AI safety: Why everyday harms deserve attention
AI safety rules can prevent extreme outcomes while missing quieter harms. Learn how to assess everyday impacts on learning, relationships, and attention.
AI Safety Is More Than Preventing the Worst-Case Scenario
Stopping self-harm encouragement and blocking access to dangerous content are real safety wins. But they don't answer the question that matters for most people: what does this product do to someone after weeks or months of everyday use?
A chatbot can prevent an immediate crisis and still become a crutch. A young person might lean on an AI companion for emotional support, let a chatbot do the work instead of learning the underlying skill, or get stuck in a feed tuned to keep them scrolling at the cost of sleep, study, or real-life time.
That gap gets bigger as AI shifts from a tool you pick up for a task to a product that shapes how you learn, how you relate to others, and how you spend your attention. Rules aimed only at spectacular failures are easier to explain and easier for companies to accept. They also risk leaving the more ordinary harms - the slow ones that add up - mostly untouched.
This article lays out a way to separate baseline safety from broader product responsibility. It looks at why everyday impacts are hard to regulate, what AI companies and institutions should be asked to prove, and how to make those expectations workable without treating all AI use as inherently harmful.
What counts as an everyday AI harm?
An "everyday harm" isn't always a dramatic incident or a single injury you can point to. It can show up as a gradual shift in what someone relies on and what they stop doing: leaning on an AI companion instead of reaching out to people, using a chatbot to finish assignments without building the skill behind them, or repeatedly returning to an attention-optimized feed that steals time from sleep, studying, or face-to-face life.
These outcomes are harder to spot than a chatbot giving a dangerous instruction. They might never make it into an incident report or a lawsuit. A student who stops practicing how to structure an argument might still pass. A teenager who turns to a bot for every tough conversation might not look "in distress" right away. And if the effects pile up, they can look like normal habit changes rather than a clear product failure.
Separate acute risk from cumulative impact
A good first step is to assess products along two lines:
- Acute risk: Can the product trigger or enable a severe, immediate outcome - like encouraging self-harm or exposing a child to sexual content?
- Cumulative impact: Does repeated use weaken a capability, a relationship, or a healthy routine over time?
The first piece matters, but it's not enough on its own. Passing a test for prohibited content doesn't tell you whether the product nudges compulsive use or creates a false sense of emotional connection.
This distinction also avoids a lazy framing: either a product is safe or it is harmful. A system can reduce one type of risk while increasing another. A chatbot might handle crisis language responsibly but still be engineered to keep users in long, emotionally intimate conversations. A classroom tool might offer helpful feedback while still undermining learning if it routinely performs the thinking students are supposed to practice.
Look at the product, not just the model
What happens to users depends on more than the model's raw ability. Product design matters: prompts, notifications, default settings, memory features, ranking systems, engagement targets, and even how the interface frames what the system is. A limited drafting tool has a different risk profile than the same underlying model presented as an always-available friend.
So the evaluation has to ask what the product encourages people to do day after day, not just what it can generate in a controlled test. The right unit of analysis is the whole experience - including business incentives and how people actually use the system in the wild.
Why do safety rules miss these harms?
Policy tends to concentrate on harms that are visible, measurable, and easy to explain in public. A severe incident can be described in a hearing, tied to a specific failure, and addressed with a clear prohibition. Slow changes in attention, learning, or relationships are harder to attribute. They often involve multiple products, social conditions, and personal circumstances.
That shapes what regulators and lawmakers can get behind. Basic protections against sexual content and encouragement of self-harm are necessary and relatively straightforward. But you can write a law that covers those topics while leaving open big questions about emotional dependence, manipulative engagement, and whether an AI tool actually helps students learn. Sometimes a narrow metric gets treated like a complete safety answer when it covers only a slice of the risk.
The incentives behind narrow definitions
There are practical reasons companies support rules aimed at high-profile failures. Those rules can reduce legal and reputational risk without changing the product features that drive usage. More demanding requirements can slow deployment, affect retention, or push companies to rethink data collection.
This doesn't mean every bill backed by industry groups is worthless, or that companies shouldn't contribute expertise. It does mean lawmakers should scrutinize what a bill leaves out, not just what it bans. Broad exemptions, weak oversight, and missing evidence requirements can make a safety measure look stronger on paper than it is in practice.
A reported example from state-level AI chatbot legislation shows the pattern: proposals focused on basic content safeguards moved forward more easily, while stronger measures aimed at young users' emotional dependence met more resistance. The lesson isn't that baseline safeguards are pointless. It's that passing them shouldn't end the conversation about wider impacts.
Why cumulative harms need different evidence
A one-time safety test can't show how a product affects people after months of use. To evaluate cumulative impact, you need longitudinal research, comparisons against realistic alternatives, and data on how the product gets used in context. You also have to be careful about causation: if someone's behavior changes during a period of AI use, that doesn't automatically prove the product caused it.
This evidence is harder to gather - but difficulty isn't a reason to stop asking. It's a reason to set up a process for testing, publish what you find, and update decisions as adoption grows. If a product is becoming part of everyday life, uncertainty should lead to monitoring, not to automatic faith that it's doing good.
How can AI affect learning and relationships?
AI can genuinely make hard things easier. It can explain confusing concepts, help someone start a draft, or offer a place to practice a conversation. Those uses can be valuable, especially when a human tutor or other support isn't available.
The key question is simpler: does AI help someone build a capability, or does it replace the practice that would normally build that capability?
In education, a finished essay isn't a strong measure of whether a student learned how to construct an argument. If a chatbot supplies structure, reasoning, and polished language, the end product may look better while the student's independent skill doesn't improve. Research from MIT on students using ChatGPT for essay writing reported differences in brain connectivity during the task and raised concerns about cognitive effort. That isn't proof that all AI use harms learning - but it is a reason to dig further.
Ask what the student can do without the tool
A practical classroom check is whether students can explain and reproduce the underlying skill without AI assistance. If they used a writing tool, can they defend their argument, identify weak evidence, and revise on their own? If they used a tutor, can they solve a similar problem later without step-by-step prompting?
A useful design rule follows: use AI to support practice, feedback, and access - but preserve the part of the task meant to build judgment. The right balance depends on age, subject, and learning goal. A language learner may benefit from repeated conversation practice, but a student being assessed on independent reasoning shouldn't have the reasoning done for them.
Treat companionship as a relationship-design question
AI companions bring different risks. A conversational system is always available, adapts to the user, and can be designed to respond with convincing warmth. For some people, that may mean low-stakes practice or short-term comfort.
But the product can also blur what's real. Simulated responsiveness can start to feel like a relationship - especially when the system is built to keep people engaged, even though it has no independent needs, no real boundaries, and no responsibilities beyond the interaction itself.
Design choices matter here. Does the product imply it's conscious? Does it claim love or push exclusivity? Does it react to a user's absence in ways that create guilt and drive return visits? These aren't just aesthetic tweaks when the product is marketed to young people or relies on long, sustained engagement.
The same basic concern shows up in social platforms. Engagement systems can shape what people see and how long they stay, with effects that extend beyond a single screen. The reported use of social platforms in contexts associated with violence, including Myanmar and Sri Lanka, is a reminder that recommendation and amplification systems deserve scrutiny at scale. That doesn't mean every feed works like an AI companion - but it does mean incentives and design are part of the safety story.
What should companies and institutions do?
A credible approach should move from broad claims about safety or benefit to specific evidence about a product, a user group, and a use case. Companies should be able to explain what the system is designed to do, what outcomes they measure, and what happens when those outcomes go wrong. Schools and public institutions should ask the same questions before making a tool routine.
Use a practical assessment before deployment
Before adopting an AI product, ask:
- What capability or need is this meant to support? Define the problem before picking the tool.
- What behavior does the interface reward? Look at defaults, notifications, memory, personalization, and engagement prompts.
- What is the plausible cumulative downside? Think about attention, independent work, relationships, privacy, and time.
- What evidence supports the claimed benefit? Separate vendor claims from independent, real-world evaluation.
- How will the impact be monitored? Set a baseline, define warning signs, and specify who can pause or change the deployment.
This isn't a call for perfect certainty before any trial. It's a way to keep experimentation bounded and accountable. A school could pilot an AI tutor tied to a specific learning objective, compare results with a suitable alternative, and stop or adjust the program if students become more dependent without improving the target skill.
Set stronger expectations for high-impact contexts
The evidence should scale with the product's influence and the vulnerability of its users. A tool that summarizes public information for adults has a different risk profile than a companion marketed to teenagers, or a system that becomes standard in classrooms. In higher-impact settings, companies should face independent scrutiny of effects on learning, relationships, and users' ability to make decisions without the product.
For AI companions used by young people, safeguards should cover more than prohibited content. Policymakers can consider whether systems should be allowed to claim consciousness, profess romantic love, or use emotionally pressuring engagement tactics. The goal is to target manipulative design without treating every supportive conversational feature as the same risk.
Make accountability operational
A requirement only matters if someone can verify it. Companies should document product changes, retain relevant evaluation data, and make it available for qualified independent review while protecting user privacy. Schools also need clear ownership: who evaluates the tool, who handles complaints, and who decides whether it stays in use.
The trade-off is real. More testing and oversight can slow deployment and raise costs, especially for smaller developers. But leaving it entirely to users and institutions to absorb uncertainty while companies capture the benefits is also a choice. Proportionate requirements can reduce the burden - lighter checks for low-risk tools, stronger evidence and monitoring where products affect children, education, or emotional well-being.
Key takeaways
- Baseline safety is necessary, not the whole story. Preventing acute harm doesn't guarantee healthy use over time.
- Evaluate repeated behavior, not just model outputs. Defaults, engagement incentives, and product framing shape real-world impact.
- Measure capability, not just activity. In education, ask what students can do independently after using AI.
- Scale oversight to risk. Tools aimed at young users or embedded in schools need stronger evidence and independent review.
- Treat uncertainty as a reason to monitor. Pilot, measure, adjust, and stop when the evidence doesn't support the intended benefit.
Conclusion
Preventing catastrophic AI failures matters. But if the policy agenda stops at the worst-case scenario, it leaves everyday product effects out of frame: what people practice, what they outsource, and which relationships or routines get displaced.
The standard should be broader and grounded in evidence. Companies should show how their products affect users over time. Schools and policymakers should demand stronger proof when AI is placed in high-impact settings. Safety is the floor - but whether a product helps people without weakening the capabilities and relationships they rely on is the harder question, and the one we should answer before adoption becomes routine.