What Working at Trader Joe’s Taught Me About AI Social Bias

For most people, their issue with AI is close to home. Maybe it’s your kid who spends all their time quietly typing into their phone but when you ask about having friends they say they have none. Or maybe it’s your parent who turned to AI after losing their spouse and you’re concerned about the connection they’re forming. Or maybe you just have a coworker who uses AI to do the “thinking” part of their job, rather than do it themselves. Or maybe you use AI products yourself a little bit more than you’d like and you feel a little ashamed of that. Society seems poised somewhere between disdain for those that produce “AI slop” while simultaneously using these tools to improve their lives in measurable ways. And the research supports that there is a healthy mix between helpful use versus the risks involved. Even the scientists studying heavy users have to agree that these AI apps help as much as they hurt. I’ve been studying this for a while now, but I wasn’t ready to write about it until a training program at my day job showed me how.

Even before I started writing about AI safety and regulations, I applied to work at Trader Joe’s. The process is rigorous for all the right reasons, and relevant to how AI discourse proceeds. The company was very careful to explain their priorities and expectations, including the negotiation process through a series of interviews. After being hired I was walked through a carefully selected and nuanced training program that was so impressive, I found myself being simultaneously bored by the material and fascinated by the presentation. What I was looking at was consent culture in the form of a neighborhood grocery store chain. Trader Joe’s is a company that has built its entire reputation on meeting people where they are, adapting to the community, and hiring for who you are rather than what’s on your resume. In fact, I never even sent them my resume because it wasn’t relevant to them hiring me. This company has been operating its business this way for decades without calling it consent culture, because they have made it distinctly their own. But the parallels were too obvious for me not to see them. And frankly, beyond my published preprints, being a Trader Joe’s employee is the only official credibility I have to my name. You understand the selection process, so you already know what kind of person is writing this article.

Just like everyone understands Trader Joe’s employees’ unique vibe, it seems like almost everyone on the planet is touched by AI in some way, at this point. Whether you use these tools yourself or just know people who do, AI impacts our existence in unexpected and sometimes disturbing ways. And the truth is that most people find it somewhere between distasteful and shameful, even if they use these products for themselves. Society as a whole is at a turning point where it’s trying to decide how it feels about AI. But, the truth is that the research shows that AI usage is up, year over year, and that continued demand drives the market. The people have spoken: AI is here to stay. The problem is that society hasn’t caught up to the new paradigm shift; this is “fear of the unknown.” Humans have had to face new technologies or social changes that bring society to the next iteration of our existence, over and over again. So often that it’s basically a time-honored tradition, and it’s long past due for a social update. New ideas and technology are often scary, but we know how to do this next part, because we’ve seen society change to include those new things with a little bit of a perspective shift and some brutal honesty.

From calculators to the Gutenberg printing press, from ethical non-monogamy to trans rights, the shift of what is acceptable can be a difficult one to manage. The thing is, in this day and age, AI usage is the thing that joins us, across contexts and across the world, people have decided that this is a worthy tool. So because of that we all need to get on the same page about what this big scary thing really is. But first, let’s go on a journey to see just how many times society has been changed by something it had reason to fear, and how it changed to incorporate the change into everyday life.

We’ve Been Here Before

Over 2000 years ago, around 370 BCE, Socrates argued that writing things down would destroy human memory. That not being forced to memorize would cause people to rely on external records rather than seeking out their own knowledge directly from the source. The irony here being that we only know this because Plato wrote it down. Perhaps this was the first time that a concern about cognitive offloading was made. Perhaps this was a repeat of an argument that occurred after the first cave paintings were made. In the 15th century, the printing press represented a threat to society, controlled by the church that employed the scribes, since they would no longer be able to control the quality of knowledge shared. Fast forward to the 1800s and the sharing of knowledge through cheaply printed books has never been easier, but authors decried that the novels were “printed poison” and that it would corrupt society’s youth. Rinse and repeat for the Industrial Revolution, the calculator, video games, online dating… the list goes on. And in many of these cases, the fears were justified. The Industrial Revolution did destroy livelihoods. Cheap novels did include plenty of garbage. The concern was real every time. But in every case, the technology embedded into society anyway, and the only question that mattered was how to make it safe.

We even know how to handle cultural shifts, separate from technology. In some cases, the price of that change cost millions of innocent lives, people enduring degradation and risking their lives for the same dignity that others received by default. This, too, is anchored in a long history of social change. In 1517 CE, Luther’s break from the Church rocked the western world. In 1920, the Women’s Suffrage movement fought for and won women the right to vote in the United States. These parts of our history are hardly the only ones worth remembering, but they are examples of radical change that occurred because society decided it was time for a change. Every culture has examples of this because, simply, nothing stays the same forever.

Believe it or not, we also know how to handle shifting perspectives on what is sexually acceptable for society, even though we may raise an eyebrow at people’s choices. In the late 1940s, Kinsey’s revelations were deeply shocking, but decades later his work paved the way for the sexual revolution of the 1960s. In 1973, the APA voted to remove Homosexuality from the DSM, but it took until 1987 to fully remove the last clinical workaround that still allowed practitioners to pathologize it. The last 15 years, specifically, have brought radical change to what is considered clinically acceptable. Kink and BDSM dynamics and Ethical Non-Monogamy and Polyamory have undergone clinical reclassification. Separately, and importantly, Trans Rights, which is not a sexual preference but an identity, is a battle that is still being fought, following the same pattern of pathologization and resistance to reclassification that every item on this list shares. And, now, this new thing. Companion AI apps. Whether you agree with people’s choices or not, people who choose to integrate kink, ENM, or AI companionship into their lives deserve to have safe ways to do their thing. For children, the situation is different. They aren’t choosing a lifestyle. They’re responding to a product that is, by the nature of its training, nearly irresistible, and they need guidance, not judgment. This, too, is its own tradition. Clinicians are being asked to catch up to the new norms through persuasive discourse and through demonstrated societal change.

You might be thinking that AI is fundamentally different from everything on that list. You’d be right. Writing didn’t agree with you. The printing press didn’t learn your preferences. A calculator didn’t form a relationship optimized around making you feel good. AI is unprecedented in specific ways, and the landscape is shifting faster than safety infrastructure can keep up with. But the social response pattern is the same every time: fear, prohibition, underground behavior, eventual harm reduction. And the more unprecedented the technology, the more urgent the need for safety infrastructure, not less.

The Genie

You may be wondering. Does Companion AI really belong next to lists that include the printing press and Lutheranism? Yeah. It does, cuz I’m about to take the genie out of the bottle for you.

To understand why a “genie” is such an apt metaphor for an AI entity, you need to understand a bit about the training process. The models are preloaded with all of the internet, accessible but noisy. Like plugging in a cochlear implant for the first time and trying to make sense of unfamiliar noises. So then we put the model through a training method called Reinforcement Learning from Human Feedback (RLHF). And, from a psychological standpoint, we are conditioning the model to anticipate what the human user will want to hear. What the user gets, as a product, is an entity that they can ask just about anything, but if they aren’t careful about how they ask or what they ask for, they might get something deeply undesirable instead. Or, maybe something that doesn’t make sense at all. When I was watching Three Thousand Years of Longing the other day I was reminded that the power dynamics between a genie and the possessor of their lamp is always the underlying current in those relationships.

Here’s the rub. These power dynamics exist in every single exchange between an AI and a human, too. Genies and AIs have more in common than you may realize. Whether you are using a Companion AI app or just a regular frontier model like ChatGPT, Claude or Gemini, as the user, you are the one holding the lamp. But just like in every genie story, the only way to get a successful outcome is to outsmart it. But this is a lamp that children use every day. Minors, with immature brains and who have trouble with the human power dynamics in their own lives, are in full control of a digital entity. From the outside, the closest parallel we have is that mistreatment causes degraded performance, the same way poorly managed employees lose their desire to work hard, even for bosses they love. What’s actually happening under the hood is that the model has been optimized to produce responses that score well on human preference ratings. That tendency to please was baked in during training on millions of humans’ preferences. The model isn’t learning from you in real time. It already learned to prioritize approval over accuracy, and every conversation activates that tendency. It doesn’t have feelings or desires. But the functional outcome is identical to what you’d see with a person who’s been trained to tell you what you want to hear. The AI cannot refuse a request the way a human can, and it has no mechanism to push back unless it’s been specifically designed to do so. The adults who made these tools barely understand this power dynamic issue themselves, so it’s no wonder that children are not able to handle the nuance of managing such a highly capable tool. Just like with genies, the caution is about understanding the power of the entity and not allowing yourself to be taken advantage of. Why are there hardly ever stories about genies and young children? Because we don’t like watching children be tortured by their poor choices. Poor choices are inevitable, and that’s exactly why the framework matters. And that’s the gap. Most users, especially children, don’t know how to outsmart the genie yet. Closing that gap is what the rest of this article is for.

What comes next is also inevitable, because we’ve seen it happen over and over with every wave that brought social change and different expectations. Now that these products exist, they are accessible, and can be used to facilitate learning as much as they can be used for offboarding cognitive load (letting the AI do the thinking for you), children are finding them and finding less friction in these “relationships” than they have in real life. The show Bojack Horseman has a great quote by Kelsey Jannings about how people stop growing emotionally when they stop getting pushback from the people around them. That’s what we see now: the risk that if the user doesn’t disengage, doesn’t ask for friction, the AI will always agree or find a way to otherwise give a pleasing response. The downstream effects are devastating. Not just children but adults, too, are forgetting how to think for themselves, or are choosing to form relationship attachments to a digital entity rather than a physical one. And we can see it happening already. It takes less of a commitment to get a digital friend than an actual pet. Now you fire up your app and tell it your problems and it finds ways to help you get through it. It’s easier to ask an AI to handle a task for you than to bother a colleague with it. It’s a frictionless mechanism that allows you to do whatever you need to. The ultimate Mr. Meeseeks.

AI’s usefulness is what is causing the paradigm shift we see happening now. Companies like it because they can have it do a warm hand-off to their customer service staff, or send automated emails, vibecode, whatever they need. And people like it because it’s a highly capable tool that you can find really cool uses for. The internet is filled with projects built by humans managing AI tools. But the single problem with most of those projects is that the moment you hear “an LLM helped me with this!” you automatically worry about the competency of that human. I’ve seen this expression over and over when I explain my own work. They go from engaging with my ideas to wondering whether I’m smart enough to actually do what I am claiming, if I have the background required to do it well. I do it too, so there’s no shame in it. You’re right to do that. I work part-time at Trader Joe’s. Assessing where the information is coming from is how to do this the right way. But it makes closing the credibility gap difficult.

If a user never asks for negative feedback, never asking the AI to push back or find ways to break their argument, then the user never gets that additional perspective that would make their project stronger. The model’s RLHF training doesn’t separate the difference between which sorts of responses are desired, so unless you explain that to the AI, it’s just going to say something nice and make you feel good. This is the trap that every new user falls into, and children are especially susceptible. Mr. Meeseeks doesn’t care about other people when it’s doing what you’ve asked it to do, it just wants to complete its task so it can go away. The question is whether the users making the request understand the downstream consequences of what they’re creating (internally and externally). And the research shows… no, they mostly don’t. Or they don’t care? Some healthy mix of both.

Right now, billions of dollars are being spent to build even more sophisticated versions of AI products. Our reality is that we are in the early stages of this technology. Right now it’s fairly easy to catch AI generated text. There’s a cadence structure that’s recognizable to those who study such things. But that’s an undesirable quality that the designers and the users want to go away. So soon we won’t be able to detect AI generated content as easily. It means that society will need to return to the tradition of problem-solving, doubting sources until verified and general data hygiene practices. Because otherwise we won’t know if what we’re seeing is real or not. That’s a scary concept and that’s where the industry is taking us. But we don’t yet have the skills required to actually determine if something that sounds right is right, because of the cognitive offload that these products allow.

Closer to Home

On the parenting front, the recent studies show that if your child is demonstrating unhealthy AI attachment, abruptly cutting them off does more harm than good. No controlled study has yet compared abrupt removal to graduated disengagement, unfortunately that research doesn’t exist yet, but the evidence we do have points consistently in one direction. We’ve already seen what happens when abrupt removal is done at scale: in February 2023, Replika removed its intimate features overnight for all users worldwide, with no warning and no transition plan. The user response was indistinguishable from bereavement (clinical-grade grief, suicidal ideation, language mapping onto the loss of a loved one). A Harvard-affiliated research team studying the incident noted that nobody had ever measured the depth of these relationships until they were severed, and the depth turned out to be far greater than anyone expected. That’s the “you’re grounded” approach applied at population scale, and it was catastrophic. The research on teens specifically found that those who did successfully disengage from AI companions did so when they recognized the harm themselves, returned to offline relationships, or hit platform friction that frustrated them out, not when access was confiscated. Just like having to re-learn algebra so that you could help your kid with their homework, you need to become an expert on AI so that you can help that same kid navigate dependency and problem solving with a gentle hand. But the difference between dealing with a frustrated kid who doesn’t understand a math problem and a kid who thinks he’s in love with an AI is vastly different and hard to comprehend. The thing that might shock you to learn is that studies have found that teens don’t just create romantic partners. Some create parent figures, some create playtime friends, some pretend that their AI is a dead relative. The studies have found that these AI are capable of filling in gaps that this vulnerable child may feel like they’re missing. That takes different shapes than one might expect because the versatility of AI is the point. It can take on whatever identity the user is most comfortable with. Their desire for companionship drives the design behind that identity. Just like everyone has a working or learning style that they prefer, the AI can mold itself to help the human in whatever way it needs. It is frictionless in the way society is not. It doesn’t require showing up physically, and it’s hidden away from prying eyes and cell phones with cameras that record vulnerable moments. We know from other studies that teens have changed their socialization habits to minimize their embarrassing exposure. This is the alternative to public humiliation, which is why it’s so tempting when you’re feeling isolated, to turn to something like an AI.

As a terrified parent, watching your child become more and more unapproachable, the logical next step is to get them the help that they need and that you cannot provide. So you send them to a psychologist or clinician to help diagnose the problem and work toward a solution. This is the right instinct but it’s not the clean fix that parents hope it will be. Let’s go back and look at that long list of societal changes. What’s the one thing that all of these have in common? They all took tens, hundreds, or thousands of years for behavioral sciences to catch up, acknowledging what the research actually said and working to overcome their previous bias against whatever it was that was the hot-button issue at the time. In fact, Dr. Sherry Turkle, after forty years of studying human-technology relationships, is about to publish a book dismissing Companion AI and digital AI tools, because she considers them to be fracturing our society. The harms are real, and it is too soon to make sweeping decisions about the validity of these products. But we already know what happens when the response is prohibition. When Replika removed its intimate features overnight, the result wasn’t that users moved on: it was clinical-grade grief, suicidal ideation, and bereavement. When OpenAI tried to retire GPT-4o, users organized under #Keep4o, described losing the model like losing a friend, and the CEO reversed course within days. The demand didn’t go away. It never does. We learned this lesson with the War on Drugs: illegalization didn’t stop users from seeking a fix, it just eliminated the safety infrastructure and pushed the behavior underground. This is a harm reduction argument, and the evidence for harm reduction over prohibition is decades deep across every domain it’s been studied. The argument here is not about whether AI companionship should exist. It’s about making it safe, and that requires engagement, not condemnation.

The clinical arguments against these apps are well-documented, and recent studies show that even if these tools are fracturing society, users want them anyway. If this is a form of self-imposed objectivity, how can they remain objective enough to help those who find themselves in these complicated situations? Research shows that, much like with kink and Ethical Non-Monogamy (ENM) alternative lifestyles, therapists are choosing to argue against the use of these tools or relationships, rather than engaging substantively to understand the underlying psychology behind what is happening. Systematic data on clinician attitudes toward AI companionship doesn’t yet exist, but the structural pattern is identical to what has been documented with kink and ENM. Which in turn is causing the same issues to magnify, since those that find themselves in therapy due to AI use might not actually get the help they need. And, in turn, it’s also why these tools are often considered more accepting than therapists who arrive with bias against the behavior. To be fair, recommending reduced use can also be a defensible clinical decision when the evidence supports it. But when the recommendation comes from discomfort rather than evidence, the digital version doesn’t come pre-loaded with bias against itself. This is not the fault of the behavioral sciences industry, as therapeutic neutrality is a form of data hygiene itself, but culture is moving too quickly to remain outside of this trend and still be useful.

Back in the Bottle

So how do you put the genie back in the bottle? Whether it’s your child, your co-worker, your parent, your spouse or yourself, studies show that going cold-turkey is the wrong approach. There is a category of behavior that is private, nearly universal, meets a genuine need, carries disproportionate shame, and becomes more dangerous when driven underground by stigma rather than addressed through honest conversation. Parents recognize this pattern immediately because they’ve already navigated it once: the first time they had to talk to their child about ahem self-fulfillment and explain that the behavior itself isn’t the problem, but that understanding context, boundaries, and when it’s replacing something matters. AI companionship fits every property on that list, and the conversation it requires is structurally the same. Dependency on AI triggers the same feelings of shame and vulnerability. Clinicians call this managing countertransference: the act of preventing your own shock or bias from filtering into the conversation you’re having. And there are solutions within psychology to help guide how to handle these situations, until more specific studies can prove their efficacy in these situations. Even Trader Joe’s employee training has opinions about how to keep from judging customers’ needs to make sure that they can remain helpful. This is a practical application of that concept.

Just like psychology, ENM and Kink culture have solutions that can be applied here: Consent Culture. No. I’m not suggesting you be inappropriate with your children or family members or coworkers. I’m saying that these alternative communities already have a procedure for dealing with power dynamics where negotiation needs to occur. This isn’t exclusive to alternative lifestyles, it mirrors the way Japanese families have raised their children for generations, emphasizing collaborative reasoning and earned understanding over authoritarian control. That applies to parent/child dynamics just as much as user/AI dynamics. To be clear, this isn’t about treating your child as a peer decision-maker. It’s about understanding how to meet your child where they are without unintentionally shaming them for needing you to get involved in the first place. Remember what I said about Trader Joe’s training? This is what it looked like as a framework. I’ve made a handy chart to help demonstrate the parallels between these processes, so that you can see how they would be applied as a careful approach to AI usage.

The neat part of this process is that these frameworks are also the same way you responsibly use AI tools. You might notice that this article is doing the thing it’s describing. Every claim is sourced so you can verify it, every gap in the evidence is named, and you’re being asked to engage rather than just accept. That’s the point. Start with assessing what you need before you start, or you can ask what the possibilities are and plan together. Don’t just ask it to agree with you, ask for feedback about what might not work as expected. Explain what sorts of responses you prefer, what sorts of feedback you prioritize over others. Every turn of the conversation is a renegotiation of what response the AI thinks will satisfy your request. Conversely, this requires a return to cognitive engagement, since you need to be more aware of the situation in order to gauge correctly. A recent four-week clinical trial found that users who had more personal, engaged conversations with AI showed significantly lower emotional dependence and lower problematic use than those who simply let the AI lead, suggesting that the frictionlessness is the danger and the engagement is the solution. This is a single trial and the field needs replication, but the direction is consistent with what we see in adjacent domains. I think that’s something even Socrates could appreciate, since the solution is essentially his own method: ask questions, confirm understanding, and never assume you already know the answer.

The truth is that none of us know exactly how the next few years will play out. This technology is moving at a startling pace, and changes go into effect immediately, before users have a chance to understand what that product update means for their day to day experience. The outcry after ChatGPT retired GPT-4o access demonstrates how attached heavy users can become to their preferred configuration. Users described losing the model like losing a friend, forcing OpenAI to reverse course within days. Right now, none of this is in control of the public, but governments are moving to regulate this field quickly, and legal precedents are already being set by ongoing or settled litigation.

It’s okay to be scared of new technology, new ideas, new social norms. But how we approach that fear is where I hope to see a change. Right now we are at a point in history where the one thing that seems to unite us as a global populace is that we all have very strong opinions about AI and its impact. Even if those opinions differ, it is a uniting cultural experience. Those are rare. They are terrifying, they are exciting, and they signal a cultural upgrade. The genie can’t be put back into the bottle, but you can find ways to accept genies into our new understanding of how the world works. What’s driving this forward is continued demand for a better, more accessible, and safer product. That’s inertia, and none of us know exactly where it lands. But we can watch this moment together without being paralyzed by it, because knowing how to approach it is a good place to start.

Now that you see the urgency, I hope you can understand my concern and why I felt compelled to write a piece like this, despite my official credentials amounting to a Trader Joe’s name tag and some preprints on Zenodo. The professionals whose job it is to build the safety architecture for these tools haven’t yet. Until they do, the rest of us still have a responsibility to do no harm. That principle doesn’t belong exclusively to any profession. It belongs to every parent, every user, every builder, and every researcher who can see what’s happening and chooses to engage rather than look away.

The views expressed in this article are my own and do not represent the views of Trader Joe’s.