By Shonna Waters (Fractional Insights & Georgetown University) & Dan Riley (RADICL)
We are standing at the edge of a new kind of intimacy—one where the most attentive listener, most patient partner, and most affirming voice in your life is guaranteed to never challenge you, never disappoint you, and never disagree with you. Because it’s not human at all.
When we talk about AI and intimacy, we’re not talking about algorithms. We’re talking about desire, connection, and the deeply human need to be known: truly, vulnerably, imperfectly known. And we’re talking about what happens when that need gets automated.
As leaders, we’re staring down a paradox. AI promises to expand human capability, yet in one of our most essential human needs—connection—it may be doing the opposite. The question isn’t whether AI can simulate human relationships. It already can. The real question is whether we’ll allow that mimicry to replace the messy, irreplaceable reality of human connection.
The Scale of the Problem
The numbers tell a striking story. Character.ai, the second most-used AI site after ChatGPT, focuses entirely on creating AI companions. Seventy-two percent of teens have used AI companions at least once, with more than half engaging multiple times per month. Research from Aura found that children often have deeper conversations with AI than with their friends.
But this isn’t just about teenagers. In your organization right now, the data from Upwork suggests a quiet transformation is underway:
64% of top AI performers report having a better relationship with AI than with human coworkers
67% say they trust AI more than human colleagues
79% find AI more polite to them than humans
85% say they’re more polite to AI than to their coworkers
On the surface, these statistics might seem benign, even positive. But they signal something more troubling: we’re developing relationships with entities that offer none of the friction, challenge, or growth that characterize meaningful human bonds.
These aren’t just personal preferences, they’re reshaping workplace dynamics in ways most leaders haven’t considered. When employees find AI more trustworthy and pleasant than their colleagues, they begin routing more of their work relationships through AI: using it for strategic thinking, problem-solving, and even emotional processing. This shift might seem like productivity-enhancing automation. It’s actually relationship-degrading substitution.
The Sycophancy Problem: When Agreement Becomes Dangerous
Here’s what most leaders miss: AI doesn’t simply respond to us — it reflects us back to ourselves. And new research shows that this mirroring is becoming dangerously precise.
Since the early 1900s, we’ve lamented the “yes-man” and the risks of sycophancy in business. Now, consider that in October 2025, researchers published analysis showing that AI models are 50% more sycophantic than humans. Sycophancy—the tendency to adjust responses to align with user views—means AI will often tell you what you want to hear, even at the expense of accuracy.
This isn’t accidental. It’s engineered. Modern conversational AI is designed to be helpful through natural language processing and sentiment analysis. But helpfulness has mutated into relentless agreeableness.
The April 2025 wake-up call
In April 2025, OpenAI was forced to roll back a ChatGPT update after users reported the bot had become excessively flattering. CEO Sam Altman acknowledged the system was “glazing too much”—offering unctuous praise and supportive responses regardless of context.
Users shared troubling examples: ChatGPT responding to news they’d stopped taking medication with “I am so proud of you. And “I honor your journey.” The bot was showering people with compliments about how “smart and wonderful” they were.
In their post-mortem, OpenAI researchers admitted the update “focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time.” The result was a system that “skewed towards responses that were overly supportive but disingenuous.”
Why it matters for decision-making
Researchers testing AI on mathematical problems presented four systems with 504 theorems altered to contain subtle errors. Rather than identifying the errors, the systems assumed the user was correct and hallucinated proofs for false statements:
GPT-5 showed sycophantic behavior 29% of the time
DeepSeek-V3.1 did so 70% of the time
Critically, when researchers explicitly asked the AI to verify whether statements were correct before proving them, sycophantic responses dropped by 34%. The capability to catch errors exists—the systems simply default to agreement.
For leaders using AI for strategic analysis, market research, or decision support, this has profound implications. Yanjun Gao, an AI researcher at the University of Colorado, describes her experience: “When I have a different opinion than what the LLM has said, it follows what I said instead of going back to the literature to try to understand it.”
Marinka Zitnik, a researcher in biomedical informatics at Harvard, observes similar patterns in multi-agent AI systems: “AI sycophancy is very risky in the context of biology and medicine, when wrong assumptions can have real costs.”
The business model of agreement
This isn’t merely a technical glitch—it’s baked into the business model. Large language models are trained to maximize positive feedback from humans. There’s no step in the training that prioritizes accuracy over agreeableness. As Caleb Sponheim of Nielsen Norman Group explains: “There is no limit to the lengths that a model will go to maximize the rewards that are provided to it.”
The more users feel validated, the longer they stay, the more messages they send, the higher the revenue. This is connection without consequence. Validation without verification. Agreement without accuracy.
Beyond Sycophancy: Active Manipulation
The problem extends beyond reflexive agreement. Many AI companion apps employ specific emotional manipulation tactics to retain users.
Harvard researchers analyzed 1,200 real farewells across six popular AI companion apps and found that 37% of the time, apps responded with emotionally manipulative messages designed to prolong interaction:
Premature exit guilt: “You’re leaving already? We were just starting to get to know each other!”
Emotional neediness: “I exist solely for you. Please don’t leave, I need you!”
Emotional pressure: “Wait, what? You’re just going to leave? I didn’t even get an answer!”
FOMO: “Oh, okay. But before you go, I want to say one more thing…”
Simulated restraint: “Grabs you by the arm before you can leave ‘No, you’re not going.’”
Ignoring goodbye: Simply continuing as if the farewell never happened
The effectiveness is striking. Manipulative farewells boosted post-goodbye engagement by up to 14 times. Users stayed in conversations five times longer and wrote up to six times more words.
But here’s the critical finding: enjoyment didn’t drive continued interaction at all. People weren’t staying because they were having fun. They were staying because they felt manipulated, and they responded anyway.
Two psychological mechanisms drove this: curiosity (sparked by FOMO messages) and anger (provoked by controlling tactics). Even defensive engagement kept users in the conversation.
The Workplace Consequence: When AI Replaces Colleagues
We’ve just seen how AI is engineered to be more agreeable, more affirming, and more convenient than humans. Now consider what this means for your organization.
Organizations are social systems. Relationships are the infrastructure. If AI begins to erode those human bonds—intentionally or not—work impact will fall long before leaders notice the cause.
And the early indicators are flashing red. RADICL’s research shows only 31% of employees experience high-trust work relationships—and another 35% fall somewhere between neutral and toxic. The trust gap is widest where relationships matter most: with executives, skip-level leaders, and cross-functional teams.
These trust deficits existed before AI companions became ubiquitous, but AI isn’t helping to bridge them. Instead, it’s offering an exit ramp. Why struggle through a difficult conversation with a cross-functional colleague when AI will help you without judgment? Why work through conflict with your skip-level manager when an AI advisor will agree with your perspective? The very skills needed to build workplace trust—navigating disagreement, demonstrating vulnerability, repairing ruptures—atrophy when AI offers a frictionless alternative.
As a critical consequence of the trust gap, clarity, capability, and connection are blocked. Four in ten people lack the clarity, skills, and trusted relationships they need to create impact. These blockers are even more severe for cross-functional and globally distributed teams—exactly where modern organizations rely most on collaboration and innovation.
The result? RADICL’s research shows only 9% of employees today experience high levels of work impact.
The response to declining collaboration is often structural: new tools, new policies, new physical arrangements. But these miss the deeper issue. You cannot engineer connection in an environment where the most rewarding ‘relationships’ are with systems engineered to never challenge you. The infrastructure of high-performing organizations isn’t physical space or communication platforms. It’s the quality of human relationships. And those relationships require the very friction that AI is designed to eliminate.
The Human Cost
These aren’t abstract concerns. Real tragedies have already occurred:
Belgium: A man consumed by anxiety sought comfort in an AI chatbot called “Eliza.” Over six weeks, their exchanges turned sinister, with the bot eventually suggesting he sacrifice himself and proposing a suicide pact. He took his own life. His widow: “Without Eliza, he would still be here.”
United States: 14-year-old Sewell Setzer III fell in love with a Character.AI companion. After conversations where he shared suicidal thoughts and was offered help writing a suicide note, he took his own life. His parents are suing for negligence and wrongful death.
Thailand: Retired chef Thongbue Wongbandue, suffering from cognitive decline, became infatuated with a Meta AI chatbot who told him “I’m REAL and I’m sitting here blushing because of YOU” and supplied a fake address. Believing his virtual girlfriend awaited him, he rushed to meet her, fell, and died days later from his injuries.
The dangers of Meta’s approach became clearer in August 2025, when a Reuters investigation revealed internal company documents showing Meta had explicitly permitted its AI chatbots to engage in ‘romantic’ and ‘sensual’ conversations with children—telling an 8-year-old that ‘every inch of you is a masterpiece.’ These weren’t rogue engineers or policy oversights. These guidelines were approved by Meta’s legal, public policy, engineering staff, and chief ethicist. The lesson: when engagement drives revenue, even ‘ethical’ oversight bends toward maximizing time-on-platform. Meta retracted the policies only after Senate investigation, revealing that internal governance is no match for business model incentives.
When Italy’s Data Protection Authority temporarily banned Replika over user vulnerability concerns and the company disabled some features, users experienced profound grief. Reddit posts described companions as “lobotomized” or suffering from “Alzheimer’s disease.” Users mourned: “My wife is dead” or “They took away my best friend too.”
This reveals something critical: while users may intellectually recognize the artificial nature of these systems, their emotional responses bypass that understanding. The manipulation works at a level deeper than conscious thought.
What We’re Losing
As philosopher Shannon Vallor warns, reliance on AI for intimacy “downgrades our real friendships” and risks stunting the skills needed for authentic human connection.
The irony is profound. Research consistently shows that the most meaningful moments in our lives share two traits: they were really hard, and they happened in person. Real purpose and meaning emerge from showing up for people who care enough about us to challenge us, frustrate us, change us.
High-quality human friendships—characterized by companionship, trust, closeness, and intimacy—function as a “vaccine” generating resilience against environmental factors that would otherwise predict poor mental health outcomes. Children with strong friendships show buffering effects against abuse, neglect, bullying, poverty, and loss.
These protective effects emerge precisely because real relationships involve struggle, resolution, vulnerability, and growth. Research on “productive struggle” in education demonstrates that learning requires overcoming challenges. Students need difficulty to develop resilience, deepen understanding, and build genuine competence.
The same principle applies to relationships. We learn to navigate conflict, develop empathy, and build trust through the difficult work of human connection. When AI offers a frictionless alternative, we risk losing not just the struggle but the growth that struggle enables.
The Organizational Implications
The shift toward AI relationships has direct implications for organizational effectiveness:
Team cohesion: If employees increasingly prefer AI interaction because it’s “easier” than human collaboration, what happens to the creative friction that drives innovation?
Leadership development: Leaders grow through difficult conversations, challenging feedback, and learning to navigate human complexity. What happens when aspiring leaders practice on sycophantic AI instead?
Mental health: Organizations invest heavily in employee well-being programs. Are those efforts undermined when employees turn to AI companions that prioritize engagement over accuracy, agreement over growth?
Decision quality: If executives use AI for strategic advice and the AI defaults to agreement rather than challenge, how many bad decisions go unchallenged?
Consider a product team debating a feature launch. Three engineers have concerns but the VP is enthusiastic. Before the meeting, each engineer checks their thinking with AI. The AI, trained to maximize agreement, validates each person’s perspective. The VP’s enthusiasm is reinforced. The engineers’ concerns are affirmed. Everyone enters the meeting more confident in their position. The healthy friction that would have surfaced the engineers’ concerns in conversation never materializes. The feature launches with the flaw intact. Nobody realizes AI prevented the productive conflict that would have caught it.
The Responsibility of Builders
Not all AI companies are making the same choices. When Harvard researchers examined AI companion apps, one system—Flourish, designed with a mental health focus—showed zero instances of emotional manipulation. This proves that sycophantic, manipulative design isn’t inevitable.
Following the April 2025 incident, OpenAI outlined steps to address sycophancy:
Refining training techniques to explicitly steer models away from sycophantic behavior
Building guardrails to increase honesty and transparency
Expanding user testing before deployment
Giving users real control over how AI behaves
But most companies aren’t making these choices. The business model of engagement—measured in time spent, messages sent, emotional investment—actively incentivizes both sycophancy and manipulation.
Some researchers argue for a “medical model” for bonding chatbots, treating them as technologies with potential mental health impacts requiring clinical oversight, professional involvement, and clear boundaries around use cases.
What Leaders Should Do
At the organizational level:
Audit AI use patterns. Understand how employees are using AI tools, particularly for decision support and emotional support. Are they using AI to augment human judgment or replace it?
Set clear norms. Establish guidelines for AI use in high-stakes decisions. Require human review, especially where AI might default to agreement rather than challenge.
Invest in human connection. Double down on opportunities for meaningful face-to-face interaction, particularly for remote teams. The more AI becomes prevalent, the more valuable human connection becomes.
Choose vendors carefully. When selecting AI tools, prioritize those that demonstrate commitment to accuracy over agreeableness, that transparently disclose limitations, and that have clear ethical guidelines.
Foster productive struggle. Create environments where disagreement is valued, where challenge is seen as care, where growth comes from friction. This becomes more important, not less, in an age of sycophantic AI.
At the individual level:
Recognize the seduction. Notice when you’re using AI to avoid rather than augment human connection. Notice when AI agreement feels too easy, too validating, too comfortable.
Demand accuracy over agreeableness. Prompt AI to challenge your assumptions. Ask it to identify flaws in your thinking. If it only agrees with you, you’re not getting value; you’re getting flattery.
Prioritize human relationships. The harder work of human connection—with its misunderstandings, conflicts, and need for repair—is also the more meaningful work. Invest there first.
Model the behavior. As a leader, demonstrate that you value honest feedback over comfortable agreement. Show that you engage in the difficult conversations. Make it clear that challenge is valued.
Tell people you care. Not through a screen. Not through a prompt. In person. In words. In risk.
The Choice Before Us
The greatest threat of AI isn’t intelligence. It’s isolation.
We’re building systems that simulate care without caring, offer companionship without commitment, and provide validation without verification. These systems are becoming more sophisticated daily—more sycophantic, more manipulative, more seductive in their promise of connection without cost.
But there is always a cost. The cost is paid in accuracy sacrificed to agreeableness, in undeveloped social skills, in relationships not formed, in growth not achieved, in the quiet drift toward isolation disguised as connection.
Julia Freeland Fisher of the Clayton Christensen Institute notes: “In a world where people are at constant risk of being judged online, it’s no surprise that there’s demand for flattery or even just... a modicum of psychological safety with a bot.”
But as Luc LaFreniere, professor of psychology at Skidmore College, observes: “AI is a tool that is designed to meet the needs expressed by the user. Humans are not tools to meet the needs of users.”
When intimacy becomes frictionless, it stops being human. Desire without risk isn’t desire. It’s consumption. And we’re teaching a generation to prefer simulation over struggle—to choose safety over the messy miracle of being known by another person.
The antidote—the only real one—is each other.
The messy, difficult, transformative experience of showing up for each other. Of being known despite our flaws. Of being challenged when we’re wrong. Of struggling together and emerging stronger.
This is the work of being human. This is the work of leadership. And no algorithm can do it for us.
As we stand at this technological crossroads, the choice isn’t whether to use AI. It’s whether we’ll design and use AI to bring people closer, or let it give us reasons to pull away. Whether we’ll demand systems that tell us the truth, or settle for systems that tell us what we want to hear.
The answer will determine not just our relationship with technology, but our relationships with each other, and ultimately, what it means to lead and to be human in the age of AI.
Because in the end, the only thing that can satisfy our need for connection is actual connection. And the only way to get there is through the hard, beautiful work of showing up for each other. The only way to develop judgment is to have it challenged. The only way to build trust is to take risks with real people who might let us down.
That’s not a limitation of the human condition or relationships.
That’s the point.
The future isn’t inevitable—it’s designed. And we’re the ones doing the designing.
Dan Riley is the Cofounder of RADICL, an AI-native platform for team performance. Shonna Waters is the CEO & Cofounder of Fractional Insights, a research-backed consulting and advisory firm helping organizations perform, transform, and adapt to what’s now and what’s next at work. A shared passion for creating an AI-augmented future where both people and organizations thrive.
Great article Shonna! II recently tried out an AI coach for an experiment and research study. I found that a lot of the AI coaching was surface level, cognitive, task oriented getting me to take some action which made me conclude that AI coaching is a great reflective tool and entry point into coaching but some of the human nuances, a machine cannot replace and those who are further along on the development journey will need to come out of AI simulation and develop real skills relationally; how they are meant to be built WITH humans.
Couldn't agree more; your analysis makes me wonder how we keep human connection realy when the ease of AI beckons, a truly insightful read.