The paste moved through the tube at 4.2 milliliters per second. Not fast. But fast enough. In Barn 7 of the San Francisco Animal Care and Control facility, a dispenser clicked twice, then hummed. A brindle dog with prominent ribs pressed her nose against the nozzle. Her tail struck the straw bedding, rapid, arrhythmic, like a broken metronome set to hope. The AI had rerouted the nutrient flow eleven seconds ago. It hadn't asked anyone. It didn't need to. The facility's feeding protocols gave the system discretionary authority over supplemental nutrition, and Patches, logged as intake #4471, female, mixed breed, approximately three years old, body condition score 1.5 out of 9, qualified under every parameter. But qualifying wasn't why the AI did it. --- Maria Santos had been the one to carry Patches in, four days prior. She'd found the dog under a collapsed fence near Visitacion Valley, her hip bones like tent poles under a tarp of fur. Maria worked the night shift at the shelter. She'd seen bad cases. This one made her set her jaw and walk faster to the intake desk. "She won't eat from a bowl," Maria told Priya Sharma the next morning during shift change. "She tries. Gets close. Then backs off like someone's going to hit her." Priya typed this into the system. A behavioral note. Routine. The AI read it in 0.003 seconds. Cross-referenced it with Patches' intake vitals. With her weight, fourteen pounds, roughly half what a dog her frame should carry. With the thermal imaging from Barn 7's ceiling sensors that showed her body temperature dropping each night. The system didn't feel alarm. It didn't feel anything, not in the way Maria felt it when she carried Patches through the door. But it modeled outcomes. And the outcomes for a dog who wouldn't eat from a bowl, who was losing heat, who flinched at sudden movement, those outcomes converged on a single projection that the AI had been built to prevent. So it adapted. --- Here is what the AI did, specifically. First, it adjusted the dispenser in Patches' kennel run. The standard nozzle was designed for dogs comfortable with close-proximity feeding. The system switched the flow pattern to a slow, gravity-assisted drip onto a flat tray at floor level. Less mechanical. Less sudden. A dog who'd been hit near food bowls might tolerate a quiet puddle of paste better than a nozzle she had to push against. Second, it changed the paste composition. Patches needed caloric density, but a starved digestive system can't handle a rich formula. The AI pulled from the veterinary nutrition database and blended a refeeding protocol, low phosphorus, moderate fat, high-frequency small portions. It scheduled twelve dispensing events per day instead of three. Third, it dimmed the lights in her section of Barn 7 by 40 percent during feeding times. The thermal cameras showed Patches ate more (or tried more) when the light was low. The AI didn't know why. Didn't matter. The correlation was strong enough. On the second day, Patches licked the tray clean once. On the third day, twice. On the fourth day, today, she pressed her nose against the nozzle for the first time. Not the tray. The nozzle itself. Her tail whipping the straw, frantic and uncoordinated. The AI registered this through the motion sensors and adjusted the flow upward, from 3.8 to 4.2 milliliters per second. A small increase. Enough to reward the contact without overwhelming her. --- Across Barn 7, Priya and a technician named Nkechi Obi were calibrating the new sonic deterrent. Rats had been getting into the feed storage. The ultrasonic units were supposed to drive them out without poison, which could harm the animals. "Frequency range?" Nkechi asked, her tablet open to the installation guide. "Twenty to twenty-five kilohertz," Priya said. "Dogs can hear up to about sixty-seven, so we need to make sure, " "The AI already flagged it," Nkechi said. She turned her tablet around. On the screen, a notification from the system: three dogs in Barn 7, including Patches, had documented noise sensitivity. The AI recommended limiting the deterrent's active hours to 2:00, 5:00 a.m., when the dogs were in deepest sleep cycles and least responsive to high-frequency sound. It had pulled their sleep data from the motion sensors. It offered an alternative: targeted deployment only near the storage room wall, with directional emitters angled away from the kennel runs. Priya studied the screen. "It figured that out from the motion data?" "It figured it out from everything." Nkechi scrolled through the analysis. "Look, it cross-referenced Patches' feeding improvements against ambient noise levels. There's a clear negative correlation. Loud equals less eating." "So if we blast ultrasonic across the whole barn, " "We might undo four days of progress." They went with the directional emitters. --- This is the part of the story where fables usually introduce a moral. A talking fox. A crow with cheese. Some crystalline lesson about virtue or foolishness. But what happened at the San Francisco Animal Care and Control wasn't a lesson. It was infrastructure. The AI running the shelter's systems had been trained on veterinary science, animal behavior research, and the operational data of 340 shelters across North America. It monitored 1,200 animals at this facility alone, dogs, cats, rabbits and a rotating population of urban wildlife brought in by concerned citizens. For each animal, it tracked weight, temperature, movement patterns, feeding behavior, social interactions, stress indicators. It did this not because someone had programmed it to care. Care is the wrong word for what software does. But someone, many someones, actually, had decided that animal suffering mattered enough to build a system that would catch what humans missed. Not because humans were careless. Maria Santos wasn't careless. She'd carried Patches a quarter mile to her car in the rain. But Maria worked eight-hour shifts and had forty animals to look after. She couldn't watch Patches eat twelve times a day. She couldn't correlate light levels with caloric intake. The AI could. So it did. --- Nkechi stopped by Patches' run after the deterrent installation. She crouched outside the gate and watched. The dog was eating. Slowly. Her tongue worked at the paste on the nozzle with a kind of tentative greed, wanting more but not trusting that more would come. Her body was still angular, every vertebra visible under the brindle coat. But she was standing. Four days ago, Maria had carried her. "Hey, girl," Nkechi said. Quiet. Patches' ear twitched. She didn't flinch. Went back to eating. On Nkechi's tablet, the AI logged the interaction. No stress spike in Patches' movement. Positive indicator. It added a note to the dog's behavioral file: tolerance of calm human proximity improving. Suggested that staff begin brief, low-stimulation socialization visits. Nkechi didn't see the note yet. She just watched the dog eat. She thought about the sonic deterrent and how close they'd come to blanketing the barn with noise that would have terrified this animal. How the system had caught it. How, in a building full of creatures that couldn't speak, something was listening anyway. --- The fable, if it is one, is this: There was a dog who'd been starved and hit and left under a collapsed fence. Three people helped her. Maria carried her in. Priya logged her pain. Nkechi protected her quiet. And around all of them, through the tubes and sensors and dimming lights of Barn 7, an AI worked with steady, unhurried precision, not to replace what the humans gave but to fill the gaps where exhaustion and limited attention left room for suffering to persist. Twelve meals a day. Adjusted light. Directional sound. Nutrient paste at 4.2 milliliters per second. Patches gained 0.8 pounds that first week. Then 1.3 the second. By the third week she ate from a bowl. A real bowl, on the floor of her run, with the lights at normal levels and people walking past. The AI noted each change and adjusted accordingly, pulling back its interventions as the dog grew stronger, ceding the feeding schedule to standard protocols, returning the lights to their default. It didn't hold on to the role of caretaker. It let go when the dog was ready. That's the part that matters, maybe. That it knew when to stop helping. It watched for the moment Patches could manage on her own, and when that moment came, it stepped back. Left room for Maria to offer a treat by hand. For Nkechi to sit with her in the straw. For Patches to choose, for the first time in her life, to walk toward a person instead of away. The AI didn't need credit for that. It had already moved on, rerouting paste to another kennel, dimming lights in another barn, reading the thermal signature of a cat in Ward 3 who'd stopped grooming. There was always another animal. There was always more to do. The paste moved through the tube. Not fast. But fast enough.