“Keep them exactly the same but make it look like a school picture.” That is the prompt TikTok creator Carlie Bond gave ChatGPT after photographing her three children against a plain backdrop at home, according to Fast Company, and the output was close enough to the standard gradient-backdrop products sold by Lifetouch and Cady Studios that the video went viral as a “MomHack”. The workflow, as gadgetreview.com reported, replaces a school photo package that commonly costs $25 to $60 per child with a phone camera and a chatbot. For Bond’s family of three kids, that is $75 to $180 per year at stake — real money, which is why the video spread.
What the workflow actually is
The mechanism is simpler than it sounds. A current-generation multimodal model accepts an uploaded photograph plus a text instruction and returns a synthesized image. The model is not pasting a new background behind a masked face; it regenerates the whole frame conditioned on the pixels it received. That distinction is where the two risks live. The face in the output is the model’s reconstruction of the face in the input, not the face itself. And the input — a child’s photograph — is now a file on the vendor’s servers, governed by that vendor’s retention policy, not by you.
Accessibility depends on plan and account configuration; image upload is a supported ChatGPT capability, but free-tier accounts have tighter limits on image generation than paid ones. The marginal cost of the edited photo itself is effectively zero if you already have a plan, or the price of one month’s subscription if you do not — either way, well under one $60 package.
Risk one: the face drifts
Bond explicitly instructed the model to keep her children “exactly the same.” Instruction is not a constraint. Diffusion and autoregressive image models have no mechanism to hold pixels fixed on request; “preserve the features” is a soft preference the model weighs against everything else in its sampling. Dexerto reported that some parents who tried the trick found their children’s faces changed in generation even after asking for fidelity — eyes, proportions, hairline shifted by amounts visible to anyone who knows the child. The output is then an AI-stylized portrait of someone who resembles your kid, not a photograph of them.
For a keepsake this may be acceptable. For anything where the image functions as a record — yearbooks, identification, a family archive — it is worth being precise about what the file now is. gadgetreview.com recommends inspecting the eyes, facial proportions and hairline before treating the output as finished. That is a checklist for detecting drift after the fact; there is no setting that prevents it.
Risk two: where the upload goes
Under OpenAI’s privacy policy, uploaded images count as user content, and per the company’s retention documentation, a file stays associated with the account for the retention period of its chat. The controls that exist are real but coarse. You can disable “Improve the model for everyone” in Data Controls, which keeps future conversations — including these uploads — out of training data. You can run the session as a Temporary Chat, which keeps it out of your history. But a Temporary Chat is not immediately deleted: OpenAI’s own documentation states the company may retain a copy for up to 30 days for safety purposes, and that if a custom GPT routes data through an external action, the third party’s retention policy governs that copy instead.
So the honest accounting is: for up to 30 days after a private session, and indefinitely for a normal saved one, a photograph of a minor sits in a vendor’s systems under terms a parent did not negotiate. OpenAI requires parental or guardian permission for users under 18 and provides a reporting route for accounts of children under 13 — the upload itself, by a parent, is permitted. Permitted is not the same as consequence-free. The child’s biometric-adjacent data — a face — has no take-back mechanism once processed, beyond deleting the conversation.
The version that is not a life hack
A second practice circulating on social media is materially different and should not be lumped in with the Bond workflow: uploading a professional proof to an AI tool to remove the watermark rather than buying the finished image. That is feeding someone else’s copyrighted work into an editor to strip the mechanism they use to get paid for it. Whether it is unlawful depends on the contract and the jurisdiction — gadgetreview.com is careful to say so — but unlike the plain-wall prompt, it is a rights question about a photographer’s image, not a privacy question about your own.
The economics that started all of this are stable: $25 to $60 per child per year, against zero marginal cost at home. Neither the face-drift failure rate nor the retention window is measured against that price in any of the viral posts. The numbers to watch are the ones nobody publishing the hack reports: what fraction of generations preserve the face recognizably, and whether OpenAI ever shortens the 30-day safety copy on Temporary Chats involving children’s images.