Analysis limits. I had Claude pull both papers and draft this review. Both are open access and were read in full from the journal pages. The PLOS ONE model tables are published as images, so effect sizes here come from the authors' text and captions, not the tables themselves. The Scientific Reports supplement (full model output) and the raw data and code on GitHub were not opened or rerun. No adversarial pass by a second model. My own commentary lives in the newsletter.
Everyone has seen the videos. A dog walks to a mat of hexagonal buttons, presses one, and a recorded human voice says OUTSIDE. Then, sometimes, a second one. MAD. The owner gasps. Four million views.
The science problem is a hundred and twenty years old. Clever Hans, a horse in Berlin, did arithmetic by tapping a hoof until a psychologist named Oskar Pfungst showed that Hans was reading the tiny release of tension in whoever asked the question. Every animal language project since has been judged by how well it keeps Hans out of the room. Washoe, Koko and Nim Chimpsky mostly did not.
Federico Rossano's Comparative Cognition Lab at UC San Diego runs the largest citizen science project on button trained pets, and these are its first two empirical papers. The first asks whether dogs understand the buttons when a human presses them. The second asks whether dogs press them deliberately. Comprehension first, then production, which is the right order.
Bastos and colleagues, PLOS ONE, August 2024. Two experiments using the same protocol. In the first, researchers visited 30 dogs at home. In the second, 29 different dogs were tested by their own owners following written, video and Zoom instructions, then filmed.
The home visit design is the best thing in either paper. The researcher covered the dog's OUTSIDE, PLAY and FOOD buttons, plus a freshly recorded nonsense word (DAXING), with randomly colored stickers. A second researcher, hidden in another room, told the first one which color to press through headphones playing music, so the presser never heard which word came out. The owner sat in the room wearing a sleep mask and noise canceling headphones. Afterward, video was trimmed and every human was blacked out before coders who knew nothing about the hypotheses scored what the dog did. Agreement between coders was acceptable (kappa 0.70). The whole thing was preregistered.
That is about as Hans proof as a living room gets.
After PLAY, dogs were roughly seven times more likely (in odds) to do something toy directed. After OUTSIDE, about seven times more likely to head for the door. Those numbers come from comparing each word against the average of the three real words, which is a fair specificity test. It asks whether PLAY produces play more than OUTSIDE or FOOD do.
What it leaves outFOOD did nothing. The authors suggest the dogs were full, since the home visits ran during working hours. Plausible, and untested. The nonsense word, which was the natural baseline, never entered the main models. And the authors' own discussion describes accuracy as only slightly better than chance and comparable to dogs following a pointing finger. Seven times the odds of a rare behavior is still a modest absolute effect.
No detectable difference between the stranger in Experiment 1 and the owners in Experiment 2. Within Experiment 2, dogs responded about the same to a pressed button and to the owner saying the word aloud, which argues the dogs are attending to the sound and not only to where the button sits on the board.
What it leaves out"No difference" here means credible intervals that cross zero in about thirty dogs per arm, so the study failed to detect a difference. Showing the responses are the same would take an equivalence test, which was not run. The spoken versus pressed comparison exists only in the owner run arm, where the owner necessarily knew the word. The authors themselves note the dog may simply have learned two parallel associations, one for the sound and one for the spot on the board.
The heading for the speech question labels Experiment 2 as the home visit study. It was the citizen science study. The discussion says dogs responded to food and outside words, while the results say play and outside. A table cross reference points to the wrong model. None of this changes the findings. It does suggest nobody read the final version slowly. The preregistered head tilt question went unanswered because the dogs tilted their heads nine times in total.
Bastos, Houghton, Naranjo and Rossano, Scientific Reports, December 2024. A completely different kind of study. Owners logged every press, theirs and their dog's, in the FluentPet phone app over 21 months. The analysis kept only dogs with at least 200 logged interactions. Owners chose the words, recorded them in their own voices, laid out the boards, and sorted each word into one of 68 preset concept categories.
The paper sets itself three tests. Presses should be nonrandom, nonaccidental, and not simple copies of the owner.
Each dog's pairs of presses were drawn as a network, buttons as nodes and pairs as links. Real boards looked like hubs (one or two favorite buttons linked to many others). A thousand random networks of the same size and density per dog did not. Bayes factor 72.7, which is strong. Most dogs, though not all, beat the random version.
What it leaves outThe comparison network treats every pair of buttons as equally likely. Any dog with a favorite button, a habit, or a board where some tiles are easier to reach will beat that, meaning or no meaning. A random model that preserved each button's popularity would have been the harder and more informative test. The paper clears the bar it set. The bar is low.
The argument is clever. Every household's board is laid out differently, so if dogs were stepping on buttons by accident, no particular pair of concepts should be favored across the population. Some pairs were. FOOD with PLAY, GO OUTSIDE with OTHER, and HELP with OTHER showed up more than their individual frequencies predict. LATER with LOVE YOU and, interestingly, NOW with WANT showed up less.
What it leaves outThe paper never reports or analyzes the layouts, so it cannot show they were independent. Owners buy the same commercial tiles and learn from the same online community, and words that go together tend to get placed together. Every press was also logged by a human who chose to open the app. The authors concede owners may preferentially record presses they find interesting, then argue the bias washes out at population level because owners differ. Owners also share a culture of what counts as interesting. Order within each pair was discarded, so nothing here bears on syntax.
How often a dog pressed a concept was barely predicted by how often the owner pressed it (coefficient 0.014). Dogs press OUTSIDE and food words. Owners press I LOVE YOU. The dogs are not simply echoing.
What it leaves outThis compares totals by concept, not sequences in time, so it says little about whether a given press followed a recent owner demonstration. The sample was restricted to experienced dogs, after most of the modeling had happened. The authors say so themselves.
The paper concludes that dog soundboard use is deliberate and that some dogs have associated different outcomes with different buttons. It states plainly that whether dogs understand the words is beyond its scope.
What the university saidThe UC San Diego release has Rossano describing dogs "pressing buttons purposefully to express their desires and needs," with two button sequences that seem to reflect specific requests. That is the claim the paper explicitly declined to make. Coverage repeated the press release.
Both papers come from Rossano's lab, with Amalia Bastos (now a postdoc at Johns Hopkins) as first author. The only declared funding is her Hopkins fellowship.
The company ties are disclosed and substantial. In the Scientific Reports paper, two authors had consulted for CleverPet, which makes the FluentPet app and buttons, and a third works there. Every press in the dataset came from the company's app under a data sharing agreement. In the PLOS ONE paper, six authors had consulted for FluentPet and two were employees. UC San Diego says the design and analysis were done independently. A company that sells buttons has an obvious interest in the finding that dogs understand buttons. That is a reason to read carefully. The lab did several things that make careful reading possible. Both studies were preregistered. Data and code are public. The blinding in the home visit experiment is better than most of the field manages.
The lab has also become the public face of the topic, with appearances in a Netflix documentary and a 2026 NOVA episode. A search in October 2026 found no published follow up study testing production under controlled conditions, which both papers name as the decisive next step.
There are two ways people have tried to get inside animal heads through language. One is to decode what animals already say to each other. That is the elephant, whale and marmoset work this newsletter covered in August as Classifier, Not Translator. The usual failure is reading too much into the machine.
The other is to teach an animal our symbols and watch what it does with them. That is Washoe, Koko, Kanzi, Alex the parrot, and now several thousand dogs in living rooms. The usual failure is reading too much into the animal, generally through the person holding the treats. Soundboards belong entirely to the second tradition. Their advantage over the apes is scale. Their disadvantage is that the trainers are pet owners who love the subject.
| Who | Channel | What was shown | Standing |
|---|---|---|---|
| Animals taught human symbols | |||
| Clever Hans | Hoof taps, Berlin, early 1900s | Arithmetic, apparently | Pfungst showed Hans read the questioner's posture. The control every study below is built around. |
| Washoe, Koko, Nim | Signs adapted from ASL | Vocabularies of hundreds of signs | Terrace and colleagues (1979) concluded Nim was largely imitating trainers. Mostly anecdote, heavily disputed. |
| Kanzi (bonobo) | Lexigram keyboard, spoken English | Comprehension of spoken words tested over headphones to strip cues | The strongest ape case. Died March 2025. How far it counts as language is still argued. |
| Alex (grey parrot) | Spoken English | Labels, colors, shapes, numbers | Irene Pepperberg received a lifetime special award from the Coller Dolittle Prize in 2026. |
| Rico, Chaser | Spoken object names | Hundreds to over a thousand toy names | Real, but confined to rare "gifted word learner" dogs. |
| Sofia (dog) | Keyboard of recorded words | Requests for walks and play | One professionally trained dog (2008). The only dog in the literature producing requests before these papers. |
| Soundboard dogs | Owner recorded buttons | Blinded comprehension of PLAY and OUTSIDE. Nonrandom pressing in self reported logs. | These two papers. Comprehension is the stronger half. |
| Animals' own signals, decoded | |||
| Elephants | Rumbles | Name like calls addressed to individuals | Careful work, modest signal. In August's piece. |
| Marmosets | Phee calls | Possible vocal labels for others | Formally disputed in the literature. |
| Bonobos | Wild vocalizations | Calls combined so one modifies the meaning of another (Science, 2025) | 2026 Coller Dolittle finalist. |
| Bottlenose dolphins | Nonsignature whistles | Shared whistle types with context specific responses | Won the first Coller Dolittle Prize, 2025. |
| Zebra finches | Calls | Mapping the meaning of a songbird's call repertoire | Julie Elie won the 2026 Coller Dolittle Prize. |
| Human to animal, in the animal's code | |||
| Honeybees | Robotic waggle dance | Foragers recruited by a robot's dance | The one case of a human sending a novel message an animal acted on. No AI involved. |
Set against that table, the soundboard comprehension result is respectable and unsurprising. Dogs have long been known to learn spoken words for routines. The new part is showing it survives a blind presser, a blind owner and blind coders. The production result is where the excitement lives and where the evidence is thinnest. A dog pressing OUTSIDE at the door is a request any dog can already make by standing at the door. The interesting claims (absent objects, the past, LATER, the two word phrase that means something neither word does) are exactly the ones these papers do not test. For the older version of this problem, where the people mistake their own reading for the animal's meaning, see Kelly, the Sassy Dolphin.
The first paper is a small, careful experiment showing that dogs trained by their owners respond to two of three button words with every human cue removed. The second is a very large pile of owner logged presses that beats pure randomness, which almost anything does.
The experiment that would matter is the one both papers promise. Create a situation the owner cannot see, and check whether the dog presses the button that matches it. Until someone runs it, Clever Hans remains on retainer.
Sources
Paper 1: Bastos APM, Evenson A, Wood PM, et al. How do soundboard-trained dogs respond to human button presses? An investigation into word comprehension. PLOS ONE 19(8):e0307189, 28 Aug 2024. doi:10.1371/journal.pone.0307189. Preregistration osf.io/tskcq. Code github.com/znhoughton/comprehension.
Paper 2: Bastos APM, Houghton ZN, Naranjo L, Rossano F. Soundboard-trained dogs produce non-accidental, non-random and non-imitative two-button combinations. Scientific Reports 14:28771, 9 Dec 2024. doi:10.1038/s41598-024-79517-6. Preregistration osf.io/kr7h9. Code github.com/znhoughton/Modeling.
Press release: UC San Diego Today, Dogs use two-word button combos to communicate, study shows (Dec 2024). today.ucsd.edu. Project scale and history: University of California, Can our pets really say "I love you"? universityofcalifornia.edu. NOVA: today.ucsd.edu.
Earlier work cited by the papers: Pfungst O, Clever Hans (1911). Terrace HS et al., Can an ape create a sentence? Science 206:891 (1979). Rossi AP, Ades C, A dog at the keyboard, Animal Cognition 11:329 (2008). Kaminski J et al., Science 304:1682 (2004). Pilley JW, Reid AK, Behavioural Processes 86:184 (2011).
Kanzi: Scientific American obituary, March 2025. scientificamerican.com
Bonobos: Berthet M, Surbeck M, Townsend SW. Extensive compositionality in the vocal system of bonobos. Science 388:104 (2025). doi:10.1126/science.adv1170
Coller Dolittle Prize: 2025 winner, Sayigh (dolphins), whoi.edu. 2026 winner, Elie (zebra finches), finalists and Pepperberg award, coller-dolittle-24.sites.tau.ac.il.
Prior WAiR: Classifier, Not Translator (Aug 2026), with sources for the elephant, marmoset, sperm whale and honeybee work. Kelly, the Sassy Dolphin (Aug 2026).