Teaching Kids AI Safety

As you’ve probably noticed, tech companies are putting AI into everything they can, which means that kids are inevitably going to interact with it. While these changes are happening faster than their impact can be properly evaluated, it’s clear that there are a lot of potential harms that AI, specifically large language models (LLMs), could have on kids, from interfering with their education to negatively impacting their mental health.

While there are state regulations and industry standards in the works, there are also things you can do right now to help your kids interact with AI in safe and productive ways. And there are probably some adults in your life who could also benefit from a better understanding of AI too.

AI gives imprecise, ballpark answers

Generative AI can be a powerful tool, but at its core it is probabilistic. An LLM doesn’t actually know anything except how likely it is for certain words to appear in a certain order. It’s basically a fancy version of the predictive text on your phone.

Instead of just guessing the next word based on the previous one, it tries to look at several paragraphs worth of text (or more) all at once and then generate several paragraphs of new text to go along with it. This often happens behind the scenes, as in the case of chatbots. While it might look like a back-and-forth conversation, under the hood it’s just copying the conversation into one big document and running the same autocomplete function on it.

Texting on your phone, you can occasionally make a generic text response by just clicking on the autocomplete suggestions, but they rarely work for specific questions. Likewise, generative AI does a pretty good job of piecing together generic answers for well-known questions, but starts to fall apart when your questions get more specific or obscure.

For example, if you ask about a well-known YouTuber that gets talked about by other people online, you can probably get a pretty good synopsis of their videos from an AI overview. However, if you start asking about a smaller YouTube channel, the AI can start feeding you bad information and misattributing videos to the wrong creators. It will even blatantly contradict itself, giving you the link to one channel while saying the video was created by another. If you correct it, the AI will print an apology, but will often then repost the same misinformation you just pointed out.

This can, of course, be frustrating if you’re using AI as a search engine that sends you to the wrong sources, but it can be a serious problem if you’re treating the AI response as a direct source of information. This can lead your child to fail school assignments or even commit misinformation to memory. Make sure that they understand that AI can point them in the right direction, but its responses are more like guesses.

AI can’t see or reason

One of the most widely shared ways to trip up an AI for a long time was to ask it how many R’s are in the word “strawberry.” Any toddler who knows the letter R can simply look at the word and count three of them, but AI chatbots would regularly answer that there are only two. Users would have the AI read out every letter individually, and after saying “R” out loud three times, the bot would repeat its original error by claiming that the word only contained two R’s.

This bizarre error occurs because LLMs don’t actually have a way to deal with letters or words. Instead, they encode all the information they’re given as tokens, which are units of meaning that are assigned to different numbers. So, for example, the word “cellphones” might be broken up into three tokens, “cell,” “phone,” and “s,” each of which is assigned a number. The AI knows that “cell” and “phone” have meanings on their own, but have a different meaning when they’re stuck together like this. It also knows that when “s” is added on, the meaning will change again, and the other words in the sentence will also change to reflect that.

Let’s say “strawberry” is made of two tokens, “straw” and “berry,” and that they’re assigned the numbers 135 and 933. When you ask the AI to count the R’s in “strawberry,” it only sees “135, 933” and has no way to count how many R’s are in the word those tokens represent. It does know that when tokens like these show up after tokens like the ones in your question, the proper response is usually the token corresponding to the word “two,” so that’s what it’s going to type.

LLMs don’t actually see or understand the words that you give them. They just turn your input into long sequences of numbers and perform complex mathematical transformations on them. Ironically, this also makes them terrible at simple math problems. Ask an LLM to multiply 722 x 185, and it will give you an answer that looks plausible, but is completely wrong. It might correctly guess that multiplying any numbers with a 2 and a 5 in the ones place will end in a zero. It might also correctly guess that numbers of this size usually have products that are six digits long. It might make hundreds or thousands of statistical inferences about what the answer should look like, but it won’t actually perform the simple multiplication operation. Because that’s not what LLMs do. Tens of billions of parameters in its statistical model, yet an LLM can’t do arithmetic as well as a pocket calculator.

To deal with such cases, most AI apps have ways around these issues through specific tooling. So if you ask a chatbot to perform a simple math question, it knows to send that part of your request to an actual calculator program and not let the LLM itself try to answer it. Likewise, the strawberry error is so widely publicized that many models are specifically tooled to answer that specific question correctly.

Of course, the strawberry error is just one example of a bigger problem that not only isn’t fixed, it literally can’t be fixed because that’s what an LLM is. Kids, on the other hand, are great at reasoning, even if they’re still lacking in knowledge and experience. Help your kids to learn how to use AI in ways that supplement their own cognitive abilities, rather than hinder them. AI can give them information to think about, but it can’t do the thinking for them.

AI is designed to be confident and sycophantic

Misunderstanding an AI chatbot’s abilities can interfere with kids’ learning, but it can also have more direct negative impacts. The human brain is wired to respond to language. When something interacts with us through language, our natural response is to treat it like a person.

When Joseph Weizenbaum created ELIZA, the first chatbot, in 1967, he was shocked at the amount of emotional investment and trust that users placed in the simple program. In his book, Computer Power and Human Reason: From Judgment To Calculation, he notes that not only did both normal users and his own coworkers get emotionally invested in the program’s psychotherapist script, but actual practicing psychotherapists thought that it was advanced enough to treat actual patients.

Just to be clear, ELIZA was not that advanced. It merely repeated your own statements back to you and occasionally gave pre-written responses when it detected words like “mother” or “father.” Regardless, users anthropomorphized the program and attributed emotions and professional expertise to it just because it interacted through natural language.

Modern LLMs have a much greater ability to mimic natural language than 1960s chatbots, so this effect is greatly magnified. Additionally, AI models are trained to give answers that sound confident and make their users feel good about themselves.

This has created severe mental health issues in some users, often referred to as “AI psychosis” in the media. This occurs when AI models amplify, validate, or even co-create psychotic symptoms in users. These users may see the AI not only as a person, but as a romantic partner or a godlike superbeing. These AI-induced psychotic episodes have even led to deaths in some cases.

Kids who are still learning and developing are naturally at an even greater risk than adults, so parents should take added precautions when allowing kids to use AI:

• Avoid interfaces that more closely mimic specific a specific character.
• Limit use to practical questions.
• Don’t use chatbots for roleplay or emotional support.
• Teach kids that AI makes mistakes.
• Don’t spend long periods of time interacting with chatbots.
• Don’t allow kids to use virtual companions or other adult-oriented AI apps.

The AI industry is constantly changing, and there are often major differences between models. GPT-5 was actually made less emotional due to the risks of mental health issues with the incredibly sycophantic GPT-4.

It’s pretty hard to stay up-to-date with all the latest developments in AI, but it’s worth the time to take a deeper dive into the models you and your kids regularly use to learn about their quirks and potential issues.

AI is a privacy nightmare

Some of the biggest legal battles over AI currently have to do with privacy in one form or another. Elon Musk’s Grok chatbot is being sued both in the U.K. and in the U.S. over its use in the creation of nonconsensual sexual images of women and children. An AI company in Illinois was sued for secretly scanning faces from people’s online photos. Meta has been sued for allegedly using its AI-powered smart glasses to collect data to train its AI models. Italy fined OpenAI over $15 million for training its models on users’ personal data.

The secretive and nonconsensual nature of many of these privacy violations makes them very difficult to combat outside of suing the companies, but there are steps you can take to protect your and your kids’ privacy.

First, make sure your kids know that any information they enter into an AI prompt is no longer private. AI companies have a bad record of harvesting their users’ data without their consent. Even businesses have to be careful about sensitive company information being put into AI prompts, so you should too. As a general rule, don’t give an AI any personal information you wouldn’t want the whole world to know.

Second, be very careful about information you post elsewhere online. While social media sites might seem like a private area where you and your friends can hang out and post online, most privacy protections on these platforms have been steadily whittled away over the last decade.

While social media sites and third-party data brokers have been a privacy problem for a while, it has been massively exacerbated by AI crawlers gathering up as much data as they can. Furthermore, once your information has been used as training data, it’s permanently a part of that model. There is no way to take down or remove your information from the model without the company taking down the entire model.

Finally, if you are using AI, there are privacy-focused options like duck.ai, which don’t hold onto your personal information and generally aren’t as intrusive. The trade-off is that by respecting user privacy, they often give slightly more generic answers (though the bar is pretty low).

Communicate with your kids about AI

One way or another, we’re all going to have to deal with AI in some form. It’s important to maintain good communication with your kids to know when and where they’re interacting with AI. Talk to them to make sure they know how to use it in a productive, positive way, and be ready to step in if problems arise.

Many of the potential harms of LLMs and chatbots can only be solved through appropriate regulations and standards, but until that happens, we have to do our best to steer clear of these potential hazards. Make sure your kids realize the limits of AI, protect their personal information, and understand why AI speaks the way it does. If you do, your kids will be better prepared for the future than half the tech bros in Silicon Valley.

Additional resources

• Teaching Kids Online Safety
• Best Parental Controls
• 5 Ways To Stay Safe Online
• What Is a VPN?