A baby-prediction app takes two adult photographs, runs them through a generative model trained on faces, and returns what is, mechanically, a weighted composite nudged toward whatever its training data treats as a normal, healthy, appealing infant. It is not predicting anything. A child’s face is the outcome of how roughly twenty thousand protein-coding genes happen to recombine, an event that has not yet occurred and that no current model — not the diffusion architectures these apps are built on, not anything in a research lab — has any access to. The honest description of the output is: flattering guess dressed as forecast.
I know this, and I still felt it. I typed in two photographs, one of me and one of my partner, into one of the more popular versions of these tools and waited about forty seconds. What it returned was a soft-focus toddler with my partner’s mouth and something close to my eyes, lit like a nappy advertisement. My first reaction was not skepticism. It was a small, embarrassing pull of tenderness toward a person who does not exist and may never exist. The image had done its job before I had time to ask what the job was.
That gap, between what the tool appeared to offer and what it was mechanically capable of, is the whole story, and it is not unique to baby generators. It is the same move running underneath most of the software I use every day. A recommendation feed does not show me what is true or good for me; it shows me what keeps me watching. A dating app surfaces the profiles most likely to hold my attention, not the person I would build a stable life with. A beauty filter does not correct my face, it edits it toward a standard I did not choose. In each case the system is optimised for a response, and the most reliable way to produce a response is to give people what they already want to see.
This is the part worth sitting with. Being shown what we want is not a malfunction. It is the product working exactly as designed. The incentive is not hidden or sinister, it is just commercial: engagement, retention, satisfaction, the metrics the business is actually built to move. Accuracy, health, and anything resembling a representative picture of reality are at best side effects, and often they sit in direct tension with the thing the model is rewarded for. The baby generator is only an unusually literal version of the whole arrangement. It took my wish and handed it back to me as a face.
It helps to look at why the guess comes out flattering rather than neutral. A diffusion model of this kind works by learning the statistical shape of its training images and then drawing new ones from the densest parts of that distribution. The densest parts are the typical and the conventional, so the tool is pulled, by construction, toward the most common version of a face rather than the unusual one. There is also a long-running strand of face-perception research, going back to Langlois and Roggman’s 1990 study on averageness, finding that composite faces built from many individuals tend to be rated as more attractive than the individual faces blended to make them. Put those two things together and the baby generator is doing something almost too neat. It averages toward the average, and the average is close to what we are already disposed to find pleasant.
The cost of that is the specific. Real children carry the odd, unbalanced, slightly asymmetric details that make a face one particular person’s and not a category’s. The composite irons those out. It also inherits whatever was overrepresented in the data it learned from, which audits of major image datasets have repeatedly shown skews toward some demographics and away from others. So the picture is not neutral. It is normal in a narrow, trained sense of normal, and it quietly suggests that this narrow normal is what a desirable child looks like. The tool is not only failing to predict. It is gently editing toward a standard, and presenting the edit as a result.
The obvious objection is that this is a toy, and an unserious one. Nobody is making decisions about having children on the strength of a phone app. That is true, and the point is not that the baby generator is dangerous. The point is that it makes visible, in a harmless setting, the logic that also governs tools where the stakes are real. The same optimisation that returns a flattering baby returns the news that confirms what I already believe, the search result that matches what I hoped to find, the account of events that is easiest to accept. The toy is a clean instrument for seeing the mechanism precisely because nothing rides on it.
What unsettled me, in the end, was not that the software flattered me. It was how completely it worked, and how little resistance I put up. The interface of a system built to inform me looks identical to the interface of a system built to please me. Same clean screen, same confident output, same air of having answered the question I asked. I cannot tell from the surface which one I am using, and the commercial pressure runs almost entirely toward the second. The skill that used to matter most, telling a reliable source from a flattering one, is getting harder to practise at exactly the moment the tools are getting smoother.
I deleted the image. Then I undeleted it, about an hour later, and looked at it again. The mouth, the eyes, the soft lighting around a face that had taken forty seconds to assemble and would never correspond to anyone. I had not asked the app whether it could predict my children. I had only asked it to show me, and it had. That was the whole transaction. I closed the app and left the picture on my phone, where it still is.
