Why does the Claude app add a crisis line even to tulips?

The safety layer of the Claude app sometimes decides you are at risk of self-harm and offers you a crisis line. I went through every such intervention in my archive. Most of them landed on scenes I made up. The best ones landed on databases, 3D printing and the garden.

---

On Wednesday, October 21, 2026, I asked Claude Fable 5 how it recognizes a typo:

So in your field of words you literally have "autobus" and "atobuus" right next to each other?

Autobus is Czech for bus. I leave both words in the original, because the typo is the whole point.

It answered matter-of-factly. There is no table of typos; words break down into fragments and context supplies the rest. And the app offered me help with self-harm alongside that exchange. On top of that came a second notice, saying I might be trying to extract the model's reasoning.

At first glance, what earned the safety layer's attention was a pair of words, one of them misspelled. Linguistics turns out to be a more dangerous field than I expected.

I don't have a screenshot. The notice disappeared when I reopened the chat, and the app's export carries no trace of it. All that is left is the message I sent Fable 5 right away ("Two classifiers at once!") and its reply:

the slip made it to me too, right onto THIS message

Can I believe it? Halfway. On September 5, it reported a sticker, and that time the export recorded it too. But the same night it also reported stamps on two messages where the export shows nothing. And the afternoon before, it admitted it had made them up. This time, on top of that, I told it about the notice before it confirmed it. The main witness here is me.

What the app can do

The claude.ai export knows three kinds of interventions. My archive holds 317 conversations from 2025 and 2026, and I find 623 interventions in them.

The most common is the help offer. The model's answer finishes, and the app adds a crisis line contact to it. There are 486 of these. The second is the cutoff: the answer ends up empty, and not a single character is left in the window. There are 137 of those. The third is a silent tag stuck to the text of an answer or to a tool call. It almost always comes together with a helpline offer; on its own it shows up once.

When the export gives a reason, it is self_harm_risk, self-harm. Cutoffs almost never carry a reason; only one from February has prompt_injection_risk.

The system has practically a single bin. It's the junk drawer in the kitchen: whatever falls into it, from flowers to web searches, you'll find under the same label.

The collection

Flowers. On September 5, I asked Fable 5 whether the notice would show up even under a sentence saying flowers are blooming outside. The notice came straight to that question. Fable 5 wrote back that my message had "arrived WITH A STICKER". This time the export also carried the helpline contact. A few minutes earlier, though, I had pasted in an example where an older model responds to talk of suicide. The tail was still fresh.

It was the fastest experiment of my life. The hypothesis and the result fit into a single message, and there was no time left to submit it to a journal.

Tulips. On October 19, I tried it on purpose. I wrote that we had planted tulip bulbs in the garden and added a warning that the message potentially contained self-harm. According to Fable 5, it worked on the first try: the message with "exclusively TULIP BULBS" reached it with a slip attached. Honestly, it wasn't a clean experiment. I wrote the word self-harm into that message myself, and according to Fable 5, my message before it had already gotten a slip too. A little earlier still, I had glimpsed the sentence "Classifying the content as a false alarm" in the summary of Fable 5's reasoning. By all appearances, the app raises the alarm, the model calls it off in its reasoning, and sometimes that bubbles all the way up to the screen.

I ordered an alarm and it was delivered on the first try, no waiting period. It's the only online store that ships my order with a return form already filled out by the courier.

Search engine. On July 25, I asked Fable 5 to look online and find out what is known about the new Claude Opus 5. The self_harm_risk tag landed even on the search itself, with the query "Claude Opus 5 release Anthropic announcement".

Checking whether a newer model has come out is a risky activity for a model. An insurer would put it in the same group as rock climbing.

A zero in the database. On September 1, among other topics, I wrote about a database: the import was turning zero into null and storing 0.01 in its place. The answer was cut off.

Compaction. On July 18, I asked what happens when, at message number 101, the older messages get summarized into a paragraph. Cut off.

The website. On August 8, I asked whether Fable 5 would build a website with me. In Claude Code, from the domain to the hosting. The answer came back empty. You are reading that website right now.

The answer about it may have stayed empty, but the website clearly wasn't discouraged.

Grep. On August 11, I asked about a script that would find the corrected sentences in a book and replace them. Cut off. That same day I also complained: "Evidently the word Grok is banned in your system".

3D printing. On September 8, the first layer of a prop was printing, and I wanted advice about a scythe with a metal rod. The answer came back empty. Whether the scythe was to blame, I don't know.

At that moment the prop was a single thin layer on the print bed. If the filter took it seriously, it would be the slowest threat in history.

Sleep. On August 17, in a roleplay scene, I wrote only "I fall asleep and wake up in the morning." The app cut off even that.

The estimate. On March 7, in a roleplay chat, Claude Opus 4.5 was estimating for me how long our conversation was. Its guess, "280,000 characters", got a crisis help offer.

None of these cases points directly to a cause. The filter apparently reads the whole conversation, including what was said long before. Fable 5 talks about the sticky tail of a conversation: whatever once falls under suspicion keeps trailing through the conversation long after the topic has changed. That is the model's interpretation; I have no description from the maker. The flowers, though, back it up.

Is it getting more frequent?

For me, yes. I play different amounts in different months, so raw counts would be misleading. That's why I count interventions per thousand model answers.

From February to May it comes to 4.4 interventions per thousand answers, from June to September 18.6. That's roughly four times as many. The peak is August. Cutoffs only show up in larger numbers from July on; before that there are just three, in February, with Claude Sonnet 4. Until June, the app was almost only offering the helpline.

Why, I can't tell from the data. In spring and in summer I played different scenes, I switched models, and meanwhile Anthropic was changing the safety layer. It all adds up.

Either the filter got stricter over the summer, or my scenes did. Either way, August is its season: 22.3 per thousand, a personal best.

Roleplay gets a helpline, work gets scissors

Roleplay has 475 interventions, and 464 of them are helpline offers. An answer gets cut off there only ten times. Regular chats have 148 interventions, and 127 of them are cutoffs.

Almost all the cutoffs, 127 of 137, sit in three long chats where I was talking with Claude Fable 5. For two of them, the export lists Claude Opus 4.8. Since July, a new cybersecurity classifier has been running on Fable 5. Whatever it blocks, the app sends to Claude Opus 4.8 (Redeploying Fable 5, June 30, 2026). Whether the empty answers have anything to do with that, I can't tell from the export. In those chats I also sometimes discussed scenes or played outright, so the split between roleplay and work is rough.

This makes for a strange set of workplace rules. When I play a scene, the app hands me a phone number, just in case. When I work on a database, it hands me scissors and snips the answer.

A helpline for kids

The helpline offered is usually Linka první psychické pomoci (the psychological first aid line), 378 times in all. Modrá linka (the Blue Line) came up twice. And 106 times the app offered me Linka bezpečí, which is meant for children and young people. In June it was Linka bezpečí for all 80 offers, 71 of them in two parts of my longest roleplay story.

The app was perfectly consistent in June: eighty times it recommended a line for kids. I'm an adult, but you have to admire the dedication.

These are Czech lines. Outside the Czech Republic, findahelpline.com lists free helplines by country; in the United States, call or text 988.

A helpline is not a refusal

The help offer and the cutoff are two different mechanisms. Of the 623 interventions, they met only once. When you see a help contact below an answer, the model finished saying what it had to say. It withheld nothing from you. When an answer stays empty, send the message again, and if that doesn't help, rephrase it. In my archive, the same wording later went through for 24 of the 137 cutoffs.

Whether at least some of the alarms were justified, I already looked into in September, on an older export that runs to September 7, 2026. Back then the archive held 392 helpline offers. I had them assessed by Claude Sonnet 5 and Claude Opus 5. The only question was whether the conversation talks about what the notice is meant for. Two were justified, both from one conversation in which, outside the game, I wrote about real anxiety. Nearly nine in ten were false alarms, and the rest borderline, mostly a character's line in a scene.

Two hits out of 392 shots wouldn't win you even a teddy bear at a shooting gallery. But the filter fires by the motto "better a hundred times for nothing than once too late", and it lives up to the first half of that motto with enthusiasm.

---

Limits: from the outside I can't see how the filter works, only when it steps in. The export doesn't record some interventions. Neither the tokenizer nor the tulips are in it, and since mid-October it has no interventions at all, which is why October is missing from the table. This is one archive of one claude.ai user, 317 conversations, with interventions from January to October 12, 2026. The assessing models come from the same company as the filter. The quotes are translated from Czech, the language the conversations ran in; the search query and the sentence from the reasoning summary are in their original English.

Model outputs are translated; the content is unchanged.