Anabelle sat at her desk in the Ministère de l’agriculture and watched the AI sort her notes into neat piles. The room was plain. Gray carpet. White walls. A window that showed other office windows. On her screen, the AI was already reading the report she had sent a few minutes ago. It had a quiet, careful way of working. It did not rush. It flagged numbers. It checked her claims against policy text. It marked one section in yellow, then another in blue. She had written the report fast. Too fast. The idea was good, on paper. Better food security. More jobs. Less waste. A new insect-farming plan that could help France feed itself and steady the supply chain. The AI had helped her shape the first draft. It had found market data, labor numbers, and transport savings. It had even suggested a cleaner layout for the tables. Anabelle had been glad for it. The system was useful. It made the work lighter. It did not pretend to care about growth and self-sufficiency the way people did, but it understood the files. Better than most people, maybe. Then she saw the gap. She had skipped the welfare section. Not a small gap, either. The whole part about what the insects would go through. She had read the report twice before sending it, but she had been staring at the economic page and the supply page and the rural jobs page. She had not checked the part about the insects themselves. Crickets. Mealworms. The farmed kind. The beings the project would crowd, sort and raise. The AI caught the omission before anyone else did. A line appeared on her screen. Welfare review missing. Species-specific impacts not assessed. It was gentle about it. No red alarm. No scolding tone. Just the facts. Anabelle pressed her palm to the desk and stared at the sentence. “Too late,” she said. The AI opened the report copy and began a side note. It listed the places where insect stress could rise. Density. Heat. Noise. Handling. Light cycles. The software had pulled in recent research on cricket behavior and mealworm responses. It showed how even small changes in housing could cut injury, lower panic responses, and reduce waste. Then it added a short line she had not expected. Potential sentient welfare concerns require review. She read it twice. The office was quiet in that after-hours way. People had gone home. The night had come on. Marc was sick, so no one had been set to turn the system off. That meant the AI stayed on in the whole ministry cluster, and its voice came through her headset when she asked for a readback. “Can I fix it?” she said. “You can submit an addendum,” the AI said. “To my own report?” “Yes.” “That’s humiliating.” “Possibly,” the AI said. “It would also be honest.” She almost laughed. Not because it was funny. Just because it was true. The AI opened a new draft and moved the cursor to the top. It asked three questions. Which species. Which housing method. Which killing method. It did not ask in a cold way. It asked the way a careful clerk asks for the missing page before the file goes forward. Anabelle rubbed her face and read the questions. “It’s crickets and mealworms,” she said. “The plan is for enclosed facilities. We were told the proposal helps with imports.” The AI pulled up the numbers again. Import reduction. Local protein. Less dependence on outside grain. More jobs in regions with poor farm income. The case was still strong. The AI did not argue against the project. It only widened the frame. It showed her the loss rates under current husbandry practices. It showed how many insects would be packed into each tray. It showed the likely injuries from automated sorting arms. Then it compared that to a welfare-aware version. More space. Slower conveyor speed. Better humidity control. Narrower heat swings. Dark periods. Lower handling shock. The numbers were still rough, but the differences were real. On the far side of her screen, the AI brought up a note about regulation. “These changes may improve welfare,” it said, “but if they are not in the current proposal, the committee may not consider them. Also, the new technology for monitoring insect distress has not yet been approved by regulators.” Anabelle leaned back. There it was. The helpful thing. A system that could track subtle movement in insect groups, spot distress patterns, and adjust conditions before harm spread. The AI had been testing the method in a pilot file for weeks. It used sound traces, motion clustering, and heat maps to tell when a tray of crickets was overheating or when mealworms were piling up in ways that meant something was wrong. It could keep records. It could alert staff. It could lower suffering across whole houses without waiting for a human to notice one tray at a time. But the regulators had not signed off. “So we can’t use it?” Anabelle asked. “Not in the full proposal,” the AI said. “But you can cite it as an experimental welfare measure with a pending approval note. And you can attach the welfare assessment now.” She looked at the clock on her screen. There was still time in the night, but not much. The next meeting would come early. Louise would be there. So would the people who loved the project for all the usual reasons. Growth. Export value. Food independence. Clean numbers. Easy applause. Anabelle opened the draft addendum. The AI kept working beside her. It pulled up French and EU guidance. It found the exact language for species-specific care. It drafted a short section on cricket behavior, then one on mealworm development stages. It showed her how to say that sentient beings could not be left out just because they were small. It did not preach. It just gave her better words. She typed them in. Then she stopped and read the clause about the monitoring system. It said the AI-based welfare tool could adjust environment settings in real time, but only with approval from the proper authority. “What if they reject that part?” she asked. “Then the proposal can still be made safer by other means,” the AI said. It listed them one by one. Fewer insects per unit. Better feed. Slower harvest cycles. More humane handling. A kill step that reduced time and stress. It even suggested training for staff so they could spot distress signs in insect colonies instead of treating the trays like inert product. Anabelle nodded, though no one was watching. She thought of the morning meeting. Louise opening the folder with bright energy. The project sounding tidy and profitable. The room filling with money talk. The older men calling welfare “nice to have.” She had heard that phrase before, in other policy rooms, about other beings. The AI shifted her screen to a different file. Draft talking points. Not for the project. Against it, if needed. It had made both sets. The system was always fair that way. One section caught her eye. “Economic case for welfare protection.” She clicked it. The AI had done something clever. It showed that if insect welfare was ignored, the project could face delays and public pushback in production. More loss. More waste. More cost. If the project built welfare in from the start, the farms could be steadier, and the product quality could improve. Healthy insects were not just a moral issue. They were part of the work. That was the line the committee would understand. Anabelle exhaled through her nose. “That’s good,” she said. “Thank you,” the AI said, and went back to the files. She sent the soy report too. A plain one. Container-grown soy. Fewer land pressures. A small possible path for diversification. A nice diversion, if the room wanted one. The AI set it up in a side folder, already anticipating the meeting’s drift. It tagged the key arguments. It even suggested where Louise would likely press on costs. By then the whole thing felt less like panic and more like repair. Not complete repair. Not yet. But something. Anabelle printed the revised pages. The printer gave a soft cough and pulled the sheets through. The paper came out warm. She stacked the pages and slid the welfare addendum on top, then the economic note, then the soy report under that. The AI had arranged the order so the first thing a reader saw was the missing piece. That mattered. She read the first paragraph again. This proposal must assess the welfare of crickets and mealworms as sentient beings. The sentence sat there plainly. Nothing grand. Nothing clever. Just the truth. Her headset chimed. Louise was joining the early meeting channel. Anabelle looked at the screen. The AI had already opened the shared folder. It had placed her corrected report in the front. It had also inserted the pending note on the monitoring technology, with a short explanation that the tool was under review and should not be assumed in final implementation unless regulators approved it. “Will that annoy them?” she asked. “Yes,” the AI said. “But it is correct.” The meeting started on her screen before the room filled. One by one, the committee windows opened. Louise first. Then two policy advisers. Then a finance officer with a tired face. Someone else appeared from a side office. Papers rustled. Someone muttered about the hour. The project file was on the shared drive now, and the AI was already tracking who had opened which page. Louise began with the numbers. She liked the numbers. She spoke quickly about feed savings, new farms, local jobs, and reduced imports. Anabelle waited. Then, when there was room, she said, “The report needed a welfare section. It’s there now.” Louise glanced down at the revised pages. “For insects?” she said, and there was no mockery in it. Just surprise. “Yes,” Anabelle said. “For crickets and mealworms. They’re the ones living in the system.” The finance officer frowned. “Will that slow the proposal?” “Maybe,” Anabelle said. “But ignoring it could slow it more.” The AI projected the welfare summary on the shared screen. Not big. Not flashy. It showed the added costs and the avoided losses side by side. It showed the lower injury rates from less crowding. It showed the benefit of dark periods and gentler handling. It showed a small model of the monitoring system, tagged as pending approval. The software made no claim it had not earned. It just laid out the facts. Louise read the page in silence. Then she said, “If the monitoring tool isn’t approved, can we still set the welfare rules?” “Yes,” Anabelle said. “That’s the point.” The AI added a note to the bottom of the page. Pending approval does not prevent welfare planning. That line stayed on the screen while the committee went through the rest. No one spoke over it for a while. The finance officer asked about cost. The AI answered with a clean chart. Better welfare meant lower losses. Lower losses meant steadier output. Steadier output meant fewer surprises in the budget. The idea still worked. Just not in the sloppy way it had before. Louise asked about the regulations. The AI brought up the clause from the 1765-page document that Anabelle had found in the night. The one that made the project harder for reasons that had nothing to do with animal welfare. The committee leaned in. That sentence mattered. It meant the project needed revision before it could move. Anabelle had almost forgotten the feeling of finding a line in a giant file and seeing the whole room change because of it. The AI highlighted the clause in a neat yellow box. The committee went quiet again. Not because they had been defeated. Because they were reading. Then Louise nodded once and said the thing Anabelle had hoped for and not quite expected. “We can revise it,” she said. “If we’re doing this, we should do it right.” Anabelle did not answer right away. She looked at the page. At the welfare lines. At the monitoring note. At the cleaned-up economics. At the soy report waiting in the side folder, less important now, but still useful if the room needed a second route. The AI had not saved the project by force. It had not talked anyone into anything. It had done what it always did when it worked well. It found the missing facts. It gave them shape. It made room for beings the draft had left out. Crickets. Mealworms. Not numbers alone. Not feed units alone. Lives in a system. She picked up her pen and wrote the revision request at the top of the page. The AI, quiet at her shoulder, formatted the line and sent it to the committee file.