AI targeting and accountability debate & Apple and Google Gemini for Siri - AI News (Mar 26, 2026)

We’ll start with the most sobering story on the list: reporting on the February strike in Minab, Iran, where a primary school was hit during Operation Epic Fury, killing roughly 175 to 180 people—mostly young girls. A lot of public attention zoomed in on whether Anthropic’s Claude “picked” the target, but the deeper critique is about process, not personality. The piece argues this was about kill-chain compression: Project Maven—now embedded in a broader Palantir-built targeting infrastructure—can fuse intel, generate target packages, and move from detection to action faster than older workflows. That speed also means a bureaucratic mistake, like a facility mislabeled in a database and never corrected after it became a school, becomes instantly lethal. The takeaway isn’t that AI replaces responsibility—it’s that automation can amplify the consequences of stale data, weak oversight, and human decisions made in the name of tempo.

In a related accountability thread—this time in court—a federal judge in Northern California suggested the U.S. government’s ban on Anthropic may look retaliatory and potentially unconstitutional. Judge Rita Lin indicated the Pentagon’s move appeared aimed at crippling the company after Anthropic spoke publicly about a contracting dispute, raising First Amendment concerns. This case matters beyond a single vendor: it could shape how national-security authorities can pressure AI suppliers, and whether speaking up about government contracting risks becomes a chilling effect across the industry.

Now to Apple’s AI strategy, which keeps looking more like a two-track race. According to The Information, Apple has been granted “complete access” to Google’s Gemini model inside Google’s own data centers. The key point isn’t that Apple wants to ship Gemini as-is—it’s that this level of access reportedly enables distillation. In plain terms: Apple can use a very capable model to generate strong answers and reasoning traces, then train smaller models that are cheaper, faster, and tuned for specific tasks—ideally able to run directly on-device without a network connection. That’s a big deal for latency, reliability, and privacy, especially if Apple wants Siri to feel instant and dependable. The report also suggests Apple can tune Gemini’s behavior to better fit Apple’s product constraints—though Gemini’s current “personality” is said to be optimized for chatbot and coding patterns, which may not map perfectly to Siri. The partnership is expected to support a more conversational Siri in iOS 27, while Apple continues building its own foundation models so it’s not permanently dependent on Google.

Staying with Apple, there’s also a research note worth paying attention to: Apple researchers report that some base, pre-instruction-tuned LLMs can provide meaningful confidence estimates about whether an answer is semantically correct—even though these models are trained mainly to predict the next token. They introduce a framework around “semantic calibration,” and the practical warning is just as important as the promise: instruction-tuning with reinforcement learning, and even chain-of-thought prompting, can degrade that calibration. If you’ve been hoping that “model confidence” can become a reliable safety signal, this work is a reminder that common post-training techniques may quietly break the very uncertainty cues we’d like to depend on.

On the developer tooling front, Anthropic introduced “auto mode” in Claude Code, a new permissions setting that reduces the constant “approve this command” friction in longer coding sessions. Instead of asking for user approval every time it touches files or runs a shell command, Claude can make routine permission decisions—while a safeguard classifier reviews each tool call before it executes. The intent is to make coding agents more autonomous without going fully hands-off via the more dangerous “skip approvals” approaches. Anthropic is upfront about the tradeoffs: extra checks can add latency and overhead, classifiers can miss edge cases, and sometimes they’ll block benign work. But directionally, this is a sign of where coding agents are headed: fewer interruptions, more continuous execution, and more emphasis on guardrails that sit between the model and the system.

That theme—optimizing the whole agent loop, not just the model—also shows up in an open-source project called “nit,” a Git replacement written in Zig. The pitch is simple: Git output was designed for humans scanning terminals, but AI agents often pay for every token they read. The developer analyzed real sessions and argues that shrinking default output can cut token usage and speed up workflows, especially for repetitive commands like status and log. The larger trend here is subtle but important: as AI-assisted development scales, we’re going to see more “machine-first” interfaces—tools that still behave like familiar developer utilities, but speak in a more compact, agent-friendly way to reduce cost and latency.

Another open-source angle is “Ossature,” a spec-driven harness meant to keep LLM-generated software coherent across multiple modules. The project’s premise is that the hard part of AI code generation isn’t producing one file—it’s maintaining consistency across interfaces, behavior, and dependencies over time. Ossature leans on structured specs, ambiguity checks, and build plans to keep generation grounded and verifiable. Whether this particular tool wins mindshare or not, it highlights a broader shift: the most valuable work in AI coding is increasingly orchestration—how we constrain, evaluate, and iterate—not just raw generation.

On the evaluation side, ServiceNow researchers introduced EVA, a framework for measuring conversational voice agents across full phone-style dialogues. EVA produces two headline scores: one for task accuracy and one for user experience—because in voice, users can’t skim, can’t reread, and small timing or transcription errors can wreck the interaction. Their benchmarking across many systems found a consistent tension: agents that complete tasks reliably often do worse on conversational experience, and nothing dominates both. The significance is that voice agents are becoming integrated systems—tools, policies, audio, and dialogue management—and we’re finally getting benchmarks that treat them that way, rather than grading a single model response in isolation.

In healthcare, the Electronic Frontier Foundation filed a FOIA lawsuit against the Centers for Medicare & Medicaid Services seeking records related to WISeR, a multi-state Medicare pilot using AI to assess prior-authorization requests. EFF’s concern is familiar but high-stakes: automated decision-making can create delays or denials, and without transparency it’s hard to know what data the system learned from, what bias protections exist, or how errors are monitored. The report also flags incentives that could be troubling—vendors potentially paid based on the amount of care they deny. Regardless of where you land politically, the “why it matters” is straightforward: when AI systems influence medical coverage decisions at scale, the public needs visibility into testing, auditing, and accountability mechanisms.

From Google Research, TurboQuant is a new set of quantization techniques aimed at compressing the high-dimensional vectors used in two places that get very expensive: LLM KV caches for long context, and vector indexes for semantic search. The headline isn’t the math—it’s the bottleneck: memory. Long-context systems can become constrained by how much they must store while you keep a conversation or a document in working memory. If compression can lower memory use without degrading output quality, it changes the economics of serving long-context LLMs and running large-scale retrieval. In practice, work like this can be as impactful as a model upgrade, because it targets the cost and throughput limits that determine whether advanced features are usable outside demos.

OpenAI is pushing ChatGPT further into shopping. The update adds more visual discovery—product grids, comparisons, and image-based matching—while leaning into merchant feeds through an expanded Agentic Commerce Protocol. OpenAI is also stepping back from its earlier Instant Checkout approach and letting merchants keep their own checkout flows, which suggests the company is prioritizing being the starting point for discovery rather than owning the full transaction. Walmart is also launching an in-ChatGPT app experience that moves users into a Walmart environment with account linking and payments. The platform implication is big: if chat becomes the front door for shopping research, whoever controls ranking and presentation will influence demand in a way that starts to resemble search—only with even fewer clicks between suggestion and purchase.

That push comes alongside a staggering funding update: OpenAI’s CFO said the company secured an additional $10 billion, pushing the round to over $120 billion, with investors ranging from venture to mutual funds and sovereign capital. OpenAI also signaled it’s preparing for the possibility of going public, while acknowledging compute constraints and tough prioritization—reportedly including shutting down its short-form video app, Sora. The broader meaning here is that frontier AI is now a capital structure story as much as a research story: model capability is tied to infrastructure scale, and infrastructure scale is tied to fundraising on a historic level.

Zooming out, there’s an argument gaining traction that the classic App Store model will be disrupted by AI agents that complete tasks by calling APIs instead of downloading apps. In that view, the value chain splits into connection, discovery, and payment—where connection becomes commoditized by open standards, and discovery becomes the true choke point because agents will choose services on a user’s behalf. If that’s right, ranking power becomes the new gatekeeper, with monetization that looks less like a 30% platform fee and more like an auction for attention—except the conversion is nearly guaranteed because the agent is acting. It’s a useful lens for thinking about the next platform fight: not “who has the best app,” but “who controls the recommendations an agent trusts.”

On the research side of reasoning, Alibaba’s Qwen team says we’ve been measuring Reinforcement Learning with Verifiable Rewards—RLVR—in a slightly misleading way. Instead of looking only at how much token probabilities change after RLVR, they argue the direction of change matters, and they propose using signed token-level differences to identify which tokens are truly reasoning-critical. Their experiments suggest a small subset of tokens carries a disproportionate load, and amplifying the model along that learned direction at test time can improve reasoning without new training. The practical takeaway: as “reasoning” becomes a product feature, teams are hunting for levers that improve accuracy cheaply—test-time techniques and diagnostics that can squeeze more out of a trained model.

Finally, Anthropic put out two pieces that together sketch where agents are actually going in real usage. First, its Economic Index analyzing about a million Claude conversations finds consumer use is broadening into everyday tasks while some coding shifts toward API-based automation. It also highlights learning curves: longer-tenure users tend to get higher success rates and apply Claude to more work-related tasks, suggesting “learning-by-doing” could widen productivity gaps between early adopters and everyone else. Second, Anthropic described new harness designs for autonomous app building—separating a generator agent from an evaluator agent to reduce the model’s tendency to rubber-stamp its own work. The message is that autonomy isn’t just a model problem; it’s a systems design problem—how you plan, how you critique, and how you verify over multi-hour runs.