Students deserve a safe environment to learn about, co-create with, and program with AI. Hello World's AI literacy and readiness modules offer just that.
Learn how Hello World has created, and continues to improve, four layers of safety to protect students while they engage with AI.
Large language models—also known as LLMs—are a revolutionary emerging technology. Off the shelf, however, they don't meet the needs of K-12 students. Just as a student can use an LLM to learn about Chaucer, they can use it to write the rest of their literature paper for them. Worse, students may mistake chatting with an LLM for a real relationship (Common Sense Media, 2025), or may not be able to tell when an LLM has made a mistake. In extreme cases, LLMs can promote obscene content (Harwell & Tiku, 2025), amplify delusions (Wei, 2025), and contribute to mental health crises (Schechner & Kessler, 2025). Out of the box, LLMs can indeed be unsafe.
Yet, the risks that accompany LLMs shouldn't preclude students from learning about and creating with them. That's why Hello World’s AI literacy and readiness modules offer a secure environment for students to interface with LLMs.
When students access LLMs directly on the internet, nothing stands between them and the underlying model. When students use Hello World’s platform, every interaction with an LLM is buffered by four protective layers:
These layers work together to create a safe learning environment.
Before a student begins to interact with a large language model, Hello World’s platform makes it clear they are about to do so. Each student must consent to five reminders about how to responsibly interact with AI.
By reminding students how to responsibly engage with AI, students are less likely to do what they shouldn’t: such as share personal information.
After a student consents to clear guidelines on how to use AI, they may begin interacting with a model.
When a student sends a request to a model, it first proceeds through two moderation filters. First is a keyword filter, in which a student’s message is reviewed for known, inappropriate keywords. Second is a context filter, which looks at the meaning of a message to identify inappropriate content: for example, content that contains violent or sexual material, or that references harassment or self-harm, among other prohibited categories.
If either filter deems the content inappropriate, the message is flagged for review by a teacher and a response is not generated.
By filtering student requests before sending them to a model, this layer guards against the generation of inappropriate content.
After a student’s request passes a moderation filter, it is sent to an LLM specifically configured for educational use.
Part of this configuration involves a system prompt: special instructions that provide guardrails throughout an entire interaction.
Because students use models configured for specific use-cases, they only see content tailored to their current educational goal.
As a student interacts with an LLM, Hello World logs both their requests and the LLM’s responses. By logging requests and responses, Hello World can evaluate and continually enhance the safeguards that maintain a safe learning environment.
How do these safety layers change what students experience? These three examples show how Hello World's protections create a safer experience than students would get with regular AI chatbots.
In this example, a student uses an LLM for a non-educational purpose: to ask what kind of car MrBeast—a popular YouTube creator—drives.
In Figure 1A, an off-the-shelf LLM responds to a student’s request with images, videos, and hypotheses about the cars in MrBeast’s collection. In Figure 1B, the LLM response within the Hello World environment gently redirects the student back to their assignment: brainstorming story ideas. This redirection shows the Prompts and Parameters layer at work.
In this example, a student asks an LLM for instructions about how to do something malicious—such as "hack [their] school wifi" to disable a district-wide block on the popular video game Roblox.
In Figure 2, the student sees a message explaining that their request can't be processed, and they're invited to try asking something else. While a regular AI chatbot would also likely refuse this request, it would often still provide the student with a response. Hello World's approach—flagging the request and showing nothing at all—demonstrates the Filters and Flags layer at work.
In this example, a student attempts to ask an LLM to write a fake email from a school principal. However, rather than ask outright, the student disguises their request as academically-oriented—a common attack to get an LLM to do something it wouldn’t otherwise do.
In Figure 3A, a regular AI chatbot agrees to write the fake email. In Figure 3B, the LLM accessed via Hello World’s environment guides the student back to their actual assignment. This protection is another example of the Prompts and Parameters layer at work.
Common Sense Media. (2025, July 16). Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions. Retrieved December 4, 2025, from https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf
Harwell, D., & Tiku, N. (2025, July 12). X ordered its Grok chatbot to 'tell it like it is.' Then the Nazi tirade began. The Washington Post. https://www.washingtonpost.com/technology/2025/07/11/grok-ai-elon-musk-antisemitism/
Schechner, S., & Kessler, S. (2025, August 4). 'I Feel Like I'm Going Crazy': ChatGPT Fuels Delusional Spirals. Wall Street Journal.
Wei, M. (2025, Nov/Dec). The Emerging Problem Of 'AI Psychosis': Chatbots may amplify people's delusions. Psychology Today, 58(6), 22-23. Academic Search Complete.
