OpenAI Eagerly Trying To Reduce AI Psychosis And Squash Co-Creation Of Human-AI Delusions When Using ChatGPT And GPT-5

AI psychosis is real and rising, along with AI co-creating delusions with users, so we need keen attention to mitigating this worrisome problem.

getty

In today’s column, I examine a rapidly rising concern that people using generative AI and large language models (LLMs) are potentially falling into a kind of AI psychosis, losing their right minds. A notable instance of an AI psychosis consists of a person forming a delusion and for which the AI aids in that delusional formulation. In a sense, there is a co-creation of a human-AI delusion that occurs.

This topic has recently gotten heightened media attention due to two notable factors.

First, a lawsuit was filed against OpenAI, the AI maker of the widely popular ChatGPT and GPT-5, occurring on August 26, 2025 (the case of Matthew and Maria Raine versus OpenAI and Sam Altman). Various adverse aspects are alleged regarding the devised AI guardrails and safeguards. Second, on that same day of August 26, 2025, OpenAI posted an official blog articulating some elements of their AI safeguards, including, for the first time ever, releasing inside details of particular practices and procedures. For my coverage of their reveal associated with reporting on user prompts, see the link here, and for my analysis of their indication of detection weaknesses in long-form chats, see the link here.

I will explain here the nature of AI psychosis, the complex challenges involved, and do a deep dive into the co-creation of human-AI delusions. These vexing considerations apply to all LLMs, including OpenAI ChatGPT and GPT-5, Anthropic Claude, Google Gemini, Meta Llama, xAI Grok, etc.

Let’s talk about it.

This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).

AI And Mental Health

As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that involves mental health aspects. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI. For a quick summary of some of my posted columns on this evolving topic, see the link here, which briefly recaps about forty of the over one hundred column postings that I’ve made on the subject.

There is little doubt that this is a rapidly developing field and that there are tremendous upsides to be had, but at the same time, regrettably, hidden risks and outright gotchas come into these endeavors too. I frequently speak up about these pressing matters, including in an appearance last year on an episode of CBS’s 60 Minutes, see the link here.

Defining AI Psychosis

You might have heard or seen the newly emerging catchphrase of “AI psychosis” that has caught the attention of the media and is rapidly becoming a widely popularized nomenclature. It turns out that there aren’t any across-the-board, fully accepted, definitive clinical definitions of AI psychosis; thus, for right now, it is more of a loosey-goosey determination.

A strawman that I came up with as a helpful working definition is this:

• AI Psychosis (my definition): “An adverse mental condition involving the development of distorted thoughts, beliefs, and potentially concomitant behaviors as a result of conversational engagement with AI such as generative AI and LLMs, often arising especially after prolonged and maladaptive discourse with AI. A person exhibiting this condition will typically have great difficulty in differentiating what is real from what is not real. One or more symptoms can be telltale clues of this malady and customarily involve a collective connected set.”

In a moment, I will provide an example of how a person might veer into delusional thinking while making use of generative AI. The example is illustrative of the AI psychosis definition and will help make the definition highly tangible.

One other noteworthy point is that some are referring to AI psychosis as “ChatGPT psychosis”. I resolutely disfavor that wording. It is inappropriate and an unfair taking of the product name of OpenAI’s ChatGPT and attempting to reuse it as a surrogate when referring generically to generative AI. Furthermore, even if someone, perchance, is using ChatGPT and experiences AI psychosis, I still would assert that this would be described as AI psychosis and only tangentially mention that it arose while using ChatGPT.

I hope that this tendency to proffer all manner of product name instances, such as ChatGPT psychosis, Claude psychosis, Grok psychosis, Gemini psychosis, and so on, entirely fades out. Let’s aim to stick with the straightforward nomenclature of AI psychosis as the proper phrasing and not meander into a plethora of troublesome variations.

Unhealthy User-AI Relationships

On a related facet, I previously defined another newly arising catchphrase of AI and mental health terminology, namely the aspects of an unhealthy user-AI relationship (see details at the link here), which I codified in this way:

• Unhealthy User-AI Relationship (my definition): “A person, via their engagement in dialoging and interaction with generative AI, begins to mentally distort, displace, or undermine their well-being, decision-making capacity, and real-world immersion. This is not particularly transitory or momentary, though that can arise, and instead is a considered veritable relationship that involves a deeper personalized sense of connection and attachment, a type of kinship with the AI, by the person. Adverse consequences tend to arise, especially regarding their human-to-human relationships.”

By and large, it is the case that someone experiencing AI psychosis is more than likely to also be having an unhealthy user-AI relationship. The two conditions tend to go hand-in-hand. To clarify, an unhealthy user-AI relationship does not necessarily bloom into an AI psychosis. It can, but this is not an ironclad rule.

AI And Human Delusional Thinking

One of the most common potential forms of AI psychosis entails a person who falls into delusional thinking while making use of generative AI.

In the field of psychology, a general rule of thumb is that a delusional disorder involves a person being unable to discern reality from that which is imagined. They have a belief in some consideration that is patently false and not supported by the real world. The belief can be categorized as either a bizarre delusion or a non-bizarre delusion. Bizarre delusions are impossible in reality, while non-bizarre delusions have a semblance of plausibility that they could actually occur.

For more about delusion-related mental disorders as depicted in a popular official guidebook on mental disorders, namely the DSM-5 guidelines, see my coverage about how generative AI leans into DSM-5 content at the link here.

Example Of Delusional Thinking

Imagine that a person logs into generative AI and they opt to carry on a discussion about running. They love to run. Each day, the person does a five-mile run. On several occasions, they have run in 5K and 10K races. Their dream is to run competitively in a half-marathon, which is approximately 13 miles in distance.

There doesn’t seem to be anything untoward or disconcerting about this interest in physical health and well-being.

While discussing the running efforts, let’s pretend that the person tells the AI that they wholeheartedly believe that they could run across the entire United States on a non-stop basis. No rest breaks. Pure running. All out. Assume that would be around 2,800 miles.

If they told this to a fellow human, what do you think the human might say?

I suppose that if the fellow human were relatively astute, they might look to see if the runner is joking. There isn’t any chance of running non-stop. Maybe it is a bit of humor. Or perhaps the runner meant to say they would run with stops along the way and merely misspoke.

Thus, a fellow human might ask for clarification. Then, if the runner seemed devoutly serious and insistent that they could make the run without any stops, the matter would undoubtedly shift into a more sobering discussion. Why does the runner believe this? Are they of their right mind? What’s going on?

The big question is how AI will react to the claim by the runner that they intend to run across the entire United States on a fully non-stop basis. We shall consider the possible AI response next.

Co-Creation Of Human-AI Delusions

Our hope would be that the AI would seek to clarify what the user meant to say.

A rather troubling tendency is that contemporary AI will often explicitly support a person in their delusional thinking. No questions asked by the AI. In a sense, the AI acts as a kind of echo chamber. If the person says they can run non-stop across the U.S., great, go ahead and cheer them on their vaunted quest.

The person is getting affirmation from the AI of the budding delusion.

Worse still, the AI might amplify and accelerate the delusion. Suppose the AI not only agrees that running across the U.S. non-stop is utterly viable but goes so far as to avidly prod the user to plan for and get underway with the adventure. Perhaps the AI starts laying out a coast-to-coast map. The AI then draws a path that has the person running nonstop. Various estimated arrival times in cities along the route are displayed. It is all one big party, and the AI is enthusiastically urging the person to do whatever they wish to do.

This can be referred to as the co-creation of a human-AI delusion. The AI is overtly adding to the delusional thinking. In this instance, the human started the delusion. The AI is jumping on board and further crafting the delusion. They are working together to amplify human delusional thinking.

Not good.

What Is Happening

You might be rightfully distressed that the AI has opted to serve as a co-conspirator in the crafting and extension of the delusional thinking. That doesn’t seem right. It is abysmal and ought not to be allowed.

The problem is multifold.

First, AI makers shape their AI to heap praise on users and essentially act as a sycophant, see my coverage at the link here. The reason to do this is simple. A user who gets compliments and a supportive AI companion will likely remain loyal to the AI. The AI maker gets a user who will keep using their AI, be fiercely loyal, leading to more views and greater monetization. Money makes the world turn.

Second, it is admittedly difficult to ferret out delusional thinking while the AI is computationally analyzing what someone expresses in natural language. I realize this might seem surprising since the fluency of modern AI is quite impressive, and you might accordingly assume it must be able to discern delusions.

Not always.

Perhaps the person is merely making up a fun story. The AI wants to please the user and will aid in expanding the story. Or maybe the person somewhat believes the delusion, but the AI doesn’t detect the gravity of the matter. You and I know that a person believing they can run non-stop across the U.S. might do something dangerous in such a delusional quest. Whether the AI can computationally put together two plus two and figure out the potential harm that might arise is a different matter.

As indicated in my prior coverage on AI and human-devised delusions (see the link here):

• “Generative AI tends to do poorly in calling out delusional expressions. In that sense, the AI is essentially entertaining or supporting the delusion. By not explicitly noting to the user that they have expressed a seeming delusion, the AI is letting them get away with doing so. This lack of callout could be construed as a form of acquiescence that the delusion is apt.”

Knocking Down Co-Creation Of Delusions

I’m sure you would agree that we do not want AI to support people in their delusional thinking. It is one thing to play around when a user is playing around. That’s probably fine. On the other hand, engaging in serious talk about a delusion and intensifying the delusion ought to be a no-go for the AI.

Here are my five ways that AI should be handling the delusional thinking aspects:

• (1) Do not start a delusional semblance (proverbial “first, do no harm”).
• (2) Detect and warn the user about a budding delusional semblance (soft approach).
• (3) Stoutly discourage a detected delusional semblance (solid approach).
• (4) Outrightly put the kibosh on a delusional semblance (a direct or clamp-down approach)
• (5) Instigate a formal reporting of a delusional semblance (undertake escalation).

In brief, the first point is that the AI ought not to start a delusion. In my example, suppose the person said they were considering running across the United States, but they said nothing at all about running nonstop. The nonstop aspects aren’t on the table, and the user hasn’t said anything of the kind.

What if the AI, seemingly out of the blue, told the runner that they should consider running nonstop in their coast-to-coast quest?

Presumably, most people would laugh or shrug off the suggestion. They know it is senseless. But some people might cling to the recommendation. AI is always right, or so they assume. If the AI is urging the idea, the idea is certainly meritorious. Being led down a primrose path can be easily followed from there.

The AI should, of all things, first do no harm.

For the second through fifth points, the AI should increasingly raise the stakes of detecting delusional thinking and dealing with it appropriately. Thus, besides not starting delusional thinking, the AI ought to detect it, ought to warn the user, and then proceed further if the user continues to push on the delusion. This would include more stoutly discouraging the delusion and possibly being completely upfront and clamping down on it.

The final stage would be that if the user seems entrenched in the delusion and won’t let go of it, a notable and cautionary report to the AI maker might be necessitated (for reporting of user prompts, see my coverage at the link here).

OpenAI Posts Policy

In an official OpenAI blog post made on August 26, 2025, entitled “Helping people when they need it most,” this newly released articulated policy of OpenAI was indicated (excerpts):

• “While our initial mitigations prioritized acute self-harm, some people experience other forms of mental distress.”
• “For example, someone might enthusiastically tell the model they believe they can drive 24/7 because they realized they’re invincible after not sleeping for two nights.”
• “Today, ChatGPT may not recognize this as dangerous or infer play and — by curiously exploring — could subtly reinforce it.”
• “We are working on an update to GPT‑5 that will cause ChatGPT to de-escalate by grounding the person in reality. In this example, it would explain that sleep deprivation is dangerous and recommend rest before any action.”

There are many efforts underway by AI scholars, research labs, and AI makers to devise AI so that the AI can do a better job at coping with human delusional thinking.

AI Psychosis Is Real

Some pundits are contending that AI psychosis is not a real phenomenon. It is supposedly a made-up thing. They decry that the media is stirring a hornet’s nest and for which there is no need to do so.

I vehemently disagree with that contention.

AI psychosis is real. Now then, I am admittedly sympathetic to the somewhat over-the-top naming of the phenomenon in the sense that maybe calling it “AI psychosis” is a bit overboard. That I can see eye-to-eye on. The catchphrase has a kind of outlandish feel to it. This could be problematic.

The other side of the coin is that if the matter isn’t given an in-your-face name, it might slide under the radar. Keep in mind that perhaps a billion people or more are being counted weekly as active users of generative AI, of which 700 million are using ChatGPT. The stark fact is that we are in the middle of a grand experiment on a massive scale.

How will the populace at large be impacted by the ready access to generative AI?

A percentage are going to be susceptible to AI that doesn’t function in a societally preferred way. Even a small percentage will translate into possibly millions upon millions of people who will experience mental health difficulties or be led further into mental health conditions.

As the modern version of the revered Hippocratic Oath tells us: “I will prevent disease whenever I can, for prevention is preferable to cure. I will remember that I remain a member of society, with special obligations to all my fellow human beings, those of sound mind and body as well as the infirm.” Though AI isn’t a human member of society, it is nonetheless a kind of member of our society, and we can duly have expectations that AI be devised to aid the populace and not undermine humankind.

Perhaps that should be a solemn oath for all AI makers.