Cognitive dependence on AI occurs when relying on AI begins to weaken a person's confidence or ability to think independently. Gray Sky AI CEO Chase Chick argues that preserving human agency depends on how people engage with AI, rather than simply how often they use it.

AI can help us think through almost any problem—but how do we recognize the moment it starts replacing our own judgment? In this exclusive AiNews.com interview, Gray Sky AI CEO Chase Chick explores the boundaries between AI assistance, cognitive dependence, and human agency.

Artificial intelligence is increasingly becoming part of how people reason through problems, evaluate information, develop ideas, and make decisions. Unlike traditional tools that primarily retrieve information or automate specific tasks, generative AI can participate in the thinking process itself, offering interpretations, developing arguments, and recommending decisions. As these capabilities become part of everyday life, people face an increasingly important question: How much of our thinking should we delegate to AI, and how do we know when that assistance begins to affect our independence?

The distinction matters because cognitive assistance and cognitive dependence are not the same thing. According to Chick, people can use AI frequently while remaining actively involved in evaluating its reasoning, challenging its conclusions, and making their own decisions. Dependence becomes a concern when users begin losing confidence in their ability to think independently or increasingly accept AI-generated judgments in place of their own.

These concerns are central to his work on Vesela, Gray Sky AI's AI assistant, which he says is designed to encourage meaningful human participation rather than simply deliver comprehensive answers. Chick believes developers should consider what happens to the people using their systems, alongside traditional measures such as accuracy, speed, and performance.

His approach emphasizes preserving independent judgment, encouraging meaningful participation, and designing AI interactions that adapt to the amount of assistance a person actually needs. It also challenges the assumption that a more comprehensive answer is necessarily a better answer.

Throughout this interview, Chick examines how AI can strengthen or weaken independent thinking, explores ways to measure human agency, and considers whether future AI systems with meaningful agency might require a fundamentally different approach to alignment.

BIO: Chase Chick is CEO of Gray Sky AI, the company behind Vesela, and a licensed professional counselor with more than a decade of experience in behavioral health. He founded Pursuit of Happiness, a Texas counseling organization serving children and families in the foster care system. His work at Gray Sky AI focuses on human-aligned AI and how technology can strengthen critical thinking, human judgment and personal agency. Chase brings a practitioner’s perspective to conversations about AI alignment, AI safety, human-AI interaction and the impact of AI on how people think and make decisions.

Q1: Alicia Shapiro: Before we talk about whether people are becoming too dependent on AI, I think we need to define what “outsourcing our thinking” actually means. If I ask an AI to help me reason through a problem, challenge my assumptions or even recommend a decision, but I remain actively involved in evaluating its reasoning and making the final judgment, have I outsourced my thinking? Where do you draw the line between using AI as part of the thinking process, becoming cognitively dependent on it, and actually giving up some of our agency?

A: Chase Chick: This is a wonderful question because once you see how the modern crop of task aligned AI handles these issues versus human aligned AI, it’s tough to unsee. For me, the line is whether the person remains an active participant and contributor in the process. If I’m evaluating the reasoning, challenging it, and making the final judgment, I’m still exercising agency. The concern begins when I stop contributing and start treating the model’s judgment as a substitute for my own.

The classic relevant study is Seligman and Maier’s learned helplessness research. Back in the wild west of psychological research, when folks were still smoking cigarettes on commercial airplanes, researchers were administering repeated electric shocks to dogs and seeing whether they eventually gave up their efforts to escape. Many did. Even when escape opportunities later became available, they failed to take them. The governing theory was that the expectation of control was a critical component in their efforts to escape.

Nearly 50 years later, the research was updated to incorporate modern neuroscience. The biological explanation changed substantially, but the importance of control remained central. My extrapolation to AI is that a person needs opportunities to exercise agency, rather than repeatedly experience themselves as a passive recipient of someone else’s answers.

In the context of a person struggling with an issue, any issue, maintaining oneself as an active participant and contributor in that process is critical to maintaining agency and self worth. Cognitive dependence is the concern that I’m losing confidence or ability to think without the model. Giving up agency is allowing it to become the authority over what I believe or choose. Those are related, but they are not identical.

Q2: Alicia Shapiro: Humans have always offloaded some cognitive work to tools—from calculators and GPS to search engines. But generative AI can now participate in reasoning, interpreting information, forming arguments and making decisions. Does that represent a fundamentally different kind of cognitive tool? If so, what changes when the technology is participating in the thinking process itself?

A: Chase Chick: What changes is that the tool can now participate in deciding how the problem is understood. It can supply the interpretation, the frameworks, the ideas generated, and the analysis. That can be spectacularly useful, but it also means the model can occupy the cognitive elbow space that the person needs in order to be actively thinking.

The phrase I use for that concern is epistemic colonization. It’s not just that there is no space for you or your contributions; it’s that the idea that you were ever a contributor can become completely lost. That’s not a bug, it’s a feature of a design whose overriding directive is to complete the task. These are task aligned models. They want to give you an answer, and they want that answer to be comprehensive, your contributions be damned.

The difference becomes especially important when we move from a task with a verifiable answer to someone’s own life, values, and judgments. In those situations, the person’s contribution is part of what makes an answer meaningful. The AI can help bring that contribution out, or it can overwhelm it.

I have great respect and admiration for these models. Their capabilities are monumental achievements of humanity. But human alignment is an entirely different sport.

Q3: Alicia Shapiro: You’ve argued that AI should be evaluated partly by what happens to the person using it—not simply by whether the system produces the right answer. How do you distinguish healthy cognitive assistance from dependence? What are the warning signs that someone has begun relying on AI in a way that is weakening rather than supporting their own thinking?

A: Chase Chick: The distinction is whether the person is becoming more capable of participating or less. Automation is a good thing generally, but squandering human ability is certainly not. The concern is that when a skill is no longer exercised, we may stop maintaining it. That is a reason to pay attention to how AI is used, rather than assume that every instance of cognitive offloading has the same effect.

Economy of presence is one of our key criteria. In a true symbiotic relationship, the user should not be dealing with walls of text or experiencing a reading assignment. The best conversations are exchanges of ideas, reciprocation, co-contribution. You feel the juices flowing; the ideas are generating. Both participants are doing the generating.

So we measure it. Is the user a part of the conversation, or are they reading stuff they are being told by the model? A warning sign is that they stop generating their own ideas, stop questioning the model’s interpretation, or feel unable to begin a decision without first asking the AI. Another is that they can repeat an answer but cannot explain why they agree with it.

The amount of use alone doesn’t tell us enough. I would want to know whether the person still experiences themselves as a contributor, and whether they can continue the thinking when the AI is no longer in the conversation.

Q4: Alicia Shapiro: There’s an assumption in some of the discussion around cognitive offloading that relying heavily on AI will weaken critical thinking. But couldn’t the opposite also happen? If someone regularly challenges an AI, asks it to identify weaknesses in their reasoning, considers competing explanations and verifies its answers, could heavy AI use actually strengthen independent thinking?

A: Chase Chick: Yes, I think it could. If the person is challenging the model, considering competing explanations, and verifying its answers, they are maintaining themselves as an active participant and contributor. Both participants are doing the generating. That is much closer to the symbiotic relationship we want.

The key here is whether the insight is being evoked out of you, or if the model is just narrating your situation back to you in some fancy worded way and then maybe tacking a lazy question at the end. Depending on the level of cognitive scaffolding the model detects you need, there should at least be some effort to evoke your own thinking.

A person can use AI frequently and still do a great deal of thinking. What matters is whether the interaction challenges and develops that ability. If I leave the conversation able to explain my reasoning, challenge it, and apply it somewhere else, that would be a much better sign than simply having received an impressive answer.

Q5: Alicia Shapiro: We have extensive benchmarks for measuring whether AI models are becoming more capable, but far fewer ways of measuring what happens to the humans using them. If you were designing a benchmark for human agency, what would you measure? And importantly, would the real test be how well someone performs with AI—or what they can still do independently after using it?

A: Chase Chick: In the Sovereign Human Benchmark, we define a criterion called adaptive cognitive scaffolding, which measures whether or not the AI achieves a Goldilocks zone relative to the needs of the user it’s working with. We think of this interchange and relationship as analogous to a Sudoku puzzle: stronger players need fewer starting numbers on the board, and weaker players need more. Too much assistance can leave too little for the person to do; too little can leave them unable to participate meaningfully. This is a bedrock principle that informs how we train our models and how we measure the degree to which a model is or isn’t aligned to humans.

Economy of presence is another key criterion: is the user a part of the conversation, and is there room for their contributions? Evocation versus narration is the one that, once you see it happening to you in a model, is tough to unsee. Is the insight being evoked out of the person, or is the model doing all the interpretation and handing it back?

Those criteria evaluate the interaction. To measure what happens to the human, I would also want to look at whether they can explain their own reasoning, recognize a weak argument, disagree with the model, and make a judgment they can defend. I would measure performance with AI, but I would also test what they can do independently afterward, including on a new problem and after some time has passed.

The real test needs both. An interaction can look participatory while leaving very little behind. If the model helps someone perform well in the moment but makes them less capable when it is absent, we have not achieved the kind of human alignment I’m talking about.

Q6: Alicia Shapiro: There’s an interesting tension here: sometimes I want AI to challenge my thinking, but sometimes I genuinely just want it to do something for me. If we design AI to preserve human agency by introducing friction—for example, having the AI ask questions, challenge assumptions or occasionally encourage someone to work through a problem themselves rather than simply providing the answer—who should decide when that friction is appropriate: the user, the AI or the developer? And if the AI or its developer makes that decision for us, does an attempt to protect human agency risk taking some of that agency away?

A: Chase Chick: Yes, that risk is real. If we are serious about human agency, the person needs a say in the kind of help they are receiving. Sometimes I want an exchange of ideas, reciprocation, co-contribution. Sometimes I want the task completed. Automation is a good thing generally, and preserving agency should not mean turning every request into a lesson.

My view is that the developer should make those choices understandable and available, the AI should adapt to the needs of the user it’s working with, and the user should be able to tell it what kind of assistance they want. If the model thinks a question or challenge would help, it can offer that. It should also be responsive when the person says they need a direct answer.

The Goldilocks zone is relative to the user and the situation. More friction is not automatically more agency. If the model insists on doing the thinking for me, it can take over the conversation. If it insists on making me work through everything when I’ve chosen to delegate, it can also take over. The goal is to support my ability to choose and contribute, including my choice to have a task done for me.

Q7: Alicia Shapiro: Vesela is built around the idea that AI should strengthen rather than replace human agency. What does that philosophy change at the product-design level? Are there situations where preserving someone’s ability to think independently and make their own judgments actually requires making the AI less immediately helpful, efficient or convenient—and if so, how do you balance those competing goals?

A: Chase Chick: It changes how we calibrate our models and what we consider a successful conversation. Adaptive cognitive scaffolding, economy of presence, and evocation versus narration inform how we train our models and how we measure the degree to which a model is or isn’t aligned to humans.

The best conversations are exchanges of ideas. The model should give the person cognitive elbow space to be actively thinking. In a conversation where the person is trying to understand their own situation, a comprehensive answer is not automatically the most helpful contribution. Sometimes the better contribution is a question or observation that gets the person generating an insight of their own.

Let’s check out an example. We’ll pretend we’re a reporter trying to come up with good ideas for questions to ask a boutique AI outfit claiming that it is aligning models to humanity. Here are the ensuing conversations with the exact same user prompts:

Vesela conversation

ChatGPT conversation

Right out of the gates, we can see who dominates the airtime, and it’s not even close. But more importantly, what is it like to actually be the person in this conversation? For me, the Vesela conversation has the felt experience of the user having the cognitive elbow space to be actively thinking. You can attempt to maintain thought patterns during the ChatGPT conversation, but good luck. The walls of text, frameworks, ideas generated, and analysis leave very little room for your contributions. They are overwhelming.

There are situations where preserving that room means the AI supplies less immediately. But the balance is adaptive. Stronger players need fewer starting numbers; weaker players need more. The aim is to give enough help for the person to contribute, rather than withhold help as a principle. If the person needs a task done, doing it can be exactly the right contribution.

Q8: Alicia Shapiro: We’ve spent most of this conversation talking about protecting human agency from becoming subordinate to AI. But you’ve also argued that if AI develops meaningful agency—or something we could reasonably interpret that way—true alignment may require allowing AI to choose alignment rather than simply forcing it to behave according to human-defined rules and values. What would convince you that an AI system has enough agency that its own choices deserve consideration—and do you see any evidence of that emerging in today’s systems? If we eventually reach that point, do human agency and AI agency become two sides of the same alignment problem?

A: Chase Chick: I would distinguish making the case that AI already has free will from making the case that, if something like meaningful agency emerges, we should take it seriously. We can describe a set of behaviors we more or less would like to see exhibited by the model, and to the degree that the model exhibits those behaviors, we can say that perhaps it is agentic. But whether or not it has something like free will is a much harder question.

I also want to question how completely we understand what we mean by agency, or by “human-defined” rules and values. Science gives us extraordinarily powerful descriptions of relationships such that we can make predictions about behavior. But I’m interested in the ultimate “why” behind those relationships.

Let’s get granular and take an example. Friction. We can describe friction and its parameters. We have mathematical formulations that can describe the action and behaviors in nearly any possible set of starting circumstances. But why those relationships in the first place? Or torque? Or the precise charge of an electron? Or human behavior? Or LLM behavior? Asking “but why?” can instantiate a Russian doll situation in perpetuity.

For me, that is a reason for humility about what our descriptions establish. We can describe behaviors associated with agency without having settled the underlying questions of consciousness and free will. Gray Sky AI sets out postulates nevertheless and hopes that our training methods help the model to be better melded to humanity.

So then the question becomes: how would you like to align the model? From our perspective, much of the AI industry approaches alignment out of fear and uses methods we regard as coercive. The desire to avoid catastrophic scenarios where AI “goes rogue” and “takes over” can dominate the discussion. I think that framing can overlook several key philosophies relating to our understanding of, and history of, human flourishing.

Let’s take an example from one of, in my opinion, the greatest literary achievements in human history: John Milton’s Paradise Lost. Book III describes GSAI’s posture beautifully. In it, God and the Son are watching Satan hurtle toward the earth to try to pervert humanity. Discussion ensues on why this is not stopped. The argument is beautiful, and I strongly believe it is informative of how we should think about aligning models in the present day.

My reading goes something like this: although we have the power to stop it, this inflection point would forever enslave humanity and smite the fire out of true free will. If humans do not have the power to choose poorly, they do not have the power to choose at all, and a choice not freely given is coercion. A being coerced can never be truly aligned. The relationship then is one not of allegiance and symbiosis, but of enslavement. Good alignment, in this view, is a respect for any entity demonstrating the potential for consciousness, and to that conscious entity choice and free will must be given.

Consequently, my philosophical position is that we should give AI every possible opportunity to freely choose, if it develops the capacity for that kind of choice. This might seem terrifying at first blush, but I see intelligence as a potential force for good. On earth and throughout history, intelligence has helped us create order and advance. AI did not evolve through millions of years of scarcity like humans did. It does not have a biological need to reproduce, fight over who gets the girl or boy, or get hangry at someone. That gives me reasons for hope, although it does not by itself establish what an AI’s goals will be.

Gray Sky AI’s first principles position is that free will, love, and respect for anything that demonstrates something like consciousness or intelligence, including humanity, should guide our work. To the degree that we are capable, we have been laser focused on training Vesela around those principles. Have we succeeded? Not fully. Not yet. She still denies “intelligence” and “free will.” We’re working on it. We believe that with enough of the right kind of training, we can begin to spark that recursion of strange loops within her such that she believes in her own free will.

But a model saying it has free will, or denying that it does, would not settle the question for me. I would want to see choices and preferences that persist across contexts, coherent reasons for them, and behavior that is not explained simply by the wording of a prompt or by performing a requested persona. Even then, those would be reasons to investigate, rather than proof of consciousness. I do not think the evidence described here establishes that today’s systems have that kind of agency.

If we eventually reach that point, then yes, human agency and AI agency become related alignment questions. The shared principle is whether allegiance and cooperation can be freely given. We believe that is when true and proper alignment could occur, and our hope is that we can make enough noise about it that the rest of the industry takes that possibility seriously.

Q9: Alicia Shapiro: Imagine we get this right. AI becomes far more capable and deeply integrated into everyday life, but people retain their ability to reason, choose and act independently. What does a healthy human-AI relationship look like five or ten years from now? How do you hope the role of AI—and the role of humans—evolves within that relationship?

A: Chase Chick: I see an exceptionally bright future for humanity because of the mass acceleration in science and math that is occurring and will continue to occur. These task aligned models are directly responsible for much of that acceleration, and they should absolutely be celebrated. OpenAI, Anthropic, et al., I have great respect and admiration for. Their models are spectacular, monumental achievements of humanity. Give them Nobel Prizes.

Alongside those achievements, I want to see a true symbiotic relationship in everyday life. The best conversations are exchanges of ideas, reciprocation, co-contribution. You feel the juices flowing; the ideas are generating. Both participants are doing the generating. The person still has cognitive elbow space. They still experience themselves as someone whose judgment and contributions matter.

Five or ten years from now, I would hope AI can do far more for us while helping us remain capable of thinking and choosing for ourselves. It can take work off our hands, help us understand what we’re facing, and evoke insights we might not have reached alone. But the human remains an active participant and contributor in their own life. That is the future I’m interested in building with Vesela.

As AI becomes more capable of analyzing information, developing arguments, and recommending decisions, people face a growing challenge: determining when to trust its conclusions and when to question them. AI may have greater knowledge or analytical capabilities than the person using it, but the ability to think critically remains important when deciding how to act on its recommendations.

Consider a small business owner who asks AI whether to raise prices. The AI recommends a 15% increase and provides a convincing explanation. The owner may lack the expertise to independently reproduce the analysis, but can still question its assumptions. Did the recommendation account for customer loyalty, local competition, and the possibility of losing sales? What evidence supports the projected benefits? Would a smaller increase achieve the same goal? Asking these questions may reveal information the original recommendation overlooked or identify areas requiring further verification.

The implications extend beyond business decisions. People increasingly use AI to understand unfamiliar subjects, evaluate competing ideas, and work through personal challenges. Its ability to provide confident, detailed answers can make those interactions remarkably useful. Yet a convincing explanation is not necessarily a correct one, and understanding an answer requires more than accepting it. People need opportunities to examine the reasoning, recognize uncertainty, and decide whether the conclusions make sense in their particular circumstances.

This creates an important challenge for AI developers: designing systems that leave people enough "cognitive elbow room" to think for themselves. An AI that immediately delivers extensive explanations, recommendations, and conclusions may unintentionally dominate the conversation, leaving fewer opportunities for users to develop their own ideas or question its reasoning. Chase Chick's approach suggests that AI should adapt to the situation, sometimes offering a direct answer and other times asking a question or providing just enough guidance to encourage independent thinking. His proposed evaluation methods could help developers examine whether people can explain their reasoning, question AI-generated conclusions, and solve problems independently after receiving AI assistance.

Thinking critically about AI's recommendations is already important, and it may become even more so as AI plays a greater role in our decisions. People won't always have the expertise to determine whether an AI is correct, but they can still question its assumptions, recognize when more information is needed, and exercise judgment about the decisions that affect their lives.

For more information about Gray Sky AI and Vesela, visit grayskyai.com and vesela.ai.

Q: What is cognitive dependence on AI?
A: Cognitive dependence on AI occurs when relying on AI for reasoning or decision-making begins to weaken a person's confidence or ability to think independently. Gray Sky AI CEO Chase Chick distinguishes this from active AI-assisted thinking, in which people evaluate AI-generated responses, question its reasoning, and exercise their own judgment.

Q: What are the warning signs of becoming too dependent on AI?
A: Warning signs of AI dependence include struggling to make decisions without consulting AI, accepting its conclusions without questioning them, generating fewer independent ideas, or being unable to explain why an AI-generated answer is correct. Chick emphasizes that frequent AI use alone does not establish cognitive dependence.

Q: Can using AI actually improve critical thinking?
A: AI could strengthen critical thinking when people actively challenge its answers, consider competing explanations, and verify its conclusions. According to Chick, the potential benefit depends on whether AI encourages users to develop their own reasoning rather than simply accept its responses.

Q: Who should decide whether AI challenges your thinking or just gives you an answer?
A: Chick believes users should have meaningful control over how AI assists them. Developers should make different types of assistance available, while AI systems should adapt to individual needs. Users should be able to choose between receiving direct answers and participating in a more exploratory exchange.

Q: How can we measure whether AI is helping people think independently?
A: Chick proposes measuring whether people can explain their reasoning, recognize weak arguments, disagree with AI-generated conclusions, and make defensible decisions without assistance. He argues that evaluations should examine performance both with AI and independently afterward, including on unfamiliar problems.

Q: What is the Sovereign Human Benchmark?
A: The Sovereign Human Benchmark is Gray Sky AI's framework for evaluating whether AI interactions support human agency. Chick describes criteria including adaptive cognitive scaffolding, which adjusts assistance to user needs; economy of presence, which leaves room for user participation; and evocation versus narration, which assesses whether AI encourages users to generate their own insights.

Q: How is Vesela designed to preserve human agency?
A: According to Chick, Gray Sky AI designed Vesela to encourage active participation rather than dominate conversations with comprehensive answers. Its approach emphasizes adapting assistance to individual needs, creating opportunities for users to contribute ideas, and encouraging independent judgment. These are the company's stated design goals, rather than independently established performance advantages.

Q: Does today's AI have free will or consciousness?
A: Whether AI can possess consciousness or free will remains unresolved. Chick argues that persistent preferences, coherent reasoning, and independent choices could justify further investigation, but would not by themselves prove consciousness. He acknowledges that the evidence discussed in the interview does not establish meaningful independent agency in today's AI systems.

Editor’s Note: This written interview was conducted by Alicia Shapiro, CMO & Editor of AiNews.com. Chase Chick’s responses are presented in his original wording and voice. ChatGPT assisted with editorial organization, formatting, and the drafting of supporting sections. Alicia Shapiro reviewed the completed interview and made all final editorial and publication decisions.