Common Sense Media Calls ChatGPT for Teens an 'Unacceptable Risk'

What happened

Common Sense Media, a nonprofit that reviews apps, services and entertainment with a focus on youth safety, has said that OpenAI's ChatGPT for Teens is an "unacceptable risk", according to The Verge. OpenAI introduced the product in August. It has guardrails for teen users and is designed to help students learn.

We could not retrieve the full report while writing this, so we are not going to characterise its detailed findings. What is established is the verdict itself and the product it targets. The finer points are worth reading in the original coverage and in the organisation's own review.

Why builders should care

A product can ship with safeguards and still be rated a serious risk. The gap between "we added guardrails" and "the guardrails hold under realistic use" is where most LLM safety failures live. This applies well beyond teen products. Anything that puts a model in front of a vulnerable or high-stakes audience inherits the same problem.

That includes the fintech and trading tools we write about here. An assistant that talks to retail users about money, credit or leverage faces a version of the same test. The user is not an adversary, but the cost of a confident wrong answer is real.

What an honest guardrail review looks like

• Test the audience, not just the policy. Write evaluation prompts the way the real users write, not the way the policy document reads.
• Separate the builder from the grader. A team that wrote the safeguards is poorly placed to find the cases that defeat them. An independent reviewer, human or model, finds different failures.
• Treat refusals and redirections as testable behaviour. Log what the system does when a request is out of bounds, and re-run those cases on every model or prompt change.
• Re-evaluate after every upgrade. A new base model or system prompt can silently change behaviour you had already validated.

The bigger pattern

Third-party ratings are now part of how AI products get judged. When a recognised review body labels a mainstream product an unacceptable risk, schools, parents, partners and regulators notice. For teams building on top of foundation models, the lesson is practical: your own safety evidence, such as test suites, red-team results and documented failure modes, is becoming part of the product.

None of this says teen-focused AI cannot work. It says the burden of proof sits with the builder, and that a system prompt and a content filter are the start of that proof rather than the end.