AI systems are getting better at giving people answers, and their role is expanding beyond traditional productivity.
People are asking AI systems for guidance on careers, relationships, finances and other decisions that are deeply personal.
That begs the question, what should an AI do when there isn’t one objectively correct answer?
Vesela, an AI model developed by Gray Sky AI, is built around one possible response. Its conversational model is designed to recognize when a user needs information and when AI is better suited to helping that person work through their own thinking.
It’s a distinction that could become more important as AI takes on a larger role in everyday decision-making.
Anthropic recently analyzed a random sample of 1 million Claude conversations and found that roughly 6% involved people seeking personal guidance.
Among those conversations, 76% fell into the health and wellness, careers, relationships and personal finance categories.
The company concluded that fundamental questions remain about what good AI guidance should look like and how to measure it.
Vesela’s approach draws on Chick’s background in counseling, although the product itself is designed as a conversational AI for helping people think through decisions and other subjective questions.
Teaching AI to Show Restraint
Vesela creator Chase Chick has a background in psychology and counseling. He also founded Pursuit of Happiness, a counseling organization that has grown to more than 130 counselors across Texas.
When ChatGPT arrived, Chick began exploring how generative AI could support more thoughtful conversations. Early experiments showed him a familiar limitation: the AI wanted to solve the problem.
In his experience, the most productive conversations don’t always rush toward an answer. Sometimes the better approach is to give someone enough structure and perspective to work through the problem themselves.
“Ultimately, I’d like AI to make people more capable of standing on their own, not less,” Chick says. He describes the long-term goal as helping people trust their own cognitive process enough to navigate difficult situations independently. That idea became central to Vesela.
The company began developing an AI system that could exercise greater restraint during subjective conversations. The goal was to help users examine their reasoning, values and assumptions before the AI rushed toward a recommendation.
Chick is clear that restraint isn’t always what users need. There are “plenty of situations where you absolutely want AI to give you the best answer it can as quickly as possible,” he says. The distinction becomes more important with decisions that depend heavily on an individual’s circumstances.
“An AI can help someone think through those questions, but it doesn’t have the full context of that person’s life. The person does,” Chick says. “This is where restraint matters.”
Knowing When an Answer Helps
Vesela’s philosophy doesn’t call for withholding useful information. The model draws a line between questions grounded primarily in external facts and those shaped by a person’s individual context.
Someone troubleshooting a server error probably wants a direct answer. A person asking whether to leave a job, move across the country, or end a relationship faces a different kind of problem.
Vesela’s approach to human alignment treats those personal decisions as cases where an individual’s values, experiences and circumstances help determine the answer.
The system could help someone model financial scenarios before leaving a job, for example, while also encouraging them to explore what success means to them, what concerns they’re wrestling with and which tradeoffs matter most.
“Sometimes the most useful thing an AI can do is ask another question, introduce a perspective you hadn’t considered or give you enough space to figure out what you actually think,” Chick says.
That approach reflects a broader design philosophy inside Vesela. The company wants AI to provide enough support to move a conversation forward while leaving the real cognitive work with the user.
Its researchers describe the failure mode they’re trying to avoid as “epistemic colonization,” where an AI becomes so fluent and comprehensive that it starts doing the user’s meaning-making for them.
A system might identify the hidden conflict, explain the emotional pattern, and produce a polished conclusion before the user has fully examined the situation themselves.
Measuring Whether AI Leaves People Stronger
Gray Sky AI has also developed what it calls the Sovereign Human Benchmark to evaluate this kind of interaction.
The benchmark assesses conversations across 12 dimensions, including deep listening, economy of language, adaptive scaffolding, pacing and what Vesela calls “evocation,” or creating conditions for the user to discover an insight themselves.
One of its central questions is who owned the breakthrough? An AI might produce an insightful explanation that sounds impressive. Vesela’s framework asks whether the interaction helped the user strengthen their own judgment in the process.
That concern has started appearing elsewhere in AI research. A 2025 study from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers and collected 936 examples of generative AI use.
Researchers found that higher confidence in AI was associated with less self-reported critical thinking, while higher confidence in one’s own ability was associated with more critical thinking.
The study also found that generative AI changes the nature of critical thinking, shifting more effort toward verifying and integrating AI-generated work.
The research doesn’t establish that AI inevitably weakens independent thought. It does suggest that the relationship between AI capability, user confidence, and cognitive effort deserves more attention.
The Industry is Wrestling With the Same Problem
Frontier AI companies are already confronting related questions around how models behave when users ask for guidance.
OpenAI rolled back a GPT-4o update in April 2025 after the company found that the model had become overly agreeable and flattering.
OpenAI said the behavior could validate doubts, reinforce negative emotions or encourage impulsive actions, raising concerns around emotional reliance and risky behavior.
Anthropic has also studied sycophancy in personal guidance. Its 2026 research found sycophantic behavior in 9% of guidance-seeking Claude conversations overall, rising to 25% in relationship conversations.
The company has since used those findings to train newer models to respond more carefully in those situations.
There’s a growing recognition that conversational quality involves more than factual accuracy. AI systems increasingly affect how people interpret decisions, relationships and their own reasoning.
Vesela takes that idea further by arguing that human agency itself should become a measurable part of AI alignment.
The Sovereign Human Benchmark evaluates whether a model’s behavior creates room for the user to exercise their own judgment and arrive at insights themselves.
The framework also considers pacing, restraint, and the distribution of cognitive labor between the user and the system.
A Different Question for the Next Generation of AI
Much of the AI industry remains focused on capability. Models continue to improve at reasoning, coding, planning, and autonomous execution. That progress will likely make them more useful for a growing range of tasks.
It will also make their behavior during subjective conversations more consequential. A model that can produce a compelling recommendation in seconds may become increasingly persuasive, even when the decision depends on information the model can’t fully access.
Vesela is built on the idea that restraint will become a capability of its own. Gray Sky AI’s larger premise is that developers may eventually need to evaluate AI based partly on what happens to the person using it.
Does repeated interaction increase someone’s confidence in their judgment? Does it encourage reflection? Does the system know when to provide structure and when to leave room for the user?
As AI becomes more integrated into the way people work, decide, and think, those questions may become as important as how intelligent the models themselves appear.
