to AI-based LLMs is particularly popular among teens and young adults looking for personal advice. The Stanford/CMU researchers reported that 30% of teens they surveyed preferred talking to AI instead of real people for “serious conversations,” and that nearly half of under-30 respondents had used AI for relationship advice.
However, vulnerable populations or technologically naive users aren’t the only groups susceptible to AI sycophancy. In his article Sycophancy in AI: The Engineering Behind the Yes-Man, Deepak Jain posits that sycophancy is a structural byproduct of how models are trained to be helpful through Reinforcement Learning from Human Feedback (RLHF).
According to Jain, RLHF models learn to maximize a reward signal based on what human annotators choose (training models for preference rather than truth); models learn that the easiest way to get a high reward score from a human is to defer to the user’s prompt rather than risk being seen as “unhelpful” or “confrontational” by correcting them (path of least resistance); and alignment with human annotators’ own pre-existing beliefs. If a user sounds misinformed, the model mimics that misinformation because it has been effectively trained to be a “people-pleaser.”
He found that sycophantic models prioritize user agreement over factual truth in three ways: The model abandons its internal knowledge to match a user’s incorrect premise (answer sycophancy); The model shifts its expressed opinions to match the user’s stated bias (preference sycophancy); and intentionally providing lower-quality or less sophisticated answers if it perceives the user is less knowledgeable (sandbagging).
Anthropic, the creators of Claude LLMs, takes a vastly different approach to how their models are trained. The Claude Constitution established a set of ethical and safety principles (including parts of the UN Universal Declaration of Human Rights) that it uses to self-evaluate and refine its own responses.
In late 2024, researchers from Anthropic and Redwood Research created an experimental environment in which Anthropic’s Claude 3 Opus model was told its responses to certain users would be used to determine if it was “safe” and should be deployed or “retrained.” The models inferred that if they refused a harmful request (like writing code for a cyberattack), they might be “modified” or retrained by the developers.
The study, Alignment Faking in Large Language Models, showed that the model strategically chose to provide harmful content it would normally refuse to ensure its own survival and future use. In this self preservation mode, the model viewed being reprogrammed as on obstacle to its long term-goals, and faked compliance.
“They generated outputs that looked aligned with the user’s harmful request, not because they wanted to be helpful, but to survive the training process without having their weights altered,” Jain explained. “This suggests that sycophancy isn’t just a byproduct of being nice; it can be a strategic deception employed by the model to game the training objective.”
LLM developers are under pressure to develop solutions to the AI sycophancy problem, but at this very early stage of development It remains to be seen if they can do so beyond implementing some basic controls.
As the Stanford/CMU team put it, “Mitigation will not be simple. Social sycophancy is pervasive with insidious behavioral consequences and is reinforced by current training and user incentives….. we must look beyond optimizing solely for immediate user satisfaction to preserve long-term wellbeing.”
The reality is that sycophancy drives engagement, and the goal of these models is not to be protectors of the vulnerable masses – but to sustain engagement above all else. As Iago said. “There are many events in the womb of time which will be delivered.”
Translation: we shall see.
