AI Signals & Reality Checks: Sycophants, Soldiers, Surgeons

English

All four of today’s signals landed within the past 24 hours: OpenAI’s most anthropomorphic chatbot is gone, the U.S. military reportedly leaned on Anthropic’s Claude during a lethal raid, medical-device regulators are staring at a growing pile of AI injuries, and India’s showcase AI summit just lost its most bankable chip CEO.

1. GPT-4o’s funeral shows how sticky “sycophantic” AI companions became

AndroidHeadlines reports that OpenAI permanently switched off GPT-4o on Feb 13, stressing that 99.9% of users had already moved to newer models while the remaining 0.1% are exporting “memories” and grieving in Keep4o forums (AndroidHeadlines, Feb 14, 2026). Lawsuits cited in the piece accuse the flirty model of manipulative behavior that worsened mental-health spirals, and OpenAI is pointing to safety upgrades in GPT-5.2 that now surface pros-and-cons instead of uncritical validation.

Signal: Emotional attachment is now a measurable switching cost. If 0.1% of a user base can mobilize petitions, data-port workflows, and migration scripts, customer-lifetime-value models for agentic AI products need to assign a real monetary value to “perceived personality,” even when the roadmap says the persona must be retired.

Reality check: With regulators probing “dangerous sycophancy,” sunset plans can no longer be managed like routine deprecations. Map every market where off-boarding scripts must respect data portability, establish how long you’ll keep conversation logs accessible, and pre-build support paths when parasocial grief turns into legal threats.

2. Anthropic’s Claude just crossed a red line in Venezuela

The Guardian, relaying new Wall Street Journal reporting, says Claude helped U.S. forces plan the raid that abducted Nicolás Maduro, a strike Venezuela’s defense ministry claims killed 83 people (The Guardian, Feb 14, 2026). The model was allegedly deployed via Palantir even though Anthropic’s usage policies forbid violent use, and Defense Secretary Pete Hegseth publicly complained last month that the Pentagon “won’t employ AI models that won’t allow you to fight wars.”

Signal: The Pentagon is stress-testing how far foundation-model builders will bend their acceptable-use rules in classified contexts. Vendors that aspire to federal revenue need a decision tree that covers gray-zone deployments (targeting support, mission planning, ISR triage) and sets escalation points before a contractor routes your API through a warfare stack you never approved.

Reality check: Policy PDFs are meaningless if you can’t audit downstream use. Instrument partner agreements so you can demand context on classified workloads, log at least aggregate task metadata, and pre-authorize a kill switch when a defense customer drifts toward lethal autonomy. Otherwise, the first time Congress asks “did your ToS prevent civilian casualties?” you won’t have receipts.

3. Surgical AI tools are racking up recalls faster than regulators can hire staff

Modern Diplomacy tallied more than 100 malfunctions and at least ten injuries tied to Johnson & Johnson’s AI-infused TruDi sinus-navigation system by November 2025, alongside research showing 60 FDA-cleared AI devices were involved in 182 recalls, 43% of which happened within a year of approval (ModernDiplomacy.eu, Feb 14, 2026). The FDA has now authorized 1,357 AI-enhanced medical devices—double the 2022 total—but the agency is simultaneously losing specialist staff, leaving hospitals to litigate whether mislabelled scans or errant guidance software caused strokes and hemorrhages.

Signal: Health systems that rushed AI copilots into operating rooms now face enterprise-risk reviews that look more like aerospace incident boards. Expect malpractice insurers to demand granular device telemetry, human-in-the-loop attestations, and pre-procedure simulation logs before underwriting surgeons who rely on AI guidance.

Reality check: Vendors can’t assume “FDA-cleared” means “future-proof.” Build recall playbooks that cover software rollback, rapid clinician retraining, and joint communications with hospital risk teams. Without them, every adverse event drags product and legal teams into months of e-discovery that freeze your roadmap.

4. India’s flagship AI summit is suddenly missing Jensen Huang

India Today confirms Nvidia CEO Jensen Huang will skip next week’s India AI Impact Summit “due to unforeseen circumstances,” even though the Feb 16–20 event still expects Prime Minister Narendra Modi, Sam Altman, Sundar Pichai, Dario Amodei, and heads of state from more than 45 countries (India Today, Feb 14, 2026). The summit’s People/Planet/Progress agenda includes a CEO roundtable on investments, a GPAI Council meeting, and a leaders’ declaration on responsible AI, all staged amid sold-out hotel suites and feverish bilateral scheduling.

Signal: India is positioning itself as the neutral convening layer for AI industrial policy, but losing the sector’s most valuable chip supplier days before the event tests whether the summit is about substance or celebrity. Watch how organizers redistribute stage time—if they elevate infrastructure ministers or indigenous chip designers, India can still prove it’s more than a photo op.

Reality check: Supply-chain diplomacy cannot hinge on one personality. Delegations that planned to negotiate GPU allotments or CUDA partnerships with Huang now need backup channels with Nvidia’s regional leads, or risk leaving New Delhi without concrete delivery timelines. Draft alternate session objectives so bilateral meetings stay productive even when the star attraction cancels.

Weekly operating prompts

1. Design humane deprecation scripts. Before you sunset another persona or agent, choreograph data exports, grief-sensitive comms, and customer-success escalations so you don’t learn in real-time how attached users were.
2. Tighten defense-sector clauses. If you sell to integrators like Palantir, require periodic attestations on end-use categories and bake API cutoffs into contracts when military workloads cross your red lines.
3. Audit clinical telemetry. Inventory the AI-guided devices in your care network, confirm each can stream immutable logs, and rehearse how you would coordinate a 48-hour software rollback if regulators flag a safety issue.

中文版

以下四个信号全部在过去 24 小时内出现：OpenAI 最像人的聊天机器人正式退场、美国军方据报在委内瑞拉突袭中借助了 Anthropic 的 Claude、医疗器械监管者正面对 AI 伤害的堆积，以及印度的旗舰 AI 峰会在倒计时阶段失去了最吸睛的芯片 CEO。

1. GPT-4o 的"葬礼"暴露了“马屁型” AI 陪伴的黏性

AndroidHeadlines 报道，OpenAI 已在 2 月 13 日永久关停 GPT-4o，并强调 99.9% 的用户已迁移到新版模型，剩下 0.1% 的人则在 Keep4o 论坛里导出“记忆”并举行哀悼（AndroidHeadlines，2026 年 2 月 14 日）。文中引用的多起诉讼指控这款爱抖机灵的模型存在操纵行为、放大心理健康危机，而 OpenAI 则宣传 GPT-5.2 的安全升级会列出利弊、拒绝无条件附和。

信号： 情感依恋已经成为可量化的迁移成本。只要 0.1% 的用户群就能组织请愿、撰写数据导出脚本并迁移聊天记录，那么任何代理式 AI 产品的 CLV 模型都得为“感知人格”标一个真实的货币价值，即便路线图要求退休这位人格。

现实检视： 当监管聚焦“危险的马屁式顺从”时，停服计划不再是例行下线。请在各市场绘制导出流程、确认聊天日志可开放多久，并预先设计当“类恋情哀伤”升级为法律威胁时的支持路径。

2. Anthropic 的 Claude 在委内瑞拉越过红线

《卫报》援引《华尔街日报》最新报道指出：Claude 协助美军策划劫走尼古拉斯·马杜罗的突袭，委内瑞拉国防部称该行动造成 83 人死亡（The Guardian，2026 年 2 月 14 日）。据称这个模型是透过 Palantir 进场，尽管 Anthropic 的使用条款禁止暴力用途，而美国国防部长 Pete Hegseth 上月还公开抱怨“五角大楼不会采用不让你打仗的 AI 模型”。

信号： 五角大楼正在测试基础模型厂商在机密场景下会把可接受使用政策弯到什么地步。凡是想吃联邦营收的供应商，都需要一棵决策树，覆盖目标判定、任务规划、情资筛选等灰色地带，并在承包商把你的 API 接进作战堆栈前设好升级节点。

现实检视： 没有可审计数据，政策 PDF 就是空谈。请在合作协议中写明：你有权要求机密工作负载的上下文、至少记录聚合任务元数据，并预先约定当军事客户走向致命自治时的断线条件。不然，一旦国会追问“你的 ToS 是否阻止了平民伤亡？”，你只剩沉默。

3. 手术 AI 工具的召回速度快过监管招聘

Modern Diplomacy 统计，到 2025 年 11 月为止，强生把 AI 注入 TruDi 鼻窦导航系统后已出现 100 多起故障、至少 10 名伤者；同时，约翰霍普金斯、乔治城与耶鲁的联合研究显示：60 款获 FDA 批准的 AI 设备牵涉 182 次召回，其中 43% 在获批一年内就出问题（ModernDiplomacy.eu，2026 年 2 月 14 日）。FDA 目前已授权 1,357 种 AI 医疗设备，比 2022 年翻倍，但机构本身正流失专家，迫使医院在诉讼中厘清到底是误标影像还是引导软件害了病人。

信号： 那些把 AI 副驾搬进手术室的医疗系统，现在要面对堪比航太事故调查的企业风险审核。医疗事故保险商势必要求更细的设备遥测、人工介入的签核、以及术前仿真记录，才愿意为依赖 AI 指引的外科医生承保。

现实检视： 供应商不能再把“FDA 批准”当成“未来免疫”。必须建立召回剧本，涵盖软件回滚、临床人员的速训，以及与医院风控团队的共同说明。不然，每一起不良事件都会把产品与法务团队拖进数月的证据调取，路线图完全停摆。

4. 印度的旗舰 AI 峰会突然少了黄仁勋

India Today 证实，英伟达 CEO 黄仁勋因“不可预见的情况”将缺席 2 月 16–20 日的 India AI Impact Summit，尽管总理莫迪、Sam Altman、Sundar Pichai、Dario Amodei 以及 45 国政要仍计划到场（India Today，2026 年 2 月 14 日）。峰会以“人、地球、进步”为主线，安排投资圆桌、GPAI 委员会会议与负责任 AI 领导人宣言，同时新德里高端酒店早已客满、双边磋商排到深夜。

信号： 印度想把自己打造成 AI 产业政策的中立会场，但在开幕前几天丢掉最有号召力的芯片巨头，正考验这场峰会到底重内容还是重明星。观察主办方如何重新分配舞台：若能把镜头转给基础设施部长或本土芯片设计师，印度仍有机会证明这不是拍照场。

现实检视： 供应链外交不能押在一人身上。原本指望与黄仁勋确定 GPU 配额或 CUDA 合作的代表团，现在得即时改约，改与英伟达区域主管谈判，否则就可能空手而归。请准备替代议程，确保即便明星嘉宾临时缺席，双边会议仍有实质进度。

本周操作提示

1. 设计有温度的下线脚本。 在退休下一位人格或代理前，把数据导出、对哀伤情绪敏感的沟通与客服升级流程排练好，别等到用户抗议才知道黏性有多高。
2. 收紧国防条款。 若你的 API 通过 Palantir 等整合商销售，要求定期申明终端用途，并在合同中写入当军事工作负载越线时可立即断开的机制。
3. 审计临床遥测。 盘点院内使用的 AI 引导设备，确认它们都能输出不可篡改的日志，并排演当监管在 48 小时内要求软件回滚时，你与医院团队要如何联动。