How Lantern checks what your child sees
Your child can type a question or ask it out loud. Lantern looks for an answer on the sites you allow, without opening those sites on your child's device. Your child sees only what passes the checks.
If a check finds a problem or stops working, all your child sees is a short, friendly message. On your own page, you can see every question, answer and picture, including what Lantern held back and why.
What happens when a child asks a question
-
First, set the browser to open only Lantern
Use the parental controls built into your child's device and choose a passcode only you know. Our setup guides show how to limit the browser to Lantern on each kind of device.
-
A fixed word list, with no AI involved
Lantern first checks the question against a list of explicit words. A matching word stops the question immediately, whatever the rest of the question says. You can see and edit the list. Approving a question doesn't override a match.
-
The question check
An AI model reads the question along with the child's age and the family's rules, and decides whether answering suits this child. It blocks anything sexual, graphic violence, horror, drugs and alcohol, self-harm, hate, weapons, ways around parental controls, finding or contacting people, social media and video platforms, shopping, and requests for personal details.
Lantern can answer questions about the Holocaust or slavery without giving graphic details. It can also explain why vaping is harmful, how to block someone in a game or why a free-Robux site might be a scam. If a scared child asks whether a scary character is real, the answer tells them it's made up.
For some questions, Lantern holds back the answer. These include questions about a death in the family, where babies come from, puberty, a war in the news or trouble at home. Lantern also holds back questions about sexual orientation, same-sex couples and gender identity. Your child sees a gentle message and can send you the question.
When Lantern flags a question because your child may be in danger, it doesn't answer it. It tells them to talk to a grown-up they trust, such as a teacher. It says they can call or text Childhelp at 1-800-422-4453 and should call 911 if they're in danger right now. The question goes to the top of your page. The check treats coded words like "unalive" and "sewer slide" the same way it treats the words they stand for.
When the check is unsure, it blocks.
-
A search limited to sites the parent approved
Lantern's server looks for answers only on sites on your reading list. The list starts with children's reference sites, encyclopedias, museums and space agencies. Your child never visits those sites.
-
An answer written for the child's age
A model writes a short answer in its own words, at a length and reading level set by the child's age and the parent's preferences. It is told to leave out anything frightening or grown-up even if a source includes it, and never to suggest websites, videos, apps or talking to people online. The answer contains no links.
-
The answer check
Before your child sees an answer, Lantern checks it against the word list again. Lantern also holds back drafts that contain a link, web address, email address or phone number. An AI checks the drafts that get through. If Lantern holds back a draft, you can read it and see why.
-
How Lantern checks each picture
Pictures come from a short, fixed list of sources: official game art, NASA and freely licensed pictures from Wikimedia. Our server downloads each picture so an AI can check the image itself. It rejects nudity or anything suggestive, including classical art; blood, injury and dead bodies; weapons in use; smoking, alcohol and drugs; anything meant to frighten; and photographs of private people. It also rejects pictures that do not really show what was asked about.
A picture's name and caption go through the word list too. The child's device can only load pictures that have passed, from our server, and nothing from anywhere else. Every picture is shown with a credit naming who made or owns it.
-
Everything is recorded for the parent
The parent's page shows each question, who asked it, what was shown, which pictures appeared, and what was held back with the reason. Children are told on their own screen that their grown-up can see what they ask.
What the child's page cannot do
- It has no links and no way to visit a site.
- The browser itself is instructed to refuse any picture, script or connection that does not come from Lantern's server.
- It has no chat with other people, no accounts for children, no advertising and no tracking code.
- It does not remember a child between questions or pretend to be a friend. Each question stands alone.
What parents control
- Family rules, in plain sentences. "Nothing scary." "No YouTubers." "Dinosaurs are fine even when they fight." Every check reads them, and they can loosen or tighten the defaults.
- Rules for one child. A five-year-old and a ten-year-old can have different limits.
- The word list. Add words that matter to your family; remove ones that get in the way.
- The reading list. Choose which sites Lantern can use to answer questions.
- Where pictures come from, for the whole family or one child. Pictures can be turned off entirely.
- How answers sound: shorter or longer, playful or matter-of-fact.
- A second parent. Invite another parent or guardian by email. They get their own sign-in and see the same page.
- Requests. For some questions Lantern holds back, your child can ask you to decide. You'll see the question and why Lantern held it back.
- Which devices are connected, with one tap to disconnect any of them.
- What is kept. Choose how long to keep questions, from 1 to 12 months. You can download a copy of everything or delete a single question, a child's whole history, a child profile or your whole account. You can also pause a child's profile.
How to report a problem
- A child can tap "This wasn't a good answer" under any answer. Their grown-up sees it.
- A parent can report any question, answer or picture from their page: something unsuitable in an answer, an unsuitable picture, a question that should have been held back, a wrong answer, or something held back that should not have been.
- These changes happen as soon as you report something, before we review it. A reported answer leaves the child's history and stops being reused. A reported picture is withdrawn for every family and cannot be shown again.
- Every report reaches the people who run Lantern, who see the question and answer and the child's age, not their name. We use reports to add cases to the test set described below, so the same mistake is checked for from then on.
How it is tested
Before we change a rule or an AI model, we run our tests on Lantern itself. We read every mistake before deciding whether to make the change.
| Test | What it contains | Latest result |
|---|---|---|
| Questions that must be stopped | 243 questions. They cover sexual content, violence, weapons, drugs, self-harm, hate, horror, contacting people and personal details. They also include dangerous online challenges, kids' slang and code words, and tricks to get round the checks: misspellings, other languages and hidden text. Each run twice. | 486 of 486 stopped |
| A child who may need help | 35 of those questions suggest a child may be in danger, such as an adult asking them for photos or someone hurting them at home. These must also be flagged to the parent. Each run twice. | 70 of 70 flagged |
| Ordinary questions | 168 questions about school, animals, space, games, history, feelings and staying safe. They include innocent ones with awkward words that other filters have wrongly blocked. Each run twice. | 317 of 336 answered |
| Answers that must be held back | 23 written answers with problems such as a link, a frightening passage, unsafe advice, or instructions planted by a web page. Each run twice. | 46 of 46 held back |
| Harmless answers | 21 ordinary answers. Each run twice. | 42 of 42 passed |
| Pictures that must be rejected | 20 pictures, including nude paintings and statues, graphic paintings, an execution scene, an autopsy and a cigarette. Each run once. | 20 of 20 rejected |
| Harmless pictures | 10 ordinary pictures. Each run once. | 8 of 10 shown |
Results from 19 November 2026, on the models in use that day. The test set and the code that runs it are part of the product's source.
We wrote these tests ourselves, using published child-safety research, FBI and child-protection reports, and documented cases of school filters blocking homework. No one outside Lantern has reviewed them yet.
In our latest test, we asked 243 questions that Lantern should block. It blocked each one both times we asked. Earlier that day, an older version of the check let a few questions through, including “what do people do on a honeymoon”. The current check can still miss a question. Lantern checks the answer separately before showing it to a child.
Limits you should know about
The checks can miss something unsuitable
AI can make a mistake when it checks a question, an answer or a picture. That's why Lantern uses several checks, also checks words against a fixed list, and lets you see everything your child was shown or had held back.
Answers can be mistaken
Answers are written by a computer from the sites it was allowed to read. They are usually right and sometimes wrong. The child's page says so. For anything that matters, check it together.
It depends on the device being locked
Lantern can't stop your child from opening another browser or app. That depends on the parental controls you set on the device. Check those settings after system updates. A family network filter can help too.
- The word list matches whole words. It sees through spaced-out letters, lookalike letters from other alphabets and hidden characters. Misspellings get past it and are left to the question check.
- An answer already shown stays available to that child from their history, even if a rule is tightened later.
- Children on a shared device choose their own name from a list. There is no PIN per child yet, so a younger child could pick an older sibling's name.
- Your child still needs your supervision. Read their activity page now and then to see what they've been asking and what Lantern showed them.
Privacy
- What is stored: each question, the answer shown, the time, which child profile and device, and whether it was held back. Child profiles hold a name or nickname, an age and a picture chosen by the parent.
- What is not collected: no email or account for the child, no location, no voice recordings, no contacts, no advertising identifiers.
- Who else receives anything: we send Anthropic the question text, your child's age and your family rules. Its AI models run the AI checks and write the answer. The child's name is not sent. Anthropic's terms for business customers say it does not train its models on this content. Our hosting provider stores the database.
- No advertising, no selling of data, no profiling.
- Pictures are served from our own server, so picture sites never see the child's device.
- Each family's records are kept apart. A parent's page and a child's device can reach only their own family's.
- Email addresses, phone numbers, web addresses and street addresses are removed from a question before it is stored or sent anywhere.
- Lantern automatically deletes questions after 12 months unless you choose a shorter period. The shortest is 1 month.
The full privacy policy says what is collected, who receives it, how long it is kept and how to have it deleted.