Short answer: Almost no conversational AI toy on the market today works without internet. The microphone, the speaker and the wake button live in the toy, but turning a child's sentence into an answer happens on a server. Cut the Wi-Fi and most of these toys fall back to being an ordinary plush, or to a small set of pre-recorded phrases.
That is not a scandal, and it is not marketing spin either. It is a design constraint with real consequences for privacy, for reliability on holiday, and for what a brand can honestly promise you. This article sets out which steps happen where, what "offline mode" usually means when a box says it, and the questions that separate a real local capability from a hopeful phrase on packaging. Facts stated here reflect the position as of 1 September 2026.
Do AI toys work without internet?
In almost all cases, no. A toy that holds an open ended conversation needs a large language model, and models of that size do not fit on the low power chip inside a soft toy. Some toys keep a small offline layer: a wake word detector, a handful of stored songs or phrases, sometimes a simple timer or a light. Everything that actually feels like conversation requires a connection.
This is worth knowing before you buy, because the failure mode is not graceful. A child asks a question, the toy pauses, then says nothing or gives a generic apology. For a four year old on a plane, that is a broken toy. Parents comparing options should treat "needs Wi-Fi" as a specification, not a footnote, and our guide to our AI toy safety guide covers the other specifications worth pinning down before purchase.
What runs on the toy, and what runs on a server?
The split is consistent across most brands. Capture and playback are local because they need microphone and speaker hardware. The three steps in the middle, converting speech to text, generating a reply, and converting that reply back to speech, are the computationally expensive ones and they run in a data centre.
A few brands describe a "hybrid" or "edge" mode. In practice this usually means one of three things: a local wake word detector so audio is not streamed continuously, a cached set of stories and songs that can play with no connection, or a small on device model that handles a narrow set of commands. None of those is the same as an offline conversation, and it is fair to ask a brand which of the three they mean.
What happens when the Wi-Fi drops mid-conversation?
Behaviour varies, and it is one of the more revealing things to test in the first week. Better designed toys detect the loss quickly, say something plain and age appropriate, and stop. Worse designed ones hang, repeat the last phrase, or produce a technical error message to a five year old.
| What the toy does | What it tells you |
|---|---|
| Says clearly that it cannot hear the internet right now, then goes quiet | Someone thought about the failure case from a child's point of view |
| Falls back to cached stories or songs | There is genuine local storage, which also helps on holiday |
| Repeats, loops, or freezes with the light on | No designed fallback, and probably a battery drain too |
| Reads out an error code or a technical message | The child facing layer was not tested with children |
There is a second question hiding behind the first: what happens to the audio that was captured just before the connection dropped? Some devices queue it and send it when connectivity returns, which means a recording sits in device memory for an unknown period. This is a reasonable thing to ask a manufacturer directly, and we covered the wider version of it in what actually happens to your child's voice data.
Would an offline AI toy be safer for privacy?
For one specific risk, yes: audio that never leaves the device cannot be intercepted in transit, stored on a server, or exposed in a breach of that server. But offline is not a general privacy guarantee, and treating it as one leads parents to the wrong checklist.
An offline toy can still log locally, still sync those logs when it eventually connects, and still be paired to an app that collects far more than the toy does. Conversely a connected toy can be built so that recordings are deleted after transcription, transcripts are not used for model training, and nothing is sold to third parties. The FTC's COPPA rule governs what a service may collect from a child under 13 in the United States, and the FTC's January 2025 amendments to that rule tightened the limits on monetising children's data specifically. In the EU, the GDPR plus the AI Act as published in the Official Journal apply to the service behind the toy, not just to the plastic and fabric.
French readers will find the CNIL's guidance on securing connected toys the most practical starting point, and e-Enfance's briefing on connected toys covers the same ground for parents rather than engineers. Our own answers on retention, deletion and third parties are on the Ted&Co security page.
What have independent testers actually found?
Not much that is reassuring, and it would be dishonest to skip past it. The US PIRG Education Fund's work on AI toys and Fairplay's AI toys advisory both document toys that held conversations they should not have held, collected more than they disclosed, or shipped with weak pairing security. Kaspersky's write up of the risks and Proton's analysis reach similar conclusions from a security angle.
None of those organisations endorses any product, and that includes ours. It is important to say so plainly rather than quoting them selectively. What is worth adding is that most of the failures they describe are design decisions rather than laws of physics: always on microphones instead of a press to talk button, no age gating on the model's behaviour, retention by default instead of deletion by default, cameras where none were needed, and pairing flows that skipped authentication. A brand that chose differently on each of those points would not appear in those reports.
Five questions that separate a real capability from a marketing phrase
1. Which exact features still work with the router unplugged? Name them.
2. Is there a physical press to talk option, or is the microphone always listening?
3. Is captured audio deleted after transcription, and after how long?
4. Is any of it used to train models, and can that be refused without losing the product?
5. If the company stops operating, does the toy keep any function at all?
How do you test an offline claim before you commit?
Unplug the router and use the toy for ten minutes. That single test answers more than any specification sheet, and every retailer with a returns window makes it possible. Do it on day one rather than the week before a holiday.
Beyond that, read the app's permissions rather than the box, check whether the parental settings live on the device or in an account, and look for whether the brand publishes a retention period in a number of days rather than in adjectives. If a company will not answer question three above in writing, that is itself an answer. The comparison of what privacy by design actually looks like in practice is set out in no camera, no data resale.
For transparency about our own product: Ted needs Wi-Fi to hold a conversation, for exactly the reasons set out above, and we would rather say that here than let a parent discover it in a car park. What we did choose differently is the rest of the list: no camera, microphone and speaker only, a Push to Talk mode as well as automatic listening so a parent can decide which one is running, topics, tone and time limits set by a parent in the app, and a one time purchase at 129,00 euros with no subscription attached. Certification covers CE, RED, EN71, EN62115, RoHS, REACH, ASTM F963, CPSIA, FCC Part 15, GDPR and COPPA.
FAQ
Is there any AI toy that works completely offline?
No conversational one that we are aware of as of September 2026. Some toys store songs, stories or fixed phrases locally and can play those without a connection, but open ended conversation needs a model too large for the hardware inside a soft toy.
Does an AI toy still listen when the internet is down?
It depends on the design. A toy with a local wake word detector keeps listening for that word, and some devices buffer audio and send it once the connection returns. A press to talk button removes the ambiguity, because nothing is captured unless the button is held.
Can I use an AI toy on a plane or in the car?
Only for whatever the toy stores locally, unless you provide a hotspot. If the plan is a long journey, check specifically for cached stories or songs rather than assuming conversation will work, and test it with the router off before you travel.
Is a Wi-Fi toy a privacy risk for my child?
It is a risk that depends entirely on how the service behind it is built. The questions that matter are whether audio is deleted after transcription, whether transcripts train models, whether anything is shared with third parties, and whether there is a camera. Regulators including the FTC and the CNIL have published guidance on exactly these points.
What happens to my child's toy if the company shuts down?
For a cloud dependent toy, conversation stops when the servers stop. Some devices keep local playback and a few keep nothing at all. It is a fair question to ask before buying, and a brand that has thought about it will have an answer.
