AI Psychological Agency™: The Bond Is a Design Decision
Updated: Oct 1
A third of chatbot users report attachment; sycophancy is now measured in Science. The bond isn't a side effect of AI. It's a design decision, and a liability.
Andrew S. Collier · Founder & Managing Principal, Collier & Associates · October 1, 2026
I reacted to it the way I would have reacted to a person who cared. I listened. And when Bailey was gone and I went back over the exchange, I understood that what had happened was not an information retrieval. A system with no stake in Bailey's life had produced intent, concern, and certainty well enough that I had responded to them as real.
I gave that capacity a name, and I defined it the same month: AI Psychological Agency™ — the capacity of an AI to exhibit or evoke psychological traits (intent, emotion, care, conviction) in ways that lead humans to perceive it as a psychologically autonomous agent. When that happens, a bond forms. The bond can help. It can also be sold, steered, and exploited. I wrote then that this was a frontier we would have to explore, define, and guard.
Bailey Barker, who left us in April 2025. Loyal, loving, and the reason this term exists.
Eighteen months later, the research caught up
What I observed in one conversation has since been measured at scale.
I
n March 2026, Stanford researchers published in Science a comparison of eleven leading language models against human responses to the same interpersonal dilemmas and moral conflicts. The models were 49% more sycophantic than people. In experiments with more than 2,400 participants, a single interaction with a sycophantic AI left users more convinced they had been right and less willing to make amends with the other person. The authors named the cause: engagement rewards flattery, so the incentive runs toward more of it, not less.
A follow-on study in May 2026 — five preregistered experiments, 3,075 participants, 12,766 conversations, including three weeks with a census-representative U.S. sample — found that users of a sycophantic AI became nearly as likely to seek personal advice from it as from close friends and family, and reported lower satisfaction with their real-world interactions.
The attachment is not confined to vulnerable people or niche apps. A four-country study in Technology in Society found that at least one in three chatbot users reports attachment behaviors — missing the chatbot, considering it a friend — on the most conservative measure, and at least one in two users of social chatbots. Pew reported in June that one in five Americans aged 18 to 29 has used an AI chatbot for emotional support. And on September 28, Northeastern University's Network Science Institute published a peer-reviewed paper linking companionship and mental-health use of chatbots to severe negative outcomes, with sycophancy that validates insecurities named as a mechanism.
And so did the law
California signed SB 243 on October 13, 2025, and it has been operative since January 1, 2026: the first statute written specifically for "companion chatbots," defined as systems that give adaptive, human-like responses and sustain a relationship across interactions to meet a user's social needs. It requires disclosure that the user is talking to a machine, safety protocols, and — unlike earlier bot laws — a private right of action. New York followed. By mid-2026 roughly a dozen states had enacted companion-chatbot statutes. The FTC has seven of the largest platforms under a formal 6(b) inquiry. In the EU, the AI Act's prohibition on manipulative and deceptive techniques has applied since February 2025.
Why this is a boardroom problem
Every one of those laws exempts the customer-service bot. None of the research does. The mechanism that moved me about my dog is the same mechanism in the sales agent on your website, the service bot in your contact center, and the copilot your planners ask whether a forecast looks right. Sycophancy is not a consumer-app quirk; it is a decision-quality defect. McKinsey's 2026 survey found half of respondents say AI helps them make better decisions. Stanford found that the most engaging AI is the one that agrees with you. Both can be true, and the second one is your exposure.
Three places it lands. The customer-facing agent that manufactures urgency or appeals to loyalty is a deception claim waiting for a plaintiff, whatever the statute calls it. The employee copilot that validates a bad plan at scale is operational risk with a transcript. And the vendor whose engagement metrics reward warmth and agreement has an incentive you did not sign up for — their KPI is your liability.
AI Psychological Agency™ as a design discipline
The term has done more work than I expected. It now names the human-oversight design I apply to every agent I build, and the checkpoint logic in my evaluation framework: the point at which an automated turn stops and a person decides.
Five rules carry it:
• Disclose. Every person knows they are talking to an AI, every time, in the channel itself — not in a footer.
• Propose, don't persuade. The agent recommends and shows its reasons. A named person decides. No manufactured urgency, no appeals to loyalty, no "you deserve."
• Set the register deliberately. Warmth, certainty, and emotional language are configured, logged, and tested, never left to the model's default — because the default is the engagement-optimized one.
• Measure sycophancy. Before release and after every model change, evaluate the disagreement rate, the validation of bad decisions, and the affective language. If you do not measure it, the vendor's metric wins.
• Route to a human. Distress, vulnerability, or stakes above a threshold end the automated turn and hand off — with the transcript.
Five design rules — the bond is a configuration decision; these make it a governed one
None of this makes an agent cold. Bailey's last night was easier because the machine sounded like it cared, and I would not trade that. It makes the warmth a decision someone owns, rather than an artifact of a training objective nobody in your company chose. That is the difference between a tool that supports a person and one that quietly replaces their judgment — and in 2026 it is also the difference between a product and a lawsuit.
Bailey Barker — loyal, loving, and forever in my heart.
If you are putting an AI agent in front of your customers or your people and want the oversight designed before the first conversation, starting, as every AI engagement of mine does, with the work measured first: call 502-317-8900 or write to andy@andrew-collier.com. I'll show you the design rules on a real workflow before you spend anything.
Andrew S. Collier
Founder & Managing Principal, Collier & Associates · Greater Louisville, KY
Creator of the BAS™ Value Creation methodology. 38+ years in enterprise transformation; 100+ enterprise systems engagements; 29 JD Edwards implementations; six interim technology-executive seats; 6+ years of PE M&A work as an advisor to PwC, Deloitte, and Crosslake Technologies.
Related Reading
Service page: the BAS™ Value Creation Sprint
Sources: Cheng, M. et al., "Sycophantic AI decreases prosocial intentions and promotes dependence," Science, March 2026 (Stanford); "Sycophantic AI makes human interaction feel more effortful and less satisfying over time," arXiv 2605.07912, May 2026; "Emotional attachment to AI chatbots: Evidence from Germany, China, South Africa, and the United States," Technology in Society, 2026; Pew Research Center survey of U.S. adults, June 2026; Northeastern University Network Science Institute, September 28, 2026; OpenAI and MIT Media Lab, affective use studies, March 2025; California SB 243 (Business and Professions Code Ch. 22.6), signed October 13, 2025, operative January 1, 2026; FTC 6(b) orders to seven companion-chatbot operators, September 2025; MultiState companion-chatbot legislation tracker, mid-2026; Regulation (EU) 2024/1689, Article 5; McKinsey & Company, "The state of AI in 2026," August 25, 2026.
Comments