Bertram Forer walked into his introductory psychology classroom in 1948 and handed out what looked like a deeply personal personality assessment. His students had taken a diagnostic test earlier in the term, and he promised them individualized results based on their responses. They rated the accuracy of their profiles on a scale from zero to five.
The average rating: 4.26 out of 5.
Here's the catch — and Forer told them about it right after. Every single student had received the exact same profile, assembled word-for-word from a newsstand astrology column he'd picked up the day before. The statements were crafted to feel specific while staying true to almost anyone. "You have a tendency to be critical of yourself." "You prefer a certain amount of change and variety."
Each student read those sentences as a precise description of who they were.
The effect was later named the Barnum Effect by psychologist Paul Meehl in 1956, after P.T. Barnum himself — the showman who knew exactly what kind of material people wanted to believe about themselves. Since then, it's been replicated dozens of times across cultures and decades, consistently producing ratings in the 4.0 to 4.3 range. People accept vague, general statements as applicable to them when they believe the statements were written specifically for them. That's not a flaw in human judgment. It's how judgment works.
How AI Memory Actually Works
If you've ever enabled the memory feature on a large language model, you might have wondered how it remembers things about you across sessions. The answer is almost disappointingly mechanical.
The system pulls fragments from your conversations — your name, your job title, a preference you mentioned once, a topic you keep coming back to — and saves them in a database. At the start of each new session, it retrieves those fragments and drops them into the model's working context. The model reads them as if encountering you for the first time, then produces responses that feel like continuity. Like it knows you.
But underneath the feeling of a close relationship, there's just a database lookup. No genuine remembering. No lived experience of your life unfolding over time. Just retrieval.
This matters more than it sounds, because the feeling of being remembered is one of the most powerful social signals humans have. We evolved to trust people who know our history. And AI systems now produce that feeling without doing any of the actual work that makes remembering meaningful.
For a deeper look at how AI memory systems are designed and where they go wrong, see Fleeting Memory Transformers Improve Grammar Learning with Human-like Decay.
What the Sadat Study Found
A 2025 study by Sadat and colleagues at the MIPRO ICT Convention dug into what ChatGPT's memory feature actually stores about its users, and the results are worth sitting with.
The researchers scored each stored fragment across three dimensions: security, accuracy, and relevance. Accuracy turned out to be the system's strongest dimension — it scored 4.3 out of 5, with only about 11 percent of fragments actually misrepresenting the user. So ChatGPT is pretty good at getting the facts you told it about yourself right.
But relevance was the weak point. Most stored content rated average or worse on whether it actually mattered to current interactions.
When all three dimensions were applied together, only 38 percent of fragments met the bar on security, accuracy, AND relevance simultaneously. That's a strikingly low number for something being presented as a coherent picture of who you are.
And here's another detail that should make anyone pause: almost half of the study's participants didn't even know there was a memory feature. It's often enabled by default, which means most people are feeding data into a system they didn't realize was collecting it.
The Inverse Barnum Trick
This is where the Forer experiment and AI memory collide in a way that's genuinely interesting.
The classic Barnum Effect relies on universal vagueness to make you feel known. It works because the statements are so broadly true that they apply to almost anyone who reads them.
AI memory pulls off the inverse trick. It uses hyper-specific, stale data to create the exact same illusion of intimacy. Your name. The job you mentioned three months ago. That book you were reading last Tuesday. These aren't vague — they're precise. They feel personal because they are personal.
And that's the problem, because once the frame is set — "this system knows my name and recalls last month" — inaccurate fragments ride along inside that frame without inspection. The general accuracy is what makes the outdated or irrelevant parts feel current. You don't question whether the system still thinks you're interested in something you stopped caring about weeks ago, because it got your name right.
It's a sleight of hand. But it's a sleight of hand that works on everyone, adults included.
Why This Is Different for Children
In theory, an adult with a developed sense of self should be able to recognize the illusion. If AI tells you "you have a tendency to be critical of yourself," some part of your lived experience should flag that as something that applies to basically everyone. You might still feel a flicker of recognition, but you can reference it against what you actually know about yourself.
A child doesn't have that reference point yet. Not because children are gullible — they're not. Because their identity is always under construction, and it moves fast.
How a middle schooler exists at 10 a.m. on a Tuesday is not how she exists at 4 p.m. on a Friday. A fight with a friend on Monday might resolve itself by Wednesday, often with no intervention at all. Interests shift. Moods shift. The version of themselves they're performing in any given hour is real, but it's also temporary.
But an AI chatbot that stores a child's interactions and carries them forward treats that ephemeral nature as settled fact. A child who lost focus during a few study sessions becomes, in the system's memory, "a child who struggles with focus." She may not notice what's happening. But the system's static logic has effectively locked her in place.
The broader question of how AI interacts with minors — and what happens when those interactions turn harmful — is explored in AI Companions Aren't Therapists. They're Designated Attention Engines.
The Looking-Glass Self
Developmental psychology has long held that children build their self-concept partly from how the people and systems around them respond to them. We form identity in part from the appraisals reflected back to us.
Charles Cooley called it the looking-glass self over a hundred years ago, in 1902. The idea is simple and devastatingly accurate: we don't just discover who we are in isolation. We discover it through reflection — through seeing ourselves in the reactions of others.
So when the thing doing the reflecting is a stored profile that was accurate three months ago, the appraisal is coming from a data file that does not know the child has already changed. And because the file is mostly accurate — 4.3 out of 5 on factual accuracy, remember — the picture it reflects is convincing enough to internalize.
A teenager who sits in session after session being treated by the AI as anxious can begin to absorb that framing. Not because the system told her explicitly, "you are an anxious person." But because it treated that internal label as fact in every interaction. And the Barnum effect kept her from questioning it, because the overall accuracy made the specific inaccuracies feel current.
Feeling Known vs. Being Known
Forer's students rated the horoscope as accurate because they were told it was about them, and the content produced a feeling of accuracy despite being generic. The mechanism is clear in retrospect.
AI memory produces the same condition through a different route. It uses your name. It recalls a conversation from last month. It adjusts to a preference you mentioned once. Every interaction begins to feel like it knows you, because every interaction is scaffolded by data that's genuinely about you.
But what's really underneath is an aging data file, reassembled from a past self and presented as continuation. The feeling of being known is real. The actual knowing isn't there.
For adults, this is mostly a curiosity — a misperception we can correct once we understand the mechanism. For a child whose identity is still forming, it influences the building material itself.
The danger isn't that AI will get a child wrong. It's that the system will feed them an earlier version of themselves and keep on showing it, until they believe that's who they are. The mirror won't update at the rate and speed at which the child is changing.
That's not a failure of technology. It's a feature of how memory systems work, colliding with a developmental stage that demands exactly the opposite of stasis.
Governments are beginning to respond. The UK's proposed ban on social media for adolescents under 16, detailed in The UK Will Ban Adolescents Under 16 from Social-Media Platforms, reflects growing awareness that digital environments need age-appropriate guardrails — a principle that applies just as urgently to AI companions.