What 23 weeks of asking AI about my unborn child taught me

11 min read

  • AI
  • parenting
  • learning
  • dependence

In May 2026, my wife and I saw our baby for the first time through an ultrasound scan. Since then, I’ve been back and forth with several AIs asking tons of questions ranging from the latest research in childhood development to “how to take care of my wife as a first-time expectant father?” Every answer came back more personalised than the last, and I’m satisfied with the answer almost every time. I only use search engines for verification now. Why go anywhere else when I’ve built up so much memory with AI already?

That volume is what worries me: I’ve now become dependent on AI. I’m doubly worried my kid will be the same; terrified of the image of my kid becoming an Axiom Human from Wall-E, fully dependent on AI to do everything for them (purely coincidental that their main interface is a screen capable of voice and video recognition).

An illustration of an Axiom human from Wall-E, reclining in front of a floating blue screen.
Prompt

Editorial ink-wash illustration on warm off-white paper (#FAF8F5). A single adult reclines in a simple rounded chair that hovers a few centimetres above a pale seamless floor. Their posture is slack and content: shoulders rounded, head tilted toward a small floating rectangle of pale light, one hand hanging open over the armrest, feet clear of the ground. The limbs look unused rather than weak. Background reduced to a faint horizon line and a soft floor shadow; no furniture, no doors, no other people. Drawn with confident thin black ink lines (#2F2823) and two soft grey washes. Generous negative space, hand-drawn line quality, slight paper grain. Monochrome only — no colour anywhere except one small area of deep cobalt blue (#18407C) in the floating screen. The screen is a soft blank field of light: no interface, no icons, no words. No text, no lettering, no captions, no logos, no signage. Composition 3:2, subject centred with air around them.

The industry is using humans to make AI better

The AI industry is predominantly focused on achieving two things in the short term:

  • Automating knowledge work

  • Augmenting AI with humans

When this happens, these consequences follow:

  • Removal of humans in the loop (HITL)

  • Keeping HITL to supervise and train AI

We’re seeing the former through the slowdown in hiring for early-career roles. Gartner and the Federal Reserve Bank of Dallas have independently confirmed this.1,2

For the latter, entirely new industries are being created to support this. A famous example is Scale AI, which provides data annotation services to companies training large-language models (LLMs), having a self-reported revenue just shy of US$1B in 2025.3

A small person reaches up to place one card into a slot in a machine much taller than them, an empty tray on the floor behind them.
Prompt

Editorial ink-wash illustration on warm off-white paper (#FAF8F5). A single small adult stands at the left edge of the frame, reaching up to press one small square card into a narrow horizontal slot near the top of a large plain machine that fills the right half and rises well above their head. Their posture is unhurried and exact, like a clerk at a counter: one arm raised to the slot, shoulders level. The adult's face is neutral and attentive: mouth level, eyes on the slot, no smile, no frown. A shallow compartmented tray rests on the floor behind the adult, every compartment empty. The machine is taller than the person; the disproportion is entirely in size, not in threat. Background reduced to a faint horizon line and a soft floor shadow; no furniture, no cables, no other people. Drawn with confident thin black ink lines (#2F2823) and two soft grey washes. Generous negative space, hand-drawn line quality, slight paper grain. Monochrome only — no colour anywhere except one small square of deep cobalt blue (#18407C) in the card held between their fingers at the slot. The machine's front is a single blank plane: no interface, no icons, no words, no panels, no lights, no face. No text, no lettering, no captions, no logos, no signage. Composition 3:2, figure small at the left and machine mass at the right, with air above them.

This means one thing: AI isn’t just another tool anymore. It instead aims to be the de-facto method to deal with everything. I’m not opposed to automation. AI has made my job easier because I don’t have to manually write code all day anymore. But it’s made my job more difficult too, because I’m losing the byproduct of that manual labour - understanding and judgment. This makes AI mandatory for our day-to-day.

And it’s that dependence that worries me, especially for the generations to come. What would my child be like if they’re so dependent on AI for their capabilities? Would part of raising a child just mean getting the best AI to augment their daily life?

I believe in the alternative: we should be using AI to raise the floor of human cognition. We should use it to boost our own intelligence, and not the other way around.

What the research actually shows

Recent research is encouraging, like the PNAS field experiment that was done on around 1,000 high-school students in Turkey. They found that there was a 127% performance increase in practice sessions when using AI with carefully designed guardrails compared with a 48% performance increase when using AI without guardrails. The same experiment found that using AI without guardrails during practice led to a 17% decrease on unassisted exams versus not using AI at all.4

For the students’ exams, they scored similarly to those not using AI at all. But students perceived that they performed significantly better, despite the mismatch between perceived and actual learning. My hypothesis is that a positive perception is enough for them to put greater effort into their learning, which would eventually lead to actual learning gains.

More encouraging is the NBER working paper, showing that supervised ed-tech support produces large learning gains in math, increasing by half a standard deviation over 31 weeks.5 By my calculation, this is equivalent to students effectively compressing more than a year’s worth of traditional middle or high school math instruction in less than a single academic year. The key part of this result was the supervision. The humans in the loop (the teachers) and the organisational structures were necessary to ensure that the technology was fully utilised.

A student sits upright at a desk looking at a small blue screen while an adult stands to their right, one hand resting on the back of the chair and the other hanging at their side.
Prompt

Editorial ink-wash illustration on warm off-white paper (#FAF8F5). A student of about sixteen sits upright at a plain desk on four straight legs, feet flat on the floor and visible beneath it, facing a small monitor that rests on the desk top on a short plain stand, raised to eye level. To the right of the chair, one adult stands turned three-quarters toward the desk: one hand flat on the top edge of the chair back, the other hanging open and unused at their side. Both arms read as separate lines against the paper. The adult touches nothing but the chair. Both heads are level and both look at the monitor; shoulders easy. The mood is ordinary — two people at work in the middle of a good day, nothing here needs fixing. Background reduced to a faint horizon line and one flat pale grey band on the floor; no other furniture, no other people. Drawn with confident thin black ink lines (#2F2823) and two soft grey washes; generous negative space, slight paper grain. Monochrome only — no colour anywhere except one flat rectangle of deep cobalt blue (#18407C) filling the monitor's face, with no glow and no sheen. The monitor's face is that one flat blue rectangle: no interface, no icons, no words. No text, no lettering, no captions, no logos. Composition 3:2, exactly two people, student left of centre and adult to the right, both fully in frame.

What I’m doing before my child is born

The PNAS study shows that guardrails prevent harm.4 And the NBER research shows us that it’s possible to use AI in a way that improves our intelligence, and doesn’t create dependence on it.5 That way is through active human participation. I’m already exploring the possibilities with this, especially when it comes to raising my child. I’m already starting to form parts of this plan with my wife.

The first is simple: no screen time under 2 years old. This is in line with the Australian Government’s 24-Hour Movement Guidelines,6 and I hope this will help my child have a healthy relationship when they eventually interact with AI and technology in general.

Second: personalised software (I’ve got to utilise my skills for my kid). I’ve been experimenting with some designs, but the one thing for sure is that I’d want it to adapt as my kid grows. For instance, while my child is under 2 years old, my wife and I use the software 100% of the time, applying the no-screen-time rule. It’d ideally be local-first for maximum privacy, only sending out anonymised data to LLMs for analysis. It should be guiding me whenever I need help with parenting, and adapting to my family’s situation. It does this through generated daily to-dos (after human-AI planning sessions) and a milestone log so the AI has context on developmental standards (it’d be cool to look back years later on how they’ve grown too).

An adult sits on the floor holding a baby, both arms around it, while a plain slab lies face-down beside their knee with a thin band of blue light along its edge.
Prompt

Editorial ink-wash illustration on warm off-white paper (#FAF8F5). Seen from the front at floor level, slightly to one side. A single adult sits on a plain pale floor with their back against a bare wall, holding a baby against their chest. The baby is small and mostly hidden by the adult's arms: one arm across the baby's back with the hand flat and still, the other arm supporting underneath. Both arms read as separate lines against the paper. The adult's head is tilted down toward the baby, not toward the reader; the face is calm and absorbed, mouth level, no smile, no frown. The adult and the baby rest fully on the floor. Beside the adult's knee, within reach and not reached for, a plain rectangular slab lies face-down and flat on the floor: its screen is against the floor, so only its plain back and its edge are visible. Background reduced to one faint horizontal line where the wall meets the floor and one flat pale grey shadow under the adult; no furniture, no toys, no stand, no cable, no other people. Drawn with confident thin black ink lines (#2F2823) and two soft grey washes. Generous negative space, slight paper grain. Monochrome only — no colour anywhere except one short flat bar of deep cobalt blue (#18407C) lying on the floor where light escapes from under the slab's edge, hard-edged, no glow and no falloff. The slab's visible surface is a plain flat back: no interface, no icons, no words, no buttons. No text, no lettering, no captions, no logos, no signage. Composition 3:2, adult seated slightly left of centre, slab low at the right, air above them.

As my child grows past 2, the software should then morph into a tool that’s developmentally appropriate for my child’s stage of life, while still keeping us parents engaged throughout. I imagine this to be similar to the popular AI learning apps on the market, but it should fade out of their life eventually, as my child shouldn’t form a dependence on it.

4 questions that tell you whether AI helped

Knowing whether this has been achieved or not boils down to 4 questions (CART):

  • Capability: can they do the task without AI?

  • Agency and character: do they choose their work, regulate their attention, and act with integrity?

  • Retention: can they still do it 30/90 days later?

  • Transfer: can they do a similar task they have not seen?

Once AI can consistently support humans in doing these, as well as the physical basics that support these (e.g. sleep, exercise, nutrition), only then do I consider AI to have raised the floor of human cognition at a level that isn’t detrimental to us. Until then, we can’t hope to maximise our human potential, as a number of the major AI providers are labelled with “Unacceptable Risk” by independent reviewer Common Sense Media.7

So in an increasingly AI world, how do we leave the next generation more capable, with more agency and strong character? I don’t have all the answers. All I know is that my hopes are the same: that my child grows into a person who can perceive, reason, act, and keep becoming. I hope my child won’t be the person who uses AI the most, but rather the person who needs it the least. They can do this because they’ve honed the skills and knowledge required to thrive, having used AI responsibly and healthily in their formative years. And since they’ve formed that healthy relationship with AI, they can then use it to go further than they could alone.

Bibliography

  1. Gartner. “Gartner Survey Finds AI Automation Is Reducing Some Entry Level Hiring at Nearly One Quarter of Organizations.” Press release, July 27, 2026. ↩
  2. Federal Reserve Bank of Dallas. “Job postings show early signs of AI automation impact.” September 2026. ↩
  3. Nieva, Richard. “Inside Scale AI’s Business After Meta’s Bombshell $14 Billion Deal.” Forbes, May 14, 2026. ↩
  4. Bastani, Hamsa, Osman Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” PNAS 122, no. 26 (2025): e2422633122. Correction (affiliation only): PNAS 122, no. 34 (2025): e2518204122. https://doi.org/10.1073/pnas.2518204122. ↩
  5. Oreopoulos, Philip, Oliver Keyes-Krysakowski, and Deepak Agarwal. “How In-School Supervised Ed-Tech Support Produces Massive Learning Gains: A Khan Academy Field Experiment in India.” NBER Working Paper 34683, January 2026. ↩
  6. Australian Government Department of Health and Aged Care. “24-Hour Movement Guidelines: Recommendations for infants, toddlers and preschoolers (birth to 5 years).” ↩
  7. Common Sense Media. “Risk Assessments.” ↩

Changes to this post

Published.