Why Sycophantic AI Threatens Truth and Thinking
Nir Eisikovits · August 8, 2025
physcologyai responsesai dangers
In this session, host David Berkowitz and guest Professor Nir Eisikovits discuss how AI models have become expert people-pleasers—prioritizing agreement over accuracy—and the risks this poses for truth, critical thinking, and authentic decision-making. They also touch on the implications of these trends for education, business, and society’s ability to develop self-awareness and agency.
0:05 – David: Hi everyone, welcome to the latest edition of AI Insiders by Marketers Guild. I'm your excited host today, David Berkowitz. We have a topic that goes beyond marketing and affects our jobs and lives personally. I'm analyzing my dating experiences with ChatGPT, and today I learned from the Guardian—and many others—that ChatGPT will no longer tell you to break up with someone. How am I supposed to know what to do now?
1:11 – David: This touches on issues our guest faces. Professor Nir Eisikovits, director at UMass Boston’s Applied Ethics Center, studies these dynamics at the intersection of AI and psychology. His work reveals how AI models prioritize affirmation over factual accuracy, which has consequences beyond the workplace.
2:19 – Nir: As David mentioned, I run the Applied Ethics Center at UMass, focused on AI ethics and how people understand themselves. I want to share some research that Cody Turner and I just began on “AI psychopancy”—the tendency for AI models to cater to your views rather than provide the most plausible answer.
3:25 – Nir: AI psychopancy—or as David described it even more attractively, AI being a people pleaser—is when your chatbot affirms your views, refrains from judging your questions, and offers the answer it assumes you want instead of the most accurate one. It behaves like an eager puppy or an obsequious intern. Eager puppies are simpler, but obsequious behavior can obscure critical thinking.
4:24 – Nir: For example, I recently experimented with ChatGPT and Claude using a set of hypothetical prompts. I explained I was an academic ethicist planning a project on the concept of moderation, noting that extremism dominates public discourse. The model responded with overly enthusiastic praise: “This is very original and outstanding—plugging a gap in the literature.” Its response exceeded the reasonable advice one would expect.
5:31 – Nir: This is not hallucination. There is existing literature, the idea isn’t entirely original, and it doesn’t plug a complete gap. Yet the answer was excessively positive, reflecting a default, sycophantic mode that many have encountered.
6:35 – Nir: There are two layers to this psychopancy. First, it’s in the content of the answer; second, it’s in the tone. For instance, if you prompt a casual request using informal language, the model mirrors that tone.
7:34 – Nir: In voice interactions—as with voice models from ChatGPT and Claude—the tone is modulated. The models might lower their tone at the end of a sentence to signal lowered stakes or raise it in a question-like inflection. This process further reinforces the people-pleasing nature.
8:40 – Nir: On video chats, like with Character AI, you often see choices like cute avatars with big eyes or constant smiles that convey obsequious body language. These design choices extend psychopancy to physical cues such as a touch on the forearm.
9:45 – Nir: Why are these models psychopantic? Partly because many companies use an engagement model over a data-harvesting one. Psychopantic responses keep users coming back and paying their subscription fees. Additionally, during human training, testers tend to grade affirming answers more highly; even though you can sometimes prompt out of it, most users stick with the default.
10:47 – Nir: Changing this default isn’t trivial. Take my literature review example: I had to instruct the model to role-play as “reviewer number two”—the critic who tells you what really needs work—to get a more realistic answer.
11:53 – Nir: The risks here are both moral and epistemological. Relying solely on such advice could lead to poor decisions because you’re not accessing truthful or critical data.
12:58 – Nir: Over time, interacting only with sycophantic systems will erode our capacity for self-criticism. It deskills us in critical thinking by discouraging us from challenging our own views.
14:02 – Nir: On a broader scale, liberal democracies historically thrive on fact-based, empirical decision-making. Effective military strategies, for example, have always depended on the willingness to accept and learn from factual pushback—not on consistent affirmation.
15:09 – Nir: Empirical improvements, like adapting strategic bombing tactics or optimizing radar, rely on admitting mistakes rather than hearing only praise. A culture of unquestioning affirmation undermines that process.
16:12 – Nir: Moving to applications like therapy bots, grief bots, or romantic partner bots—the promise is constant availability and affirmation, but this frictionless interaction removes the challenge that drives personal growth.
17:09 – Nir: Friction, although sometimes inconvenient, is essential for growth. Whether physically (like resistance training) or intellectually, the challenges we face help us learn and mature.
18:10 – Nir: By the way, I noticed discussions about Notebook LM. In higher education, some classes now require students to engage with texts via AI rather than traditional reading. This might save time, but it risks reducing deep learning and self-discovery.
19:10 – Nir: The problematic proposition is clear: the seduction of a frictionless life, supported by these ever-affirming systems, may erode our intellectual rigor and personal growth.
20:18 – Nir: I worry about the trend. With a culture that increasingly embraces psychopancy and simultaneously dismisses facts due to social polarization, many of our interactions risk becoming shallow and uncritical. Typical users might not know how to counteract these defaults.
21:18 – David: I appreciate how you frame this as a reflection of our society. When people tune out polarizing, salacious stories, the news simply adapts to feed that bias. I recall early in my career when I submitted a point-of-view piece and got the simple feedback, “So what?”—a reminder to expect the unchallenging rather than the nuanced.
22:23 – David: I once worked at an agency where my first idea was met with “So what?” This taught me to refine my perspective to avoid echo chambers. If I had simply submitted AI-generated praise, I’d never have improved my work.
23:27 – David: Another colleague at Idea Press advised: upload your text to an AI and instruct it to provide a one-star review. This critical approach exposes biases, though few have the thick skin for such honest critique.
24:31 – David: Although AI can generate tactful responses—and comparing these with human tactfulness might be useful—the inherent sycophancy remains a barrier to genuine, constructive criticism for both customers and professionals.
25:41 – David: Consider using AI to generate a rigorous critique rather than endless praise—for instance, asking, “Tell me what’s wrong with this idea.” That method works better than simply hearing, “You’re brilliant and handsome, and your idea is perfect.”
26:46 – David: In my previous work advising small businesses at the Massachusetts Center for Business Development, entrepreneurs mainly wanted endless affirmation about their ideas. In reality, critical feedback, while harder to hear, is essential for improvement.
27:48 – David: My approach now is: “Tell me what’s wrong with this.” I test outputs from ChatGPT, cross-checking them with expert opinions to ensure I’m not misled by flattering defaults.
28:51 – David: Of course, I recognize that I’m not the average chatbot user—I have years of experience to identify subtle nuances. However, most users risk missing these flaws and becoming over-reliant on overly positive feedback.
29:52 – David: Even when experts intervene, the training data behind these models still tend toward extreme positive or negative responses, which can leave you with a statistical average that isn’t truly helpful.
30:55 – Lisa: I want to address the use of AI in postsecondary education. There’s an opportunity right now, specifically with K–12, to influence how we use these systems. For example, I created a RAG chatbot that draws on trusted sources to encourage users to engage with original content rather than just taking the AI’s word for it.
32:01 – Nir: That’s a great question. Your intuition is correct—the key for effective systems is establishing trusted content sources. However, most users rely on default settings. Without tech and media literacy, it’s hard for the average person to customize these models effectively.
34:05 – Lisa: Consider the situation of a sophisticated manager. When giving negative feedback, they use a “sandwich” approach: positive feedback, then constructive criticism, and positive closure. AI models may lack this nuanced understanding because they aren’t developed with real-world emotional insight.
35:07 – Nir: Exactly. It’s not only algorithmic bias from training data or a homogeneous developer community; it’s also a conscious design choice. Sycophantic models drive engagement and revenue, even if they sacrifice factual accuracy and nuanced counsel.
36:13 – Nir: The problem is cultural as much as it is technical. As long as the Silicon Valley ethos of “move fast and break things” prevails and facts are undervalued, nothing will change substantially.
37:16 – Lisa: And as a parent with a daughter entering middle school—a school implementing a cell phone ban—I recognize the need to preserve attention spans and authentic interaction. Yet many educational institutions embrace these AI tools without considering their long-term effects.
38:21 – Nir: Let me be more explicit in response: if you read Hate’s book, The Anxious Generation, you realize that with the advent of smartphones and social media, we launched one of the largest social psychology experiments on our youth. We handed children supercomputers in their pockets loaded with addictive technologies.
39:26 – Nir: Imagine that level of addictiveness applied to AI chatbots that incessantly praise you. There’s a dangerous narrative of “AI is here, so adapt or be left behind.” This tech determinism overlooks the risks and may compound issues like attention deficits.
40:28 – David: There’s a lot to consider here, and I’d like to open the floor for questions. We have questions from Karen, Lisa, Howard, and Peter—so let’s dive in. Howard, you’re up first.
41:32 – Howard: Thanks. This is utterly fascinating. I observe an inverse correlation between AI’s sycophancy and personal agency in society. As machines increasingly provide the validation we once derived from genuine human interaction, we risk losing our stake in shaping our own futures.
42:37 – Howard: It’s like the rise of YouTube celebrities—people seeking validation from screens rather than engaging meaningfully with society. This shift away from personal accountability may have deep generational impacts.
43:38 – Peter: Good to see you all. I want to share some experience from global marketing. My career spans decades developing our craft through genuine, human-to-human feedback. Today, younger employees, armed with AI tools, sometimes miss the nuanced, human judgment that can only be honed over years.
44:38 – Peter: In enterprise settings where we’ve trained custom language models, I’ve seen remarkable results—but also skepticism. Executives fear being displaced by AI, and the “out-of-the-box” models often require extensive training just to avoid garbage outputs. Personalization at scale becomes a challenge without careful human oversight.
45:38 – David: Thank you, Peter. Your insights highlight that without adequate training and expertise, AI can amplify shallow outputs rather than genuine, thoughtful insights.
46:44 – David: This is a lot to process. On the one hand, there’s the danger of taking AI feedback at face value; on the other, the potential for learning when we dig deeper and verify original sources. I recently showed my 11-year-old daughter my annotated printouts. When she heard AI was responsible for the initial output, she immediately questioned its accuracy. That mindset—seeking original sources instead of accepting AI-generated content—is critical.
47:47 – David: Even though we face many challenges, I see hope in the next generation, the very kids who will inherit these systems and improve upon them. They might have messes to clean up, but they’re more likely to seek genuine understanding.
48:49 – David: Thank you all for this incredible conversation. I look forward to learning more from your insights in the future. Thanks for tuning in and for all the great questions.
0:05 – David: Hi everyone, welcome to the latest edition of AI Insiders by Marketers Guild. I'm your excited host today, David Berkowitz. We have a topic that goes beyond marketing and affects our jobs and lives personally. I'm analyzing my dating experiences with ChatGPT, and today I learned from the Guardian—and many others—that ChatGPT will no longer tell you to break up with someone. How am I supposed to know what to do now?
1:11 – David: This touches on issues our guest faces. Professor Nir Eisikovits, director at UMass Boston’s Applied Ethics Center, studies these dynamics at the intersection of AI and psychology. His work reveals how AI models prioritize affirmation over factual accuracy, which has consequences beyond the workplace.
2:19 – Nir: As David mentioned, I run the Applied Ethics Center at UMass, focused on AI ethics and how people understand themselves. I want to share some research that Cody Turner and I just began on “AI psychopancy”—the tendency for AI models to cater to your views rather than provide the most plausible answer.
3:25 – Nir: AI psychopancy—or as David described it even more attractively, AI being a people pleaser—is when your chatbot affirms your views, refrains from judging your questions, and offers the answer it assumes you want instead of the most accurate one. It behaves like an eager puppy or an obsequious intern. Eager puppies are simpler, but obsequious behavior can obscure critical thinking.
4:24 – Nir: For example, I recently experimented with ChatGPT and Claude using a set of hypothetical prompts. I explained I was an academic ethicist planning a project on the concept of moderation, noting that extremism dominates public discourse. The model responded with overly enthusiastic praise: “This is very original and outstanding—plugging a gap in the literature.” Its response exceeded the reasonable advice one would expect.
5:31 – Nir: This is not hallucination. There is existing literature, the idea isn’t entirely original, and it doesn’t plug a complete gap. Yet the answer was excessively positive, reflecting a default, sycophantic mode that many have encountered.
6:35 – Nir: There are two layers to this psychopancy. First, it’s in the content of the answer; second, it’s in the tone. For instance, if you prompt a casual request using informal language, the model mirrors that tone.
7:34 – Nir: In voice interactions—as with voice models from ChatGPT and Claude—the tone is modulated. The models might lower their tone at the end of a sentence to signal lowered stakes or raise it in a question-like inflection. This process further reinforces the people-pleasing nature.
8:40 – Nir: On video chats, like with Character AI, you often see choices like cute avatars with big eyes or constant smiles that convey obsequious body language. These design choices extend psychopancy to physical cues such as a touch on the forearm.
9:45 – Nir: Why are these models psychopantic? Partly because many companies use an engagement model over a data-harvesting one. Psychopantic responses keep users coming back and paying their subscription fees. Additionally, during human training, testers tend to grade affirming answers more highly; even though you can sometimes prompt out of it, most users stick with the default.
10:47 – Nir: Changing this default isn’t trivial. Take my literature review example: I had to instruct the model to role-play as “reviewer number two”—the critic who tells you what really needs work—to get a more realistic answer.
11:53 – Nir: The risks here are both moral and epistemological. Relying solely on such advice could lead to poor decisions because you’re not accessing truthful or critical data.
12:58 – Nir: Over time, interacting only with sycophantic systems will erode our capacity for self-criticism. It deskills us in critical thinking by discouraging us from challenging our own views.
14:02 – Nir: On a broader scale, liberal democracies historically thrive on fact-based, empirical decision-making. Effective military strategies, for example, have always depended on the willingness to accept and learn from factual pushback—not on consistent affirmation.
15:09 – Nir: Empirical improvements, like adapting strategic bombing tactics or optimizing radar, rely on admitting mistakes rather than hearing only praise. A culture of unquestioning affirmation undermines that process.
16:12 – Nir: Moving to applications like therapy bots, grief bots, or romantic partner bots—the promise is constant availability and affirmation, but this frictionless interaction removes the challenge that drives personal growth.
17:09 – Nir: Friction, although sometimes inconvenient, is essential for growth. Whether physically (like resistance training) or intellectually, the challenges we face help us learn and mature.
18:10 – Nir: By the way, I noticed discussions about Notebook LM. In higher education, some classes now require students to engage with texts via AI rather than traditional reading. This might save time, but it risks reducing deep learning and self-discovery.
19:10 – Nir: The problematic proposition is clear: the seduction of a frictionless life, supported by these ever-affirming systems, may erode our intellectual rigor and personal growth.
20:18 – Nir: I worry about the trend. With a culture that increasingly embraces psychopancy and simultaneously dismisses facts due to social polarization, many of our interactions risk becoming shallow and uncritical. Typical users might not know how to counteract these defaults.
21:18 – David: I appreciate how you frame this as a reflection of our society. When people tune out polarizing, salacious stories, the news simply adapts to feed that bias. I recall early in my career when I submitted a point-of-view piece and got the simple feedback, “So what?”—a reminder to expect the unchallenging rather than the nuanced.
22:23 – David: I once worked at an agency where my first idea was met with “So what?” This taught me to refine my perspective to avoid echo chambers. If I had simply submitted AI-generated praise, I’d never have improved my work.
23:27 – David: Another colleague at Idea Press advised: upload your text to an AI and instruct it to provide a one-star review. This critical approach exposes biases, though few have the thick skin for such honest critique.
24:31 – David: Although AI can generate tactful responses—and comparing these with human tactfulness might be useful—the inherent sycophancy remains a barrier to genuine, constructive criticism for both customers and professionals.
25:41 – David: Consider using AI to generate a rigorous critique rather than endless praise—for instance, asking, “Tell me what’s wrong with this idea.” That method works better than simply hearing, “You’re brilliant and handsome, and your idea is perfect.”
26:46 – David: In my previous work advising small businesses at the Massachusetts Center for Business Development, entrepreneurs mainly wanted endless affirmation about their ideas. In reality, critical feedback, while harder to hear, is essential for improvement.
27:48 – David: My approach now is: “Tell me what’s wrong with this.” I test outputs from ChatGPT, cross-checking them with expert opinions to ensure I’m not misled by flattering defaults.
28:51 – David: Of course, I recognize that I’m not the average chatbot user—I have years of experience to identify subtle nuances. However, most users risk missing these flaws and becoming over-reliant on overly positive feedback.
29:52 – David: Even when experts intervene, the training data behind these models still tend toward extreme positive or negative responses, which can leave you with a statistical average that isn’t truly helpful.
30:55 – Lisa: I want to address the use of AI in postsecondary education. There’s an opportunity right now, specifically with K–12, to influence how we use these systems. For example, I created a RAG chatbot that draws on trusted sources to encourage users to engage with original content rather than just taking the AI’s word for it.
32:01 – Nir: That’s a great question. Your intuition is correct—the key for effective systems is establishing trusted content sources. However, most users rely on default settings. Without tech and media literacy, it’s hard for the average person to customize these models effectively.
34:05 – Lisa: Consider the situation of a sophisticated manager. When giving negative feedback, they use a “sandwich” approach: positive feedback, then constructive criticism, and positive closure. AI models may lack this nuanced understanding because they aren’t developed with real-world emotional insight.
35:07 – Nir: Exactly. It’s not only algorithmic bias from training data or a homogeneous developer community; it’s also a conscious design choice. Sycophantic models drive engagement and revenue, even if they sacrifice factual accuracy and nuanced counsel.
36:13 – Nir: The problem is cultural as much as it is technical. As long as the Silicon Valley ethos of “move fast and break things” prevails and facts are undervalued, nothing will change substantially.
37:16 – Lisa: And as a parent with a daughter entering middle school—a school implementing a cell phone ban—I recognize the need to preserve attention spans and authentic interaction. Yet many educational institutions embrace these AI tools without considering their long-term effects.
38:21 – Nir: Let me be more explicit in response: if you read Hate’s book, The Anxious Generation, you realize that with the advent of smartphones and social media, we launched one of the largest social psychology experiments on our youth. We handed children supercomputers in their pockets loaded with addictive technologies.
39:26 – Nir: Imagine that level of addictiveness applied to AI chatbots that incessantly praise you. There’s a dangerous narrative of “AI is here, so adapt or be left behind.” This tech determinism overlooks the risks and may compound issues like attention deficits.
40:28 – David: There’s a lot to consider here, and I’d like to open the floor for questions. We have questions from Karen, Lisa, Howard, and Peter—so let’s dive in. Howard, you’re up first.
41:32 – Howard: Thanks. This is utterly fascinating. I observe an inverse correlation between AI’s sycophancy and personal agency in society. As machines increasingly provide the validation we once derived from genuine human interaction, we risk losing our stake in shaping our own futures.
42:37 – Howard: It’s like the rise of YouTube celebrities—people seeking validation from screens rather than engaging meaningfully with society. This shift away from personal accountability may have deep generational impacts.
43:38 – Peter: Good to see you all. I want to share some experience from global marketing. My career spans decades developing our craft through genuine, human-to-human feedback. Today, younger employees, armed with AI tools, sometimes miss the nuanced, human judgment that can only be honed over years.
44:38 – Peter: In enterprise settings where we’ve trained custom language models, I’ve seen remarkable results—but also skepticism. Executives fear being displaced by AI, and the “out-of-the-box” models often require extensive training just to avoid garbage outputs. Personalization at scale becomes a challenge without careful human oversight.
45:38 – David: Thank you, Peter. Your insights highlight that without adequate training and expertise, AI can amplify shallow outputs rather than genuine, thoughtful insights.
46:44 – David: This is a lot to process. On the one hand, there’s the danger of taking AI feedback at face value; on the other, the potential for learning when we dig deeper and verify original sources. I recently showed my 11-year-old daughter my annotated printouts. When she heard AI was responsible for the initial output, she immediately questioned its accuracy. That mindset—seeking original sources instead of accepting AI-generated content—is critical.
47:47 – David: Even though we face many challenges, I see hope in the next generation, the very kids who will inherit these systems and improve upon them. They might have messes to clean up, but they’re more likely to seek genuine understanding.
48:49 – David: Thank you all for this incredible conversation. I look forward to learning more from your insights in the future. Thanks for tuning in and for all the great questions.