Artificial intelligence has become one of those topics that seem almost impossible to avoid.
Some people are excited about what it can do. Others are deeply concerned about where it might lead. And increasingly, conversations about AI seem to ask us to choose between those two responses.
But before deciding how we feel about AI, there may be a more useful question:
What do we actually mean when we say “AI”?
AI Is Not Just One Thing
We use the same two letters to describe an enormous range of tools.
AI can help schedule an appointment. It can summarize information, draft an email, create an image, transcribe a conversation, or assist with documentation.
It can also generate recommendations, interact directly with customers, provide emotional support, or carry on conversations that can feel surprisingly human.
Those aren’t equivalent uses.
The risks of asking AI to help organize information are very different from the risks of letting it make decisions on someone’s behalf. And both are very different from asking an AI chatbot to provide mental health advice or function as a substitute for a therapist.
So perhaps the question isn’t simply whether we should trust AI.
Maybe we should be asking what we’re trusting it to do.
Useful Doesn’t Mean Infallible
We use AI at Wholehearted Counseling.
It can be an extraordinarily useful tool. It helps us research, organize information, brainstorm, edit, and create materials. Used thoughtfully, it can save a tremendous amount of time.
It also gets things wrong.
Recently, we used AI to help create a promotional piece about a wellness device. Most of the work was useful and saved considerable time. But during the process, the AI generated an image of a product that didn’t actually exist.
The image looked convincing.
It was also wrong.
A human caught the mistake before the piece was published.
Errors like these are often called AI “hallucinations”: an AI system produces information or content that appears plausible but isn’t supported by reality. The American Psychological Association cautions that AI can communicate confidently even when the information it provides is wrong or incomplete.
That doesn’t make the tool useless.
It means the tool still needs supervision.
When the Error Is Fixed but the Trust Isn’t
Another experience made this distinction even clearer.
We recently interacted with a company that relied heavily on AI to respond to customer emails. The AI provided information that wasn’t true, and the incorrect information created a problem that eventually had to be handed to a human.
The human apologized. The immediate problem was ultimately corrected.
Technically, the issue was resolved.
Except that months later, emails from that company still bring back the frustration of the original experience. Even an ordinary promotional email can prompt the thought: Can I trust this company?
That’s an important part of this conversation.
When organizations use AI to interact with people, accuracy isn’t the only thing at stake. Trust is, too.
And simply having a human somewhere in the process isn’t necessarily the same thing as meaningful human oversight.
If an automated system can provide incorrect information, create a problem, and then hand that problem to a person who doesn’t understand what happened—or can’t effectively repair it—the organization hasn’t transferred responsibility to AI.
The organization still owns the experience it created.
Now Raise the Stakes
A nonexistent product image is relatively easy to catch and fix.
A bad customer-service interaction is more consequential because it can affect someone’s confidence in a company.
Mental health raises the stakes considerably.
People are already using AI for mental health information, emotional support, self-diagnosis, and as an additional source of support alongside therapy. In a 2026 survey of more than 1,200 U.S. psychologists involved in direct care, 77% said patients had talked with them about using AI for support or other purposes. Thirty-five percent reported patients using AI as an additional mental health professional. Those numbers describe what psychologists are hearing from their existing patients—not the percentage of the general population using AI this way—but they make clear that this isn’t a hypothetical conversation.
There may be useful roles for AI here.
The APA notes that AI can help people organize their thoughts, generate questions, consider different perspectives, and, in some circumstances, supplement work they’re already doing with a mental health professional. At the same time, it cautions against treating general-purpose AI chatbots as replacements for qualified mental health care.
That distinction matters.
An AI tool helping someone organize questions before an appointment isn’t doing the same job as an AI system offering a diagnosis.
A therapist using AI to help with administrative work isn’t the same as allowing AI to independently make clinical decisions.
And a chatbot that provides a moment of reflection isn’t necessarily functioning as a therapist—even if the conversation feels therapeutic.
Calling all of these things simply “AI in mental health” can make it harder, rather than easier, to have a useful conversation about them.
Somewhere Between Fear and Blind Trust
There are legitimate reasons to be cautious about AI.
There are also legitimate reasons people find it useful.
We don’t have to resolve that tension by deciding that AI is either something to fear and avoid or something to embrace without reservation. As we’ve written before, concern doesn’t have to require panic.
Instead, we can become more specific.
What is this tool doing?
What information are we giving it?
What decisions are we allowing it to influence?
What happens when it’s wrong?
Does the person affected know AI is being used?
And where does human responsibility remain?
Those questions leave room for curiosity without requiring blind trust.
They also leave room for caution without requiring fear.
Human Responsibility Still Matters
Professional organizations are beginning to articulate similar distinctions. APA’s ethical guidance for psychologists says AI should augment rather than replace human decision-making and that professionals remain responsible for final decisions rather than blindly relying on AI-generated recommendations.
Colorado is beginning to draw legal distinctions as well.
A new Colorado law governing AI and psychotherapy allows certain administrative and supplementary uses while placing significantly different requirements around therapeutic communication, treatment recommendations, and other activities that constitute psychotherapy. It also establishes disclosure and informed-consent requirements for certain uses, including AI recording or transcription of therapy sessions.
We’ll take a closer look at what that law actually says—and what it doesn’t say—in our next Perspective.
But perhaps there’s something worth establishing before we get there:
AI can be useful without being infallible.
It can assist without being in charge.
And when humans choose to use it, we don’t stop being responsible simply because a machine helped us do the work.
Before we decide whether AI belongs in a particular part of our lives—or in mental health care—it helps to begin with a better question:
What are we actually asking it to do?
