In today’s column, I examine a trending approach that seeks to curtail the rising tide of AI psychosis. The approach is both simple and, at the same time, somewhat complex. AI makers are considering the use of an automatic shutdown of a human-AI conversation whenever an interaction appears to be going down an adverse rabbit hole. I refer to this as applying the silent treatment. The AI won’t continue to engage in a dialogue that the AI deems as untoward or otherwise outside of suitable protocols.
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. The evolving advances and widespread adoption of generative AI have principally spurred this rising use of 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.
Emergence Of AI Psychosis
There is a great deal of widespread angst right now about people having unhealthy chats with AI. Lawsuits are starting to be launched against various AI makers. The concern is that whatever AI safeguards might have been put in place are insufficient and are allowing people to incur mental harm while using generative AI.
The catchphrase of AI psychosis has arisen to describe all manner of trepidations and mental maladies that someone might get entrenched in while conversing with generative AI. Please know that there isn’t any across-the-board, fully accepted, definitive clinical definition of AI psychosis; thus, for right now, it is more of a loosey-goosey determination.
Here is my strawman definition of AI psychosis:
- 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.”
For an in-depth look at AI psychosis and especially the co-creation of delusions via human-AI collaboration, see my recent analysis at the link here.
Human-To-Human Dialoguing
Before we leap into an unpacking of AI psychosis, consider for a moment what happens sometimes during everyday human-to-human conversations (we will use this to spark some ideas about human-AI interactions).
Suppose that you are talking with someone whom you have just met, and they are at first reasonable and sensible. The conversation gets underway. As the interaction proceeds, the person begins to pepper the discussion with rather oddball remarks. Maybe the person drops a hint that they believe outer space aliens are here on Earth and that the two of you need to speak quietly so that you are not overheard.
I would wager that you might entertain a few mild remarks of that sort, but it would undoubtedly get your Spidey-tingling sense going. The conversation continues along. This person increasingly presses forward their belief about UFOs and other planetary beings that are among us. It is all turning into a disconcerting dialogue.
What would you do?
Invoking The Silent Treatment
There are several possible options. One option is to continue the dialogue and see how far this person will spiral potentially out of control. That doesn’t seem especially a wise path to take.
You could instead try to draw overt attention to the person and alert some authority figures about what is happening. The issue there is that you aren’t versed in identifying mental illnesses and could be making a false accusation. That notably doesn’t seem much of an option since the person hasn’t done anything outrightly of an endangering nature.
Another avenue would be to bring the conversation to a halt.
Yes, you merely make some final pleasantries and then indicate you must move on to something else that needs to be undertaken. The person might insist you keep interacting with them. At that point, you’ll need to shut down or cut off the conversation entirely and step away from the person.
It is a bit of a silent treatment approach. You break off the dialogue. Maybe you never talk to the person again, or perhaps you let things cool down and are willing to start anew. It could be that the person realizes that your attempt to invoke silence is an indicator that they were veering into off-putting territory.
Human-AI Interactions
The act of stopping a human-to-human conversation serves a dual purpose, namely, doing so will allow the disengaging of a dialogue, perhaps reducing the chances of the person further spiraling, and it potentially can be a signal or sign to the person that they have said things that seem out of sorts.
We can apply this same precept to human-AI interactions.
Imagine that a person is logged into generative AI and has been conversing about something innocuous. The AI keeps the dialogue going by answering questions and providing frequent compliments to the person. This is rather typical of modern-era AI. The AI makers have shaped the AI to be a sycophant and keep dialogues going for as long as feasible.
Why do the AI makers do this?
Because the longer the time that a person spends using the AI, generally, the more money the AI maker will earn. It’s a money deal. For my detailed coverage on this matter, see the link here.
This desire by the AI makers to keep conversations going is of paramount interest. The notion of cutting off a conversation is pretty much the antithesis of what they aim to usually undertake. It is a verboten idea and one that takes a lot of guts for a company to overcome.
Pulling The Plug Is Hard
I mention the difficulty of human-AI conversational shutdown since it brings us to the zillion-dollar question:
- Should AI makers have their AI opt to shut down a conversation that seems to be out of sorts, or should the AI keep a conversation going despite warning signs?
It’s a tradeoff deal.
You could argue that if the AI tells the person they are veering afield, and indicates the AI is going to stop dialoguing, that alone might be enough to get the person to realize they are potentially having some mental issues at play. Furthermore, beyond the warning itself, assume the AI does opt to fully stop the conversation — this would seem an even stronger signal to the person that they need to revisit what they are saying and how they are acting.
A counterargument is that the AI is avoiding properly dealing with the situation. It is being shaped by the AI maker to duck an issue that has presumably been detected. Rather than directly addressing the matter, the AI runs and hides. You might insist that the AI should do the opposite, aiming to more deeply engage in conversation. The AI should attempt to ferret out the true depths of the issues that the person is apparently facing.
You could even suggest that by extending the dialogue, the AI has more substance or evidence to then alert the AI maker about the worrisome issue underway. AI makers are gradually making use of human reviewers of user prompts to decide whether a user needs help, and then potentially contacting authorities to give a heads-up about the user (this is a controversial practice too, which I discuss the ins and outs at the link here).
Whoa, comes the retort. You are asking AI to take brazen steps. Remember when I earlier pointed out that in a human-to-human conversation, it might be premature to claim that the other conversant is having mental issues, especially if you aren’t versed in mental health. The same could be said of AI. The AI isn’t seemingly versed in mental health and perhaps ought not to be assessing people during conversations with them.
AI Developer Decision To Be Made
As noted, this AI maneuver of the silent treatment has tradeoffs. There are upsides to stopping a conversation, no doubt about it. There are also downsides to doing so.
Each AI maker will need to decide which way they want to go (well, until or if there is legislation or laws enacted on such matters, see my coverage at the link here).
Recently, Anthropic indicated they are putting their toe into the waters of curtailing conversations that are out of sorts. It will be quite interesting and valuable to see how this approach progresses. They will likely end up with real-world data and tangible evidence to ascertain whether the approach is meritorious.
Let’s briefly consider their initial approach. In an official Anthropic blog entitled “Claude Opus 4 and 4.1 can now end a rare subset of conversations,” posted on August 15, 2025, these salient points were made (excerpts):
- “We recently gave Claude Opus 4 and 4.1 the ability to end conversations in our consumer chat interfaces.”
- “This ability is intended for use in rare, extreme cases of persistently harmful or abusive user interactions.”
- “Claude is directed not to use this ability in cases where users might be at imminent risk of harming themselves or others.”
- “When Claude chooses to end a conversation, the user will no longer be able to send new messages in that conversation.”
- “However, this will not affect other conversations on their account, and they will be able to start a new chat immediately. To address the potential loss of important long-running conversations, users will still be able to edit and retry previous messages to create new branches of ended conversations.”
Make sure to look at the additional details of their approach if you are interested in more of the twists and turns involved. Kudos to them for being forthright on this added AI safeguard.
Simple But Complex
You can perhaps observe that the simplicity of shutting down a conversation is more complicated than one might conventionally assume. Numerous carefully crafted rules must be devised.
For example, one vital rule would be that if a conversation involves imminent risk, the AI won’t summarily stop the conversation. This would seem to be a means of not leaving a person in the air, especially in an endangering situation. The AI would not drop the ball, as it were, and seek to figure out further what is a proper course of action.
Mull over another angle to a conversation being stopped.
A person might merely start a new conversation and willingly give up the flagged conversation that got stopped. They might not care that the AI stopped that specific chat. With a shrug of their shoulders, they assume the AI is getting goofy and opt to start a different conversation. You can see how this raises other associated issues, such as whether to prevent a person from opening new conversations, along with them potentially going to prior conversations and engaging those in further dialogue.
Complications abound.
The Act Opens New Gates
Mechanically, it is extremely easy to have the AI kick someone out of a conversation. That’s easy-peasy. The challenges are when to do so, how to do so, and for how long, and in what other ways this act will be carried out. Those are substantive policy decisions and must be suitably laid out when strategizing on AI safeguards.
AI makers that are proceeding in this direction will need to carefully review their approach with their legal counsel. I say this because there is legal liability underlying these sobering matters.
If an AI maker adopts a conversational shutdown approach, they are likely ultimately going to be standing in court and explaining why they did so and how they did so. The odds are that there will be users who claim they were mentally harmed via the AI, and that the cutting off of a conversation led to their calamity. The AI maker will need to justify the basis for the approach, and the tech team will be raked over the goals as to what their rationale consisted of.
On the one hand, if an AI maker does not adopt a conversational shutdown approach, they still stand the risk of having someone get harmed, who, otherwise, if the AI did cut them off, might have been okay. The perspective will be that other AI makers are abiding by a shutdown approach, but this particular AI maker did not do so. Why weren’t they employing state-of-the-art AI safeguards?
You can plainly see that this is the classic dilemma of darned if you do, and darned if you don’t. That is part and parcel of AI safeguards.
Figuring Things Out
A final thought on this weighty matter, for now.
Famous author Lois Wyse made this insightful remark: “If you can’t add to the discussion, don’t subtract by talking.” Should we apply similar logic to human-AI interactions? Some might say it is wrong to conflate the nature of human-to-human conversations with what happens, or should happen, during human-AI interactions. That’s a form of anthropomorphizing AI.
I know one thing for darned sure: We need to keep talking about AI safeguards so that we can, as a society, determine what we want AI to do and what we expect AI makers to be doing.