Can AI Help Us Become Less Dependent on It?
Last week, we looked at students whose homework improved while their performance on exams without AI got worse. That leaves parents and teachers with a pretty difficult problem. The help is available, the student likes using it, and the finished work looks good to everyone involved. Asking people to go back to doing things the old way is going to be a hard sell, especially when there are other things AI genuinely helps them understand.
What I want to know is how much we can change what happens while someone uses it. Is there any world in which the help AI gives can leave them better prepared to do the next problem themselves? And if someone turns to AI because they are upset, could that conversation help them figure out what to do when the same feeling comes back, without having to use AI again?
This week's research offers some early answers. We have an experiment that changed how people used an AI math assistant, another that changed the kind of emotional support a chatbot offered, and a study of why a friendly answer can be easier to believe even when it is wrong. We also have a look at what parents and teenagers say about AI companions, which suggests they may be worrying about different parts of the same conversation.
01 / THE RESEARCH
Showing people which part they let AI do
New preprint • September 17 • Preregistered randomized experiment
|
704
adults practiced fraction arithmetic
Feedback about their own use of AI showed promise for reducing complete-answer requests and improving a later test without assistance.
|
Sebastian Maier and colleagues gave adults fraction problems to practice, followed by a test where they could not use AI. Some had an assistant that would provide a complete answer if they asked for it. Others also received short feedback describing how they had used the assistant and which parts of the method they had worked through themselves.
That feedback is the part that I personally find interesting. Getting the answer right would normally let someone move on to the next thing. Here, they also got to see how much of the work behind it they had actually done before starting the next problem.
The feedback reduced requests for complete answers and improved later test performance under the researchers' planned statistical tests. Rewards for using less assistance did not produce those benefits. But the estimates were uncertain, and this was one session with adults, not a semester with children who have less overall knowledge. The experiment also did not establish that access to AI, by itself, made people perform worse than having no AI at all.
I would be careful about calling this a solution to what I put in last week’s Reckonings. Still, as a former teacher, I like that it gives the learner something more useful than another warning about cheating. If I read a worked example and every step makes sense, I can easily believe I would have known what to do. Having to choose that first step myself is where I might discover that I didn't really have a full understanding.
For a parent helping with homework, that is a useful place to pay attention. A child may honestly feel that they understand the explanation. Asking them to try a similar problem gives both the child and the parent a better idea of where they still need help. And the AI could be designed to make that part of the process, instead of leaving the parent to add it after the assignment already looks finished.
Read the preprint →
02 / THE RESEARCH
Feeling better and being able to handle what comes next
Peer-reviewed experiment • Published September 8 • Included in this week's reading
|
380
adults received one of two kinds of AI support
Both groups felt better immediately. The differences appeared in a later task without AI and in what participants said they would need next.
|
In a study by Chun Wang, adults in Guangzhou described an upsetting everyday experience and received one of two styles of chatbot response. One helped them name what they were feeling, consider another interpretation, and choose a coping step. The other emphasized reassurance and the AI's continued availability, whenever they needed to call on it.
Both groups reported immediate relief, with no clear difference between them. But in a later task without AI, the first group wrote better coping strategies and reported greater emotional recovery. The reassurance group expressed a stronger expectation that they would need the AI again in the future.
Before we turn that into a claim about addiction, the author is clear about what was measured. These were short-term responses in a single session. The study did not establish emotional dependence, and the difference in whether people actually requested another AI response was inconclusive in this study.
Even with that limit, I think this is a useful question to bring into the conversation about companions. If someone comes away feeling calmer, we can all acknowledge that has value. I don't want to dismiss the relief just because it came from a machine, because it's real all the same. But I also want to know what the person thinks helped them, and what they believe they can do the next time they feel that way.
Suppose someone is upset after an argument with a friend. They may need reassurance before they are ready to think about the conversation without bias. If the help eventually gets them to consider the friend's point of view and decide how to talk to them, there is a chance for the relationship to improve. If every disagreement becomes something they only discuss with the chatbot, the friend never gets that chance to repair that relationship, and even make it stronger. That is a possible consequence I want researchers to follow, beyond whether the AI user felt better at the end of the chat. Because what happens after that probably matters more to both parties.
Read the study →
03 / THE RESEARCH
The answer can feel personal and still be wrong
Peer-reviewed experiment • Earlier paper receiving new September coverage
|
477
participants tested a simulated AI health assistant
Conversational delivery increased the credibility of false information. Active verification helped people become more skeptical.
|
Mengqi Liao and S. Shyam Sundar tested how people responded to an AI health assistant that supplied false information. Some saw a more conversational version, with responses tailored to what they had said, greetings, and even the appearance of someone typing. Others received a less conversational version of the AI assistant.
The more conversational delivery made the information more credible to participants. The researchers also tested verification features. A visible cue alone was not enough. Following a verification link to information that corrected the false claim helped reduce misplaced trust. When checking was optional, nearly half chose to do it.
I can understand why an answer that refers to the details we just provided would feel more trustworthy. It looks as though the system has taken our circumstances into account. If it remembers what we said and explains itself in language we find easy to follow, checking can start to feel unnecessary.
Imagine using AI to make sense of a claim, then telling a friend that you looked into it. You may genuinely believe you did. The friend now has a reason to trust it because it is coming from you, and neither of you has actually dug in to the original source. As these conversations become more natural, I think we have to be more deliberate about that step, especially when someone else will be making a decision based on what we tell them.
Read the study →
Read Penn State's explanation →
04 / THE RESEARCH
Parents may be missing why the child wants the chatbot
New preprint • September 17 • Analysis of Reddit posts
|
1,489
Teenager-community posts
|
139
Parenting-community posts
|
These were separate online accounts, not interviews with parents and their own children.
|
Thomas Berkane and colleagues compared discussions of young people's AI companion use in parenting and teenager communities on Reddit. The topics that came up most often differed. Parenting posts focused more on romantic and sexual interactions. Posts in teenager communities more often described getting emotional support from the bot.
The teenagers' accounts were not simply enthusiastic, either. Attachment and dependence were prominent concerns there too. The picture was more complicated than adults seeing the danger and children seeing only the benefit of using AI for support.
There are some limits to this analysis that we should consider. First, the researchers could not verify everyone's age. Also, the groups were very unequal in size, and parents posting for help may be dealing with something unusually concerning. This is not a representative survey of families, and it cannot tell us what any particular parent knows about their child.
What it does make me wonder is how often the conversation at home starts with a question the child doesn't think explains why they want to use AI, or any technology for that matter. A parent might want to know whether there is sexual content or how many hours the app is taking up. Those are reasonable things to ask. But if the child is going there because they feel left out at school, answering those questions may never get either of them to what is actually happening at the core of their child's AI use.
I would want to ask what they are getting from the conversation that they are having trouble getting elsewhere, namely from other humans. That doesn't mean accepting everything the chatbot says or abandoning a boundary that needs to be there. It means knowing what else needs attention. If the child is lonely, changing an app setting still leaves a lonely child, and the people around them need to know enough to help.
Read the preprint →
|
THE RECKONING
I don't think most people want to become dependent on AI. They want help with something, and there is usually a reasonable explanation for why they went looking for it. We have spent a lot of these first ten issues talking about what can happen after that becomes a habit. This week's papers give us more to consider about how that habit might develop differently.
The math experiment, in particular, makes me think about how much responsibility we are putting on the person using the AI tool. We give someone a system that can finish the work, then ask them to decide which parts they should still work through themselves. For a child who wants to get through homework, that is asking quite a lot, if we are being honest. For an adult who has work due, it may not feel like much of a choice either. Time is precious and many of us have more to do than ever.
I would like more of the companies building these products to take that problem seriously. If an assistant is sold as something that helps people learn, whether they can do the next problem without it should matter. If it is meant to help someone through a difficult conversation, I want to know whether the person eventually has that conversation, and how it goes for both people.
At the same time, I don't want being independent to become another way of saying we should need nobody. Asking a friend for advice can be part of how we stay close. A teacher explaining something again is doing their job, and a child should feel comfortable asking, not waiting to ask an AI. The concern I have is that we make the machine so easy to turn to that the other forms of help gradually get less room in our lives.
There is some encouragement in this week's research. How a tool responds appears to make a difference in at least some of these settings. I want to see those experiments followed for longer, including what happens to the people who would otherwise have been asked to help. In the meantime, there are parts of that process we can keep doing ourselves, including deciding when someone we know should be part of it.
|
Also this week
A few more from this week's reading, with the source linked if you want to look further.
An AI tutor improved science test results
A four-week randomized trial in English secondary schools found better test results with the Medly AI tutor, although only 644 of the 929 students assessed at the start completed the final assessment, a substantial limitation in this preliminary preprint.
Read the preprint →
AI mediation helped in a simulated dispute
James Hale and Jonathan Gratch found that AI mediators reduced negative emotion more effectively than novice human mediators in a buyer-seller role-play, but the buyer was always an AI, so the preprint does not establish how well the system would mediate an argument between two people.
Read the preprint →
AI peers influenced dishonest behavior
Two randomized experiments involving 640 people found that dishonest cues from AI group members encouraged dishonest reporting, with influence comparable to human peers in these tasks, not a measure of how often people would cheat in everyday life.
Read the preprint →
The researchers used AI even when the rules said not to
Among survey respondents assigned to an AI ban in an ICML peer-review experiment, 22.5 percent reported using it anyway, while the study found almost no effect of policy assignment on paper scores or final decisions.
Read the preprint →
Work can still feel like yours when AI helps
An exploratory survey by Megan Wei and colleagues found that people described more ownership when they directed or revised AI-assisted work than when they merely approved its suggestions, including projects they said they could not have completed without AI.
Read the preprint →
Florida announced new classroom AI restrictions
Florida's education department announced adopted rules restricting companion AI in public education and requiring greater transparency about classroom tools, a policy development whose effects on students still need to be evaluated.
Read the department's announcement →
|
PRACTICE
Check an answer that sounds convincing
This week, try Signal at The Quiet Cost Practice app. It gives you a claim and asks you to inspect the evidence behind it. Sometimes the claim holds up. Sometimes it doesn't, and sometimes you need more information before you can decide. You get an explanation after each answer.
Have a go and see which details you notice. Then, the next time AI gives you a summary you plan to use, choose one important claim and open the source. Find the part that supports it. If you can't, that is worth knowing before you pass it along.
You can also pick one of the daily games if you have a few minutes and want to play. I am building these as places to keep using skills that AI makes easier to skip. We have not established that doing better in a game prevents skill loss elsewhere in life, but we can give ourselves more opportunities to practice.
Try Signal or play today's games →
|
|
YOUR TURN
Think about the last time AI helped you with something you didn't know how to do. If a similar situation came up tomorrow, what could you do yourself that you couldn't do before? What would you still need help with?
Take five minutes to write about the actual situation before asking AI to help with your answer. If another person used to help you with this, include what has changed between the two of you.
|
Pass it on
If someone came to mind while you were reading, forward this to them. Reckonings is free and arrives each week. They can subscribe and read the earlier issues at quietcostweekly.com. My longer essays are at At a Cost on Substack.
Subscribe and read earlier issues →
At a Cost →
Michael McNamara
quietcostweekly.com →
Illustrations created with AI. They are editorial images, not depictions of research participants.