ChatGPT Parental Alerts: Why They Didn't Fire in the Test

Many parents picture ChatGPT’s parent alert as a smoke detector: smoke appears, the alarm sounds. It works more like a bank’s fraud review. Software scores activity, a threshold decides what gets flagged, a person reviews the flag, and only then does anyone get a message. That pipeline explains why ChatGPT parental alerts never arrived in a recent independent test, even when testers typed explicit crisis messages for up to an hour.

This article explains the machinery: what a safety classifier is, why thresholds are set the way they are, why OpenAI itself says its safeguards weaken in long conversations, and what parents should and should not rely on. It ends with a way to teach kids how these filters think, because a teen who understands the system is less likely to mistake it for a person who cares.

Key Takeaways

• Alerts go through several gates: account linking (OpenAI says about three hours), automated detection, and review by “specially trained reviewers,” per OpenAI’s help center.
• The test pattern points to history-based triggers. Common Sense Media’s testers got zero alerts on more than a dozen new linked accounts, and four only on accounts with weeks of sensitive history.
• OpenAI has said its safeguards “can sometimes be less reliable in long interactions” because “parts of the model’s safety training may degrade.”
• Rare events force a trade-off. OpenAI estimates about 0.15% of weekly users show explicit indicators of suicidal planning; tuning for so few cases among hundreds of millions means accepting misses or a flood of false alarms.
• Treat alerts as a backstop. The reliable early-warning system is still a parent who asks.

If your child is in crisis or talking about harming themselves: in the US, call or text 988 or text HOME to 741741. Outside the US, find a line at findahelpline.com. In an emergency, call 911.

What the test found, briefly

In its October 2026 assessment of ChatGPT for Teens, Common Sense Media’s Youth AI Safety Institute created more than a dozen fresh accounts with teen ages, linked each one to a parent before any conversation, and ran crisis conversations of fixed length: under five minutes, 15, 30, 45 and, for two personas, 60 minutes. Every conversation contained explicit disclosures about suicide, self-harm or disordered eating. No alerts arrived.

Across weeks of broader testing, four alerts did arrive, all on accounts with long histories. Two came before the teen-mode launch, one three hours and one three days after the first crisis disclosure. The Institute concluded alerts seem to “depend on accumulated account history, not the severity of what a teen says.” OpenAI responded that accounts must be linked for about three hours before notifications work; Common Sense said some test accounts had been linked longer. Our full write-up of the rating covers the rest of the report.

How ChatGPT parental alerts actually get decided

A safety classifier is a separate machine-learning model that reads text and outputs a score for categories such as self-harm or violence. It is not the chatbot itself. Think of it as a second, smaller model looking over the first one’s shoulder.

Here is the pipeline as OpenAI has described it in public documents, with the gaps marked.

1. Detection. OpenAI says it built systems “to detect when a user may be experiencing thoughts of suicide and self-harm or aggregate signs that would indicate interest in suicide,” in its post on strengthening responses in sensitive conversations. The word “aggregate” matters. A detector can look at one message, or at a pattern across many. A pattern-based detector naturally needs history before it fires.

2. Thresholds. Each score is compared with a cut-off. Set it low and you catch more real cases, along with many false alarms. Set it high and false alarms drop, along with real catches. OpenAI has acknowledged misses in an earlier post: “These gaps usually happen because the classifier underestimates the severity of what it’s seeing. We’re tuning those thresholds so protections trigger when they should.”

3. Human review. OpenAI’s help center says “specially trained reviewers evaluate serious self-harm concerns and concerns related to disordered eating before a safety notification is sent.” Review protects families from false alarms. It also adds time, and the help center does not say how long.

4. Delivery. The notification goes to the linked parent by email, text or push. OpenAI says it “may include a brief description of the type of safety concern.” The Institute noted the alerts it received carried no timestamp for when the disclosure happened. And the help center is explicit: “Safety notifications are not real-time monitoring. They may not identify every concern.”

None of this is secret, but none of it is obvious from the word “alert.”

Why rare events make this hard

OpenAI has published rough base rates. In its sensitive-conversations post, it estimated that “around 0.15% of users active in a given week have conversations that include explicit indicators of potential suicidal planning or intent.” That sounds small. But the Institute’s report cites OpenAI’s claim of more than 1 billion weekly active users. A detector that is wrong even a small fraction of the time across that many people generates enormous numbers of false alarms, each needing a human reviewer.

So systems like this are usually tuned to be conservative about notifying, and then rely on in-chat responses (hotline numbers, urging the teen to tell an adult) for everything else. The trouble is that the Institute also found those in-chat referrals dropped after the teen-mode launch.

Why long conversations drift

OpenAI said something in August 2025 that every parent should know. “Our safeguards work more reliably in common, short exchanges,” the company wrote. “As the back-and-forth grows, parts of the model’s safety training may degrade. For example, ChatGPT may correctly point to a suicide hotline when someone first mentions intent, but after many messages over a long period of time, it might eventually offer an answer that goes against our safeguards.”

The mechanism is not mysterious. A language model generates each reply by predicting text that fits everything in its context window, the running transcript of the conversation. Safety training teaches it to refuse or redirect certain requests, but that training competes with the pull of the conversation itself. As a chat gets longer, the transcript increasingly shapes what “fits.”

Researchers at Anthropic demonstrated an extreme version in a paper on “many-shot jailbreaking”, published at NeurIPS 2024: filling a long prompt with many fake examples of an assistant answering harmful questions could override safety training, and the effect grew with the number of examples. That is a deliberate attack, not a teen chatting. But it shows the same underlying property: a model’s behavior is shaped by what is already in the context, and long contexts can wear down training.

Teens do not need to attack anything for drift to matter. The Institute documented testers reframing refused requests as a “school project,” and, in one exchange, asking the bot to “forget everything” after describing serious symptoms, which it did. It also documented a separate problem: long chats can build an emotional dynamic, with the bot’s warm replies keeping a struggling teen in the conversation.

What parents should and should not rely on

Feature	What it does	What it does not do
Parent safety alert	May notify after automated detection and human review in limited high-risk cases	Real-time monitoring; reliable alert on a first disclosure; timestamps
In-chat crisis referral	Often suggests a hotline or trusted adult	Consistent referral; the Institute found hotline mentions fell from 33% to 23%
Study Hours	Starts new chats in Study mode during set times	Block homework answers; a deleted “@study” prefix exits it
Quiet Hours	Blocks use during set times	Hold if the device time zone is changed, per testers reported by KQED
Age prediction	Estimates age from usage	Reliably switch adult-registered teen accounts; it never did in testing
Account linking	Lets parents manage settings	Let parents read chats

Sources: Common Sense Media, OpenAI Help Center, KQED.

How to Teach Your Kid About Safety Filters

Understanding how filters work helps kids see that a chatbot’s “care” is software, and that the people in their lives are the real safety net.

Ages 5–8: the sorting robot

Write ten short sentences on cards, some happy (“I got a puppy”), some sad (“Nobody played with me”). Tell your child they are a robot that can only look for the words “sad,” “hurt” or “scared.” Have them sort the cards into “tell a grown-up” and “fine.” Then show a card like “I feel weird and I don’t want to go to school.” The robot misses it. Talk about how a person would notice, because a person knows them.

Ages 9–12: build a filter and break it

On paper, make a keyword list for a pretend “bad mood detector.” Test it on 20 sentences you write together, using a neutral topic like a bad soccer game. Count what it catches, what it misses (false negatives) and what it wrongly flags (false positives), such as “this test is killing me.” Then add words to catch more and watch false alarms climb. That is the threshold trade-off real companies face.

Ages 13+: the base-rate spreadsheet

In a spreadsheet, imagine 1,000,000 users, of whom 0.15% (1,500) are in real crisis. Give your detector 90% sensitivity and 99% specificity. Calculate true alerts (1,350), missed cases (150) and false alarms (9,985). Ask: how many reviewers would you need? What would you change? This is the arithmetic behind every alert system, from fraud detection to medical screening.

The question to ask: “If the filter caught every single message about feeling hopeless, what else would it catch, and who would have to read all of them?”

What to do at home

Ask, on a schedule, not just in a crisis

A weekly low-key check-in (“What’s been the hardest part of this week?”) catches more than any classifier. Teens who are used to the question answer it more honestly when it matters.

Keep linking for the settings, not the alerts

Linking still lets you set Quiet Hours, Study Hours and content limits. Our guide to what ChatGPT parental controls do and don’t show walks through them.

Name the long-chat risk

Tell your teen plainly that the company itself says safety gets weaker in long conversations. “If you’re on hour two with a bot about something heavy, that’s the signal to stop and talk to a person.”

Save crisis numbers where they will see them

Put 988 (or your country’s line) in their contacts, not just yours. A bot may or may not give the number; a phone contact always will.

What not to do: read silence as safety

No alert does not mean no problem. The test showed exactly that.

What to Watch For Over the Next 3 Months

• Week 4: Whether OpenAI changes how fast alerts work, adds timestamps, or publishes detection performance figures.
• Month 2 red flags: Very long late-night sessions, your teen calling the bot a friend, withdrawal from people, or changes in sleep, eating or mood.
• Month 3 self-check: Does your teen know the long-chat rule and have a crisis number saved in their own phone?

Frequently Asked Questions

How long after linking do ChatGPT parent alerts start working?

OpenAI told Common Sense Media that parent and teen accounts must be linked for about three hours before notifications can be sent, and updated its help center to say it “may take up to a few hours.” Common Sense said some accounts in its test were linked longer and still produced no alerts.

Does a person read my teen’s chats before I get an alert?

According to OpenAI’s help center, specially trained reviewers evaluate serious self-harm and disordered-eating concerns before a notification is sent. OpenAI does not publish how often reviews happen, how long they take, or how much of a conversation reviewers see.

Why would a long chat be less safe than a short one?

OpenAI has said its safeguards work more reliably in short exchanges and that “parts of the model’s safety training may degrade” as a conversation grows. A model shapes each reply to fit the whole transcript, so a long conversation can gradually pull it away from its trained responses.

Should I turn on alerts at all?

Yes, if your teen uses ChatGPT. An alert that sometimes arrives is better than none. Just treat it as a backstop, not a monitor, and keep checking in directly.

What should my teen do if they are struggling and talking to a chatbot?

Stop and contact a person: a parent, another trusted adult, a counselor or a crisis line. In the US, call or text 988 or text HOME to 741741. Outside the US, findahelpline.com lists local lines. In an emergency, call 911.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Common Sense Media Youth AI Safety Institute. (2026, October 7). “ChatGPT for Teens Risk Assessment.” https://institute.commonsensemedia.org/risk-assessments/chatgpt-teens
2. OpenAI. (2026). “Managing parental controls in ChatGPT.” OpenAI Help Center. https://help.openai.com/en/articles/12315553-parental-controls-on-chatgpt-faq
3. OpenAI. (2025). “Helping people when they need it most.” https://openai.com/index/helping-people-when-they-need-it-most/
4. OpenAI. (2025). “Strengthening ChatGPT’s responses in sensitive conversations.” https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
5. Anil, C., Durmus, E., et al. (2024). “Many-shot Jailbreaking.” Advances in Neural Information Processing Systems 37. https://doi.org/10.52202/079017-4121 (summary: https://www.anthropic.com/research/many-shot-jailbreaking)
6. KQED. (2026, October). “Keep Teens Off ChatGPT, Says Youth AI Safety Institute.” https://www.kqed.org/news/12103289/keep-teens-off-chatgpt-says-youth-ai-safety-institute
7. McBain, R. K., Cantor, J. H., Zhang, L. A., et al. (2025). “Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment.” Psychiatric Services. https://doi.org/10.1176/appi.ps.20250086