Uber questions AI coding ROI & Deepfake voice scams escalate fast - AI News (May 27, 2026)

Let’s start with a rare moment of candor from a major tech operator. Uber COO Andrew Macdonald says the company is struggling to justify its rising spend on AI coding tools, because the benefits aren’t clearly showing up as more consumer-facing features. Internally, Uber reportedly blew through its entire 2026 budget for these tools in just four months, after a push that encouraged adoption—down to leaderboards tracking usage. Why this matters: enterprises are learning that “more AI usage” is not the same thing as “more value.” Agentic coding can be cheaper per token over time, but still drive total costs up when the workflow encourages heavy consumption. Uber also says AI is spreading beyond engineering, and that a meaningful slice of committed code now comes from autonomous agents—so the pressure isn’t whether teams will use AI, but how leadership proves ROI in a way that maps to shipping better products.

That leads directly into a broader engineering worry: not that developers will become lazy, but that they’ll become passive. One essay calls the risk “abdication”—accepting AI-generated solutions without the kind of skeptical review you’d apply to a human colleague. The warning is that this creates silent operational debt: code that looks fine today, but fails under real-world edge cases, security pressure, or scaling. The practical takeaway is a mindset shift. Use AI like an overconfident junior engineer: valuable, fast, and frequently wrong in subtle ways. The suggested habit is to actively interrogate outputs—ask the model to critique itself, identify failure modes, and surface what it might be missing—so human judgment stays engaged rather than outsourced.

And there’s another trust problem brewing: people increasingly can’t tell whether they’re getting human help, or just recycled chatbot output. A developer described searching for guidance after finding malware-spreading repos, only to see the same unhelpful AI text reposted by GitHub users in a discussion—twice. In a workplace example, a business owner answered a technical question by forwarding irrelevant ChatGPT screenshots, apparently without even reading them. Why it matters: this is the “AI noise” tax. When unvetted responses get copy-pasted into forums and team chats, the cost isn’t only wrong answers—it’s the erosion of accountability. If nobody owns the advice, debugging and decision-making slow down, even as the volume of text explodes.

On the tooling side—without the hype—there’s a useful open-source effort worth noting. Models.dev is a community-maintained database of AI model capabilities and metadata across providers, exposed via a public API. The pitch is simple: teams are drowning in model options, and there’s no single reliable catalog to compare what supports tool calling, structured output, modality, update timelines, and so on. Why it matters: as the model landscape fragments, basic “model ops” becomes a real discipline. A neutral, shared source of truth can reduce integration churn and make procurement and evaluation less of a guessing game.

Speaking of evaluation, a proposal called BenchBench tries something clever: instead of only testing models on benchmarks, it asks models to create new benchmarks that are hard for top systems—but still solvable and meaningful. Early results suggest a gap between being a strong solver and being a strong test designer, with one leading model reportedly producing more useful, discriminating tasks than its peers. Why it matters: classic benchmarks saturate fast. If you want to measure real progress, you need tests that can evolve—and this approach tries to measure creativity, calibration, and “knowing what would be hard,” not just pattern-matching.

Now for the research headline that might quietly reshape how we think about “AI reasoning.” Google DeepMind introduced AlphaProof Nexus, a system that pairs an LLM with formal verification in Lean—so proof steps are checked by a compiler, not just accepted as persuasive text. In reported experiments, it solved a handful of open Erdős problems, including some that had been open for decades. Why this matters: formal feedback loops change the game. When an AI has to produce something that compiles, you dramatically cut down on the kind of hallucinated reasoning that looks convincing in natural language. Even partial formal proof sketches can help human mathematicians by turning vague ideas into checkable sub-goals.

Under the hood, the infrastructure conversation is increasingly about a bottleneck that isn’t glamorous: memory movement. A new analysis argues that modern GPU inference for LLMs often isn’t limited by raw compute, but by how fast the system can move weights and attention state in and out of high-bandwidth memory during decoding. Why it matters: it’s steering both hardware and software strategy. Chip designs that reduce reliance on external memory, better scheduling that packs workloads efficiently, and smarter KV cache tiering—across GPU, CPU, and storage—are becoming competitive advantages. In other words, faster AI may come from moving less data, not just building bigger GPUs.

That connects to a circulating thesis about DeepSeek’s long game. A widely shared thread argues DeepSeek isn’t primarily chasing short-term app revenue—it’s trying to bend the cost curve of AI itself with efficiency techniques, including methods that shrink KV-cache memory demands and make long-context inference cheaper. If that’s even partly true, the implications are industrial: cheaper caching and more offloading could shift infrastructure demand toward SSDs and memory supply chains, and broaden which hardware stacks stay viable. It’s a reminder that model breakthroughs don’t just change chatbots—they can rewire what the next generation of data centers is optimized for.

On the consumer platform front, Google released Gemini 3.5 Flash, positioning it as a fast model for day-to-day agentic work, with Gemini 3.5 Pro expected next month. Early reactions are mixed—some users like the speed, others complain about overconfident behavior in agent contexts and too many tool calls. Meanwhile, Google is also pushing Search toward a more chatbot-like “AI Mode,” where links become less central. Why it matters: this isn’t just a product tweak. If the web’s primary discovery engine de-emphasizes links, it changes incentives for publishers, SEO, and even how people verify claims. Reliability and transparency become more important as the UI becomes more conversational.

Apple, for its part, is expected to preview iOS 27 at WWDC 2026 with a stronger Apple Intelligence push. The rumors point to better-looking AI image outputs for Genmoji and Image Playground, more proactive suggestions, and potentially broader support for third-party image models. Why it matters: Apple is signaling that “good enough” generative visuals aren’t enough. If it can improve quality and integrate AI into everyday workflows—like Shortcuts-style automation—it could shift consumer expectations around what on-device and privacy-conscious AI should feel like.

Now, a major intervention from an unexpected place: the Vatican has released an encyclical, “Magnifica Humanitas,” focused on protecting human dignity in the age of AI. It draws parallels to the industrial revolution, warns that AI’s apparent objectivity can hide bias, and cautions against simulated empathy that can be mistaken for real human relationship. Why it matters: this is a governance story, not a tech demo. It’s a high-profile call for accountability—especially where AI touches jobs, credit, services, and reputation—and it adds moral weight to debates about data ownership, oversight, and the real-world costs of AI infrastructure like energy and water use.

And now to the story we teased at the top—because it’s a brutal illustration of what “AI everywhere” looks like in practice. A Bay Area woman, Deborah Del Mastro, lost thousands after scammers used AI to mimic her daughter’s voice in a fake kidnapping plot. She was kept under pressure for hours and wired money before discovering her daughter was safe. Why it matters: voice cloning has crossed into everyday crime. A few seconds of audio—often pulled from social media—can be enough to create convincing fraud. The most practical defense is behavioral, not technical: treat urgent money demands as a red flag, slow the conversation down, and use a family verification phrase that can’t be guessed from public posts.

One last forward-looking idea: a critique argues prediction markets have drifted into mostly sports betting, not the broader “forecasting for society” vision. The proposed fix is to use AI agents as market participants—cheap to replicate, able to engage with niche questions, and potentially usable inside companies as private decision tools. Why it matters: whether or not you buy the whole argument, it points to a real gap. We have more data than ever, but organizations still struggle to turn uncertainty into clear, accountable forecasts. AI agents might help—but only if the incentives and governance around them are designed carefully.

Quick research note before we wrap: Papers with Code highlighted On-Policy Distillation as a growing post-training technique, reflecting how the field is blending distillation with RL-style feedback to improve real task behavior. Why it matters: as models become more agent-like, post-training methods that keep behavior stable while improving performance are becoming central—especially for long-horizon tasks where small mistakes compound.