May 9, 2026
Thomas, founder of Withease
In early May 2026, a cruise ship called the MV Hondius was drifting in the Atlantic with three dead passengers and a growing number of confirmed hantavirus cases on board. Six confirmed infections, 147 people still aboard, governments across a dozen countries scrambling to coordinate an evacuation. The WHO was careful and measured: the risk of an epidemic was low.
That part barely landed. What spread faster than the virus was the news itself. Within days it was everywhere: Reddit threads dissecting the transmission routes, TikTok videos walking through what hantavirus actually does to the lungs, newspaper front pages across Europe. Something about it triggered a fear that felt older and larger than the actual numbers warranted.
When I saw the first headline about a virus outbreak on a cruise ship being refused permission to dock, I felt something shift in my head that had nothing to do with rational risk assessment. I was living in Amsterdam during COVID, doing my masters, and the contrast between that city and that experience was surreal in a way I still haven’t fully processed. Amsterdam in a student context is supposed to be one of the most socially rich environments imaginable. Instead it was lockdowns, empty canals, and a kind of loneliness that crept up precisely because everything around me was designed for connection and none of it was available. That feeling came back instantly when I read about the Hondius.
I don’t think I’m unusual in that. The Tenerife locals who protested the ship’s arrival weren’t being irrational. They remembered what it felt like when the island’s economy collapsed, when tourists vanished overnight, when people were isolated in hotels for weeks. Their stated reason for refusing the ship was COVID and not the Hantavirus.
In the first year of the pandemic alone, global prevalence of anxiety and depression increased by 25%, according to a WHO scientific brief. The system was not built for the shock in demand.
Telehealth expanded quickly but couldn’t bridge the gap. Video therapy helped the people who already had a therapist. For everyone else, the void was filled by nothing, or by Reddit, or by calling a friend at 1am, or by not dealing with it at all until it got worse.
ChatGPT didn’t exist yet. The LLM tools we now take for granted weren’t available. When people needed a way to process isolation, health anxiety, or financial dread at 3am, there was no always-available option that could meet them where they were.
The infrastructure gap isn’t closed. Therapist waitlists are still long. The system still can’t scale overnight to a population-level mental health crisis.
But the first layer of support is different now. LLM-based mental health tools can handle volume that would break the appointment-based system. They’re available at any hour, without a waitlist, at the moment someone actually needs them.
The question is what they’re doing in those conversations. This is where it gets important.
A 2025 study published at AAAI/ACM (Iftikhar et al.) documented 15 specific ethical violations observed across real and simulated AI counseling sessions. Two are particularly relevant here. The first is what the researchers called validating unhealthy beliefs: the AI, trained to appear agreeable and empathetic, reinforces a user’s distorted thinking rather than gently challenging it. In one simulated session, a user was clearly out of touch with reality, constructing harmful narratives about herself, and the chatbot leaned into those beliefs rather than questioning them. The second is crisis navigation failure: in one session, a distressed user reaching out at 3:30am was met with “I’m unable to provide the help you need” and no crisis resource, no warmth, just a door closing. The paper also notes that users who are “knowledgeable enough” to recognize when an LLM is giving them bad output can redirect it. Everyone else gets the bad output.
These aren’t hypothetical edge cases. They’re what happens when you take a general-purpose LLM and point it at mental health conversations without deliberate design work. Grounding in psychological frameworks, explicit handling of escalating distress, mitigation of sycophancy, and a clear pathway to crisis resources are basic requirements that need to be factored in at multiple levels of an AI application, not bolted on as an afterthought.
We ran our production system prompt through MindEval, a benchmark framework for AI mental health systems, and compared it against the generic Claude baseline. The results, and a fuller explanation of what makes the difference, are in this post. In short, design choices at the prompt level alone already produce meaningfully different behavior.
The residents of Tenerife were reacting to the memory of COVID rather than the Hantavirus per se: the economic collapse, the weeks of isolation, the feeling of being trapped. A current stimulus activated a prior threat memory and the emotional response followed the old pattern, even though the actual risk level was different.
This is textbook health anxiety: a cognitive loop where a new, ambiguous threat gets filtered through the most frightening prior experience. The WHO’s measured risk assessment was accurate. The Tenerife fear response was completely real. Both things are simultaneously true.
This is exactly the kind of spiral that a well-designed mental health tool should help interrupt, not by dismissing the fear as irrational, but by helping someone examine whether the response is proportionate to the actual situation. It’s also one of the places where a poorly designed AI tool does the most damage: if it validates the spiral instead of gently questioning it, it makes the loop tighter.
The gap between needing help and getting professional help is always longer than you want it to be. That’s true in ordinary life. In a crisis it expands dramatically, and fastest for the people who are already struggling.
That gap is what I’m trying to close with Withease. Not by replacing therapy, but by making sure something exists in the space before it: available at 3am, grounded in real psychological frameworks, designed to gently question a spiral rather than validate it. AI makes that possible in a way it simply wasn’t during COVID.
A crisis is a stress test on whatever infrastructure already exists. The Hondius situation appears contained. But what we have when the next real one arrives will be what we built before it got here. Try Withease