Table of Contents
AI Risks for Families, Ranked by Evidence Not Headlines
AI risks for families ranked by how much evidence actually supports each one, what reduces each one, and why the media order is almost exactly upside down.
The order in which a parent encounters AI risks is almost exactly the reverse of the order in which they should act on them. Here are the AI risks families ranked by two criteria I can defend: how much real evidence supports the concern, and how much a household can actually do about it this month. Existential risk tops the headlines and sits ninth here. Unsupervised companion chatbot use by a struggling teenager barely makes the news and sits first. Nearly three in four teens have used an AI companion, and about a third have chosen one over a person for a serious conversation.
Key Takeaways
- Ranked by evidence and actionability, the top three are companion use by a vulnerable teen, image-based abuse involving minors, and over-granted permissions on AI agents.
- The strongest single dataset available to parents is Common Sense Media’s July 2025 report: nearly three in four teens have used AI companions, half use them regularly, about a third have chosen an AI over a human for a serious conversation, and a quarter have shared personal information.
- Regulators have ranked these too. The FTC’s September 2025 inquiry went to seven companies about companion chatbots and minors, not about superintelligence.
- The media order is roughly inverted. Coverage volume correlates with how dramatic a risk sounds, not with how much evidence exists or how tractable it is at home.
- The honest entry on learning harm is “mixed.” Anyone telling you the research is settled on whether AI use damages learning is overstating it in either direction.
How I ranked them
Three criteria, applied openly so you can disagree with the weighting.
Evidence quality. Is there published research, regulatory action or documented incident data, or is there a plausible story? Plausible stories can be right. They rank lower.
Household actionability. Can a family reduce this measurably in one afternoon with settings, conversations or configuration? A risk you cannot touch belongs lower in a parent’s priority list even if it is real.
Severity when it happens. Not frequency. The worst case, weighted by how recoverable it is.
Anything that scores high on all three goes to the top. That is the whole method, and it produces an order that looks nothing like a news feed.
AI risks for families ranked, one to nine
1. Unsupervised companion chatbot use by a teenager who is already struggling
This is first on every criterion and it is not close.
Common Sense Media’s “Talk, Trust, and Trade-Offs”, published July 16, 2025, found that nearly three in four teens have used AI companions, half use them regularly, about a third have chosen an AI companion over a human for a serious conversation, and a quarter have shared personal information with these platforms. Nearly half view them as tools rather than friends, which is reassuring and does not describe the other half.
Regulators have acted on this specific risk rather than any other. The FTC’s 6(b) inquiry, opened September 11, 2025, sent orders to Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI, asking how they “measure and monitor potential harms to minors before and after product launch,” how they restrict children’s access, and what steps they took “to evaluate the safety of their chatbots when acting as companions.”
What reduces it: account age configuration, so the right safety classifiers apply; knowing which platform your child uses for conversation rather than tasks; and one stated, repeated household message that telling a parent about an upsetting exchange is always safe. Our walkthrough of safety settings on the major assistants covers the configuration in about twenty minutes.
2. Image-based abuse involving minors
Second because severity is extreme, the capability is widely available, and the legal system has moved faster here than on almost any other AI harm.
The EU AI Act implementation timeline shows prohibitions on AI systems generating non-consensual intimate imagery and child sexual abuse material becoming applicable on 2 December 2026, carved out separately from the main February 2025 prohibitions. When a legislature separates a category and gives it its own date, that is a statement about priority.
What reduces it: a conversation about what to do if it happens, before it happens, including that the child is not in trouble and that the images should be reported rather than deleted. Our parent’s guide to AI-generated fake images covers reporting routes.
3. Over-granted permissions on AI agents
Third because it is rising fast, almost nobody has configured it, and the fix takes forty minutes.
The security framing is formal. The OWASP Top 10 for LLM Applications 2025 ranks Prompt Injection first and lists Excessive Agency at LLM06. Together those describe the actual failure: an agent reads untrusted content, treats it as instruction, and acts with whatever permissions it was given.
What reduces it: read-only by default, no payment method attached, per-action approval for anything that sends or spends.
4. Acting on a confidently wrong answer about something consequential
Fourth because the failure mode is well documented and the mitigation is a habit rather than a setting, which makes it slower to install.
OpenAI’s own research on why language models hallucinate attributes part of the problem to evaluation: “standard training and evaluation procedures reward guessing over acknowledging uncertainty.” The practical consequence is that fluency carries no signal about accuracy, and the categories where this costs most are health, medication, legal questions and anything with a number in it.
What reduces it: one hard household rule, narrowly scoped. In our house nothing about health or dosage comes from a chatbot, which is a rule kids accept because it is specific.
5. Degradation of the information environment
Fifth because it is measurable and diffuse. NewsGuard’s AI tracking centre counted 3,749 AI content farm news sites as of June 23, 2026, using criteria that include “strong evidence that the content is being published without significant human oversight.”
What reduces it: teaching the accountability question, “if this is wrong, who loses something,” rather than teaching detection, which does not work reliably.
6. Learning substitution, where the evidence is genuinely mixed
Sixth, and this is the entry where I want to be most careful.
Pew Research Center found that 26% of US teens aged 13 to 17 had used ChatGPT for schoolwork as of its survey run from September 18 to October 10, 2024, double the 13% in 2023, with 31% among eleventh and twelfth graders. That is exposure, not harm.
On harm, the research is early and contested. Stanford’s 2025 AI Index documents rising incidents and rare standardised evaluation, which is a statement about measurement rather than about learning outcomes. Teachers report concern. Studies disagree. Anybody claiming a settled answer in either direction is ahead of the data.
What reduces it: assignment design at school, and at home the distinction between summarising something you read and summarising something you did not.
7. Attention fragmentation from proactive agents
Seventh because the mechanism is real and the measurement comes from adults in offices. Mark, González and Harris at CHI 2005 found 57% of working spheres interrupted and about 11 minutes spent in one before switching. Nobody has run the equivalent study on children with proactive AI assistants.
What reduces it: route proactive output into a digest or a folder rather than a notification.
8. Labour-market disruption
Eighth on actionability, not on importance. It is a genuine economic question with a long time horizon and almost nothing a parent can do this quarter beyond the obvious: keep a child’s skills broad, keep them building things, and resist optimising a fourteen-year-old’s choices around a job market nobody can forecast.
9. Catastrophic or existential risk
Last, and I want to be precise about why.
It is not last because it is fake. The Future of Life Institute’s AI Safety Index, published July 17, 2025, graded seven developers and found none scoring above D in existential safety planning, with the panel concluding that “the industry is fundamentally unprepared for its own stated goals.” That is a serious finding about institutions.
It is last because household actionability is close to zero. There is no setting, conversation or configuration in your home that changes it. The appropriate response is civic rather than domestic: it belongs in how you vote and what you ask of institutions, not in your Sunday-evening device audit. Our piece on what to actually worry about with superintelligence separates the research question from the panic.
The whole thing in one table
| Rank | Risk | Evidence quality | Fix time at home | Media coverage rank |
|---|---|---|---|---|
| 1 | Companion use by a vulnerable teen | Strong survey data, regulatory action | 20 minutes | 5th |
| 2 | Image-based abuse of minors | Legal action, dedicated statutory dates | One conversation | 3rd |
| 3 | Over-granted agent permissions | Formal security taxonomy | 40 minutes | 7th |
| 4 | Acting on confidently wrong output | Published research from developers | A habit, weeks | 4th |
| 5 | Information environment degradation | Counted, 3,749 sites | Ongoing teaching | 6th |
| 6 | Learning substitution | Mixed and contested | School-dependent | 2nd |
| 7 | Attention fragmentation | Strong, but on adults | 10 minutes | 8th |
| 8 | Labour-market disruption | Speculative, long horizon | Not actionable | 9th |
| 9 | Catastrophic risk | Real debate, institutional evidence | Not actionable at home | 1st |
Media coverage ranks are my own estimate from reading this beat, not a measurement. The point stands without precision: the two risks a family can most effectively reduce receive less attention than the two it cannot touch.
What to do with this ranking
Spend your first hour on rows one to three
Not on reading more. Twenty minutes on account configuration, one conversation about what to do if an image appears, forty minutes auditing what every AI tool in the house has permission to touch. That is the highest-value hour available to a parent on this topic in early 2027.
Pick one hard rule, not five soft ones
Households that write comprehensive technology policies do not follow them. Households with one specific prohibition, stated with a reason, usually do. Choose the one that matches your highest-ranked worry.
Let the unactionable things stay unactionable
You are allowed to find existential risk interesting, worrying and not your job this month. Treating every risk as a household project produces exhaustion and no safety. The civic version of that concern is real and lives somewhere other than your settings menu.
Re-rank yearly, not weekly
The evidence base moves slowly and news moves fast. A ranking rebuilt every news cycle is a mood, not a plan.
What not to do
Do not use this ranking to dismiss a specific concern about your specific child. These are population-level priorities. If your eleven-year-old is having long conversations with a chatbot about something that worries you, that outranks everything on this list for your family, regardless of where the category sits in general.
What to Watch For Over the Next 3 Months
- Week 4: Watch for the FTC inquiry producing public output. When a regulator publishes what seven companies told it about minors and companion chatbots, that will be the best evidence any parent has had on risk number one.
- Month 2 red flags: Watch for your own household’s rank order changing for a reason other than a headline. A specific thing your child said or did should be able to move a risk up the list. A news story should not.
- Month 3 self-check: Ask whether rows one to three are actually done. Most families read a ranking, agree with it, and configure nothing. The difference between knowing and doing is one calendar entry.
Frequently Asked Questions
What is the single biggest AI risk for a family right now?
Unsupervised use of companion chatbots by a teenager who is already having a hard time. Common Sense Media found nearly three in four teens have used AI companions and about a third have chosen one over a person for a serious conversation. It also happens to be among the cheapest risks to reduce.
Should I worry about superintelligence?
As a citizen, reasonably. As a parent configuring devices this weekend, no. The Future of Life Institute found no major developer scoring above D in existential safety planning, which is a serious institutional finding and still leaves nothing for a household to adjust.
Does AI hurt kids’ learning?
The evidence is mixed and the honest answer is that nobody knows yet. Exposure is documented: Pew recorded 26% of US teens using ChatGPT for schoolwork in late 2024, double the 2023 figure. Whether and how that affects learning depends heavily on assignment design, and the research disagrees with itself.
What about deepfakes of my kid?
Second on this list, because severity is extreme even though frequency is lower. The EU AI Act gives prohibitions on AI-generated non-consensual intimate imagery and child sexual abuse material their own applicability date of 2 December 2026, separate from other prohibitions. Have the conversation before you need it.
Is it worse to over-worry or under-worry?
Both fail, differently. Over-worrying produces bans that get reversed and kids who do not tell you things. Under-worrying leaves configuration undone. The way out is to be specific: one or two named risks, with named actions, rather than a general posture.
How do I keep up without reading AI news every day?
Re-run a configuration check quarterly and ignore the rest. The underlying evidence base changes over months, not days, and almost nothing in a given week’s coverage changes what a family should do.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Common Sense Media. (2025). “Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions.” July 16, 2025. https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
- Federal Trade Commission. (2025). “FTC Launches Inquiry into AI Chatbots Acting as Companions.” September 11, 2025. https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions
- Future of Life Institute. “EU AI Act Implementation Timeline.” https://artificialintelligenceact.eu/implementation-timeline/
- OWASP. (2025). “OWASP Top 10 for LLM Applications 2025.” https://genai.owasp.org/llm-top-10/
- OpenAI. (2025). “Why language models hallucinate.” September 2025. https://openai.com/index/why-language-models-hallucinate/
- NewsGuard. “AI Tracking Center.” Updated June 23, 2026. https://www.newsguardtech.com/special-reports/ai-tracking-center/
- Pew Research Center. (2025). “About a quarter of U.S. teens have used ChatGPT for schoolwork – double the share in 2023.” January 15, 2025. https://www.pewresearch.org/short-reads/2025/01/15/about-a-quarter-of-us-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/
- Future of Life Institute. (2025). “AI Safety Index: Summer 2025.” July 17, 2025. https://futureoflife.org/ai-safety-index-summer-2025/
- Stanford Institute for Human-Centered AI. (2025). “The 2025 AI Index Report.” https://hai.stanford.edu/ai-index/2025-ai-index-report
- Mark, G., González, V. M., & Harris, J. (2005). “No Task Left Behind? Examining the Nature of Fragmented Work.” Proceedings of CHI 2005, ACM. https://www.ics.uci.edu/~gmark/CHI2005.pdf
