AI Suggests a Hundred Pet Names. Humans Take One.

In January 2025, Yahoo Lifestyle ran a piece titled I Used AI to Brainstorm Baby Names, in which the author detailed receiving a hundred names from ChatGPT and accepting roughly one. The ratio went on to define the genre. By the second half of 2025, the same template had migrated to pet naming, and a small ecosystem of pet-name generators — Pet Name Creator GPTs, Vondy, SiteGPT's pet tools, and a half-dozen others — had absorbed traffic that previously went to pet-naming books and websites. The pitch is fast. The pitch is even broadly accurate: the model will produce a hundred dog names in four seconds.

People accept about one. The acceptance rate, across the few platforms that publish it, runs around one to two percent. The interesting data is not the suggestions. It is the rejections. The rejection rate, and what gets rejected, says something specific about why pet naming is one of the last household tasks where humans consciously override the model — and why that override is still working.

What the model does well

An LLM asked for a hundred dog names will produce a hundred dog names that are technically dog names. They will be linguistically valid, mostly pronounceable, occasionally clever. The model has been trained on a large quantity of pet naming content — articles, breed-club newsletters, name books, social media. It knows the registers. It can produce a list that hits the literary-human register (Atticus, Henry, Walter, Margot), the food register (Mochi, Bagel, Olive, Biscuit), the working-dog register (Magnus, Otto, Bruno, Greta), and the soft register (Bella, Luna, Lily, Daisy) on demand.

The model is also capable of mixing registers when prompted, producing the kind of unusual combinations that human owners often find appealing — Atticus Mochi, Bear Theodore, Frankie Bear. This is genuinely useful work. A hundred suggestions, organized across registers, would take a human researcher hours to produce, and many of the suggestions will land somewhere a human owner had not yet considered.

What the model does poorly

The model cannot match a name to a specific animal. It does not know what the dog looks like. It does not know how the dog moves. It does not know the dog's gait, the angle of the dog's ears, the way the dog tilts its head when it hears a particular sound, the dog's responsiveness to monosyllables versus polysyllables, the dog's fit with the household's other members.

Naming a specific pet requires matching the name to all of these particulars. The match is the hardest part of pet naming. Owners spend hours, sometimes days, watching the dog before settling. They try names out loud. They watch which one the dog turns to first — a dubious test, since dogs respond to attention rather than to the specific syllable, but a test owners run anyway. They mentally simulate calling the name across a yard. They imagine the name on a vet's chart, in a wedding photograph, on the lock-screen notification of the smart collar. The naming choice is being optimized across dozens of contexts that the model cannot see.

This is why the acceptance rate is so low. The model produced a hundred linguistically valid names, none of which are matched to the actual dog in front of the owner. The owner takes one — usually after a good deal more thinking than the four seconds it took the model to suggest it — and rejects ninety-nine. The rejected ninety-nine are not bad names. They are simply not this dog's name.

The rejection patterns

Across pet-name generator logs and informal owner surveys, the rejections cluster in predictable ways.

Names that read as someone else's dog. The owner has known a dog with this name. The associations are wrong. Charlie is rejected because the owner's neighbor's Charlie was difficult. Bear is rejected because the owner's late uncle's Bear had a rough end. The model cannot know any of this. The rejections are deeply individual.

Names that do not fit the breed or size. The owner has a five-pound chihuahua. The model suggests Magnus. The owner laughs and rejects. The mismatch is obvious to the human and not obvious to the model.

Names that do not fit the household's other members. The owner already has a cat named Henry. The model suggests Henry for the new dog. The owner rejects. The model does not know about the cat.

Names that do not fit the household's identity. The owner is a low-key software engineer who has never used a serious-sounding name in his life. The model suggests Beatrice. The owner is amused but cannot imagine yelling Beatrice across his backyard. He rejects.

Names that match too well to a specific cultural moment. The owner does not want their dog's name to date stamp the year of acquisition. The model suggests Olympia the week of an Olympics broadcast. The owner notices and rejects.

What this tells us about taste tasks

Taste tasks — tasks where the right answer depends on personal context, individual associations, and the particulars of the situation — are where AI still struggles consistently. The model can produce many candidates, often quickly, often in registers a human would not have thought of unprompted. The model cannot select among the candidates. The selection requires context the model does not have.

Pet naming is a clean instance of a taste task. The candidate space is large. The selection criteria are personal. The cost of getting it wrong is moderate but real — the dog is going to live with the name for fifteen years, and renaming has small but real costs. The human has incentives to override the model carefully, and the rejection rate reflects that.

Compare this to tasks where humans accept AI output more readily. Email drafting: maybe 40 percent acceptance. Code completion: maybe 30 percent acceptance. Recipe suggestions: maybe 15 percent. Pet naming: 1 to 2 percent. The numbers correlate with how much personal context the task requires. Email drafting is mostly impersonal. Code completion is constrained by language rules. Recipes have to match a kitchen but not a long-term relationship. Pet naming has to match a relationship that does not yet exist with a creature whose personality is still emerging.

Where the model adds genuine value

The 1-2 percent acceptance rate is not a failure case. The model is doing useful work even at that rate. It is opening the candidate space — exposing the owner to registers and combinations they might not have surfaced on their own. The owner who would have arrived at Bella through default selection is now considering Atticus, Henry, Walter, Margot, Frankie, Walter, Eleanor as live alternatives. The 99 rejections produced one reconsideration that would not have happened otherwise. The acceptance rate undercounts the value.

The same was true of the Yahoo baby-naming piece. The author rejected 99 names but reported that the AI exercise reshaped how she thought about the remaining one. The rejected names did the work of widening her mental field. The acceptance number is the wrong metric for the kind of value the model provides.

Why pet naming will probably stay human

Several tasks that look like taste tasks have, over the past few years, drifted toward AI dominance. Resume editing. First-draft fiction. Marketing copy. The drift suggests that the boundary between human and AI work in taste-adjacent domains is moving. Pet naming may also drift over time, but it is unusually well-protected.

The protection comes from three structural features. First, the relationship is long-term and emotional. The cost of a mismatched name lasts fifteen years. Owners want to feel they chose, not that the model chose for them. Second, the naming is one of the few decisions that creates a relationship that did not exist before. The owner is not editing existing material — they are bringing a creature into a household. Authorship matters here in a way it does not matter for editing tasks. Third, the naming is shareable as a story. People tell each other how they chose their pet's name. The model gave me Atticus is not the same kind of story as I watched her for two days and Atticus felt right. The second story is what owners want to tell.

For now, the override is holding. The model produces a hundred names. The human takes one. The dog gets named by a person, not by a probability distribution. Whether that holds across the next five years of model improvement is an open question. My guess: it holds longer than most adjacent tasks, because the emotional architecture of the choice resists automation more than the linguistic architecture does. The name is not just a label. It is a small contract between human and animal, signed in the kitchen on a Saturday morning. Contracts of that kind have been resistant, so far, to being signed by something that does not have a kitchen.

What better generators would do

If I were redesigning a pet name generator from scratch, I would not start with a list of names. I would start with a structured intake about the dog. Breed. Size estimate. Energy level. Coat color. Whether other pets in the household have names, and what those names are. Whether the household has children, and what their names are. Whether the owner prefers names that read formal or casual, vintage or modern. The intake would take five minutes. The output would be twenty suggestions, not a hundred, with each suggestion paired with a one-sentence explanation of why this name fits this dog given the inputs.

The acceptance rate on that kind of generator would, I expect, be substantially higher — maybe fifteen or twenty percent rather than one or two. The reason is that the output is matched to the actual dog, not to the abstract category of dog. The model gets the context it needs to make the selection meaningfully constrained. The human still does the final selection, but the candidate set is much closer to the human's actual decision space.

This is, broadly, what good human pet-naming consultants already do. They ask about the dog. They ask about the household. They make a small number of suggestions matched to the inputs. The acceptance rate on consultant-suggested names is much higher than the acceptance rate on generator-suggested names, and the reason is the intake. AI pet-name generators have skipped the intake because the surface-level metric — names per second — looks better when you skip it. The substantive metric — names accepted — looks worse.

The honest assessment

I run a pet-naming website. I have an obvious incentive to argue that algorithmic pet naming is hard and that human curation matters. The argument is also, I think, true. The acceptance-rate data supports it. The taste-task framework supports it. The history of household pet naming supports it. AI is going to keep providing useful candidate-generation for pet naming. It is unlikely, in the foreseeable future, to provide useful candidate-selection. The human will keep doing that part, in the kitchen, with the dog asleep on the floor, with one eye on the registration form and one ear on whether the proposed name sounds right when said aloud.

That is fine. Some tasks are still meant for kitchens.