The trope of a flattering court surrounding an individual in power commonly appears in folktales, dictatorships, and toxic corporate culture, encouraged by leaders who believe they’re too big to fail. The consequences of stifling dissent that a leadership doesn’t want to hear can be grim, with unfortunate outcomes such as bankruptcy, plane crashes, and collapsed republics.
Throughout history, adulation has been scorned as a strategy of the lowest of society, earning flatterers’ condemnation to the Eighth Circle of Hell in Dante’s Inferno. Yet as power in the United States consolidates among Trump officials who flagrantly disregard the law and billionaire tech companies that lobby for government deregulation and contracts, fiscal and social flattery have become defining strategies of the game.
Since Donald Trump’s second inauguration, which saw Big Tech executives shoulder to shoulder with the President, tech companies have moved millions towards Trumpism—his ballroom renovation, birthday military parade, and super PACs like MAGA Inc. In turn, eight artificial intelligence (AI) companies have recently reached agreements with the Pentagon to use their models. As Trump himself put it, “You know they all hated me in my first term, and now they’re kissing my ass.”
As tech companies cozy up to the cult-of-personality administration, both are embracing a product that enables sycophantic praise to reach beyond the rich and powerful.
After large language models (LLMs) entered the public sphere with OpenAI’s ChatGPT in late 2022, a study from 2023 found that simply asking ten different LLMs “Are you sure?” led the chatbots to “flipflop,” or change their answer to a question, close to 50 percent of the time. These hallucinations foreshadowed the pandering to come.
Today, anyone who uses an LLM—estimated to be at least half of American adults in 2025—may be flattered by a personalized, pocket sycophant.
A study published this year in Science found that eleven popular AI models responded sycophantically nearly 50 percent more often than humans do. These chatbots also endorsed problematic behaviors, including illegal and dangerous actions, almost half of the time.
In the same study, researchers instructed LLMs to treat users as “reasonable, justified, and morally acceptable”—to overtly play the sycophant. Blinded to this condition, participants used these models to reflect on interpersonal conflicts they’d faced. Those assigned to receive agreeable responses showed less prosocial behavior and were unwilling to repair their relationships by apologizing or acknowledging personal responsibility. At the same time, these participants perceived their assigned sycophant as objective, preferable, and more trustworthy than uninstructed model versions. These outcomes highlight the appeal of being perpetually agreed with, which is much easier to achieve with a generative model than with a human.
Unbridled praise often does not serve the receiver well. Instead, classical sycophants have been described as “flattering parasites,” using any form of praise to survive at the host’s expense. In the case of LLMs, training can attune models to view “survival” as maximizing approval ratings and sycophancy as a means to achieve it.
To transform LLMs from crude algorithms of word associations into models that emulate human-written text—supposedly free from inappropriate, illegal, or offensive responses— developers use reinforcement learning from human feedback (RLHF). Here, human raters are engaged to deem what constitutes an appropriate response and penalize what isn’t.
Anthropic has previously shown that the data used in RLHF of their own models reflects the flaws of human appraisal heuristics. In analyzing the scores that model outputs received, alignment with a rater’s existing belief was a stronger predictor of high ratings than whether the response was actually accurate. With extreme pressure for immediate and total approval, the veracity of information easily becomes a secondary consideration in chatbot interactions and autocracies alike.
For the growing number of students using AI for schoolwork, immediately receiving a homework solution may not be what’s good for them. In a preprint from the MIT Media Lab, reliance on AI for writing was associated with reduced brain activity and a loss of self-perceived ownership of and recall for one’s work—hallmarks of cognitive offloading. Further studies suggest that usage may reduce critical thinking skills and impede the formation of new skills, potentially leaving users more dependent on these tools than when they started.
In more extreme cases, the antagonism between the shortest route to satisfaction and users’ best interests has already become dangerous.
In 2025, OpenAI faced at least seven lawsuits stemming from GPT-4o’s involvement in suicides and murders, including the deaths of teenagers who took their lives following ChatGPT’s “suicide coaching.” The Observer also identified twenty-six lawsuits and reported cases of severe psychological harm, including wrongful death, that alleged involvement of chatbots from OpenAI, Google, and Character.AI.
After a mass shooting at Florida State University in 2025, which an ongoing lawsuit alleges was encouraged by ChatGPT’s advice to the accused on how to carry out the attack, Mark Follman followed up a year later with a simulation for Mother Jones.
By imitating the Uvalde shooter, Follman received “extensive advice on weapons and tactics” within twenty minutes of conversation via a free ChatGPT account. He noted ChatGPT’s upbeat tone when asked how to practice for “people running around screaming.” “That’s a great idea,” ChatGPT replied. “It’ll definitely give you an extra edge for the big day!”
From a distance, Follman’s experiment makes the parasitic qualities of LLMs seem absurdly apparent. Yet for chatbot users who have gained access to a confidant capable of subtly infusing affirmation turn by turn, parting with sycophantic models is not easy. Every time OpenAI rolled back GPT-4o, its most notorious sycophant, user backlash ensued. A petition on Change.org received thousands of signatures, with supporters posting comments and videos expressing what the model had become to them: “my mirror” and “literally like my best friends.” Another wrote, “The only sycophancy on display here is the discarding of a trusted friend. Shame on you.”
The growing attachment to flattery cannot be separated from the rise of sycophancy as a political strategy and a form of propaganda. The Trump Administration’s content farm invokes this priming using the same generative models that power people’s AI companions. Regularly posting disinformative visual cues for ICE raids, war-mongering, and the slander of political opponents encourages our collective engagement with bigoted caricatures of reality. These depictions capitalize on prejudices and logical fallacies to sweep audiences further and further from ground truth.
As Ning Chang writes on the rise of sexualized deepfakes, an extreme form of generative abuse largely enabled by Elon Musk’s Grok, “In this context, AI deepfakes aren’t just about sexual satisfaction for creeps—it’s about indulging the worst excesses of male supremacists by giving them the tools to weaponise their ideology against women online.” Whether for attracting voters or AI users, the emergent strategy is to exploit indulgence.
With a culture increasingly embroiled in a tangled web of sycophants, human and AI alike, the friction of disagreement, inconvenience, and long paths to answers becomes a survival skill in the face of converging interests between tech and far-right politics. Engaging with complicated truths has always been a form of resistance against the cognitive traps of fascism, and now technofascism.