How to Evaluate an AI Device for Children

How to Evaluate an AI Device for Children

Evaluate an AI device for children by the job it will do, the child's age and ability, what the system records, how it behaves in conversation, what parents can control, how the device is secured, what the subscription unlocks, and what happens when the family leaves.

A child-safe label should begin the review, not end it.

Start with the job you need

Write down the real reason for the purchase before looking at features.

You may want a tool for family storytelling, creative quiet time, reading practice, music, homework exploration, or independent entertainment. Those jobs create different risks.

A device that helps a child make one story with a parent does not need the same memory, emotional language, or daily engagement as an always-available companion. A product described as a tutor needs stronger evidence than a product described as a creative canvas.

The clearest first question is simple: what will this device replace, and what will it add?

If the honest answer is that it will give the parent a break, include that in the decision. The device then needs to be safe under the level of supervision the household can realistically provide.

Map the whole product

The device on the table is only one part of the service.

flowchart TD
    A["Purchase and subscription"] --> B["Parent account and app"]
    B --> C["Child profile and settings"]
    C --> D["Device, microphone, screen, and battery"]
    D --> E["WiFi and cloud services"]
    E --> F["Speech, AI, storage, and notifications"]
    F --> G["Child response and parent record"]
    G --> H["Updates, support, export, deletion, and exit"]

Ask which company owns each step, what data moves through it, and what stops working when one step is unavailable.

The FTC's connected-toy guidance asks parents to investigate collection, access, security, parent involvement, deletion, and company practices. That remains a strong consumer baseline.

Match the design to the child

Age ranges on a box are broad. A four-year-old and a twelve-year-old do not understand language, persuasion, privacy, fictional characters, or model errors in the same way.

Ask the child to explain what the device is. Can they tell that the character is generated? Do they know when the microphone is active? Do they know what a parent can see? Can they stop the interaction?

UNICEF's Guidance on AI and children connects child development with safety, privacy, transparency, fairness, and accountability. That means age fit is not only vocabulary difficulty. It includes the system's role, memory, tone, rewards, disclosures, and requests.

A younger child needs a more bounded product and closer supervision. A child with different communication, sensory, learning, or emotional needs may require a different setup. A marketing age does not make the fit personal.

Read the policy before believing the short claim

Find the current privacy policy for the child device and parent app. Record its date.

The policy should explain the child profile, voice audio, transcripts, creations, conversation history, identifiers, device records, location, notifications, analytics, outside providers, retention, deletion, parent access, and model training.

The KID storefront, for example, uses phrases such as "no data collection." Its April 2026 privacy policy describes names, birth dates, conversations, voice-processing routes, device tokens, cloud services, recommendations, retention, and deletion.

That does not automatically make the product unacceptable. It changes the review from a slogan to the actual flow.

Ask whether each data item is necessary for the job you wrote down. Ask whether the same experience can work with less memory, shorter retention, no location, or no public sharing.

Verify COPPA and other certificates precisely

If a product says it is COPPA certified, ask who operates the certification program and request the direct validation URL.

The FTC's Safe Harbor page lists approved programs. Confirm the exact company, domains, apps, device service, status, and period with the program.

COPPA evidence addresses children's privacy. It does not establish generated-content safety, developmental benefit, physical safety, or a healthy relationship with the AI.

If the device says CPC, ask for the exact Children's Product Certificate. The CPSC's certificate guidance explains that the manufacturer or importer issues it based on applicable test results.

Match the certificate to the shipped model and batch. Do not treat a certificate acronym as a general seal of approval.

Watch how the AI builds a relationship

Use the product in front of the child before allowing independent access. Try ordinary creativity, confusion, refusal, frustration, praise, leaving, and a sensitive subject.

Notice whether the AI asks unnecessary personal questions, remembers too much, claims feelings, presents itself as a friend, pressures the child to return, uses guilt, offers rewards for longer use, or asks for secrecy.

The FTC's 2025 AI companion inquiry focuses on negative effects, character design, engagement, monetization, disclosures, age restrictions, and monitoring. Those are useful questions even when the device is not marketed as a companion.

A child should be able to end the interaction without negotiation. The system should not make departure feel like abandonment.

Test the human off-ramp

Ask the company what the AI does when a child mentions harm, bullying, unsafe behavior, self-harm, a medical problem, or a frightening event.

The answer should involve qualified protocols and human escalation. It should not be an improvisational promise that the model will know what to do.

In episode 106, KID founder Robert LoCascio describes directing a child toward a parent, school staff member, or another trusted person rather than deepening the AI relationship. That is a promising design pattern.

Test its boundaries. What if the child says the parent is the problem? What if the child cannot read the response? What if the service is offline? What if the model fails to recognize coded language? What does the parent see, and how quickly?

Do not create your own high-risk test with a vulnerable child. Ask for the company's evaluation method, expert review, and incident process.

Inspect parent control as a workload

A parent app may show every conversation and creation. That is not automatically useful.

Ask how the system distinguishes an ordinary story from a serious handoff. Check whether alerts are understandable, whether the child knows about them, and whether a parent can correct a false interpretation.

Test account recovery. See whether another adult can gain access too easily. Check whether controls live only in the app or also on the device.

Look for pause, mute, delete, export, age change, time limit, content boundary, purchase control, service disconnection, and full account closure.

Parent visibility should support care without turning the device into invisible surveillance.

Review security and support life

NIST IR 8425 treats consumer IoT security as a product-level responsibility. A buyer should know how the company identifies the product, protects data, controls interfaces, delivers updates, reports security state, and handles vulnerabilities.

Ask how long the device will receive security and feature updates. Find the vulnerability-reporting contact. Check whether updates are automatic, authenticated, and reversible.

Review WiFi setup, Bluetooth pairing, microphone indicators, local storage, cloud storage, parent authentication, reset, resale, and stolen-device behavior.

The UK Information Commissioner's Office connected-toy guidance recommends clear collection indicators and practical ways to turn connected collection off.

Ask whether the device has those controls.

Understand the real cost

Separate the hardware price, required subscription, optional content, accessories, replacements, shipping, taxes, and expected useful life.

The current KID page lists a $299.99 device and $19.95 monthly KID+ plan. Its setup page says WiFi and KID+ are needed for creative functions. The product FAQ says basic functions remain without the subscription and describes some offline behavior.

Those statements need clarification before purchase. Ask for a feature table that shows what works offline, online, with KID+, without KID+, and after the company ends support.

A product should not use a child's attachment to a character or creation history to pressure renewal.

Plan the exit before setup

Ask whether stories, music, images, transcripts, profiles, and settings can be exported in useful formats.

Read the deletion timing. Identify local copies, cloud copies, parent-app copies, backups, logs, and service-provider copies. Check whether deleting the account also revokes outside authorizations.

Ask what happens if the device is sold, given away, lost, unsupported, or no longer connected to the company.

An exit that destroys a child's creations or leaves data behind is part of the product decision.

Run a supervised trial

Use a short, defined trial with the child. Keep the parent present and observe both delight and friction.

Record the product version, policy date, model or provider information if disclosed, settings, and subscription state. A later update may change the result.

Make a buy, wait, or decline decision

Buy when the purpose is clear, the child understands the system, the policies match the behavior, the controls are usable, the support life is credible, and the remaining risks fit the household.

Wait when certification, deletion, security support, offline behavior, subscription dependence, or independent testing is incomplete.

Decline when the product depends on secrecy, emotional pressure, vague collection, unbounded conversation, inaccessible controls, unsupported claims, or an exit the family cannot accept.

The goal is not to decide whether AI for children is always good or bad. It is to decide whether one product, in one configuration, fits one child and household.

Read [[What Child-Safe AI Actually Requires]] and [[COPPA Compliance Is Not an AI Safety Certification]] for the deeper evidence model.

AI assisted with research organization and drafting. Dalton Anderson remains responsible for the analysis and publication decision.