23 May 2026: Google Maps AI’s Next Phase as Microsoft Pulls Claude Code Access

Google’s I/O Dialogues stage mapped the near-term direction for AI products, Microsoft has begun cancelling internal Claude Code licences, and researchers used AI to reconstruct crash pilots’ voices, forcing a US safety agency to take its records system offline.

Google’s annual I/O conference put its AI roadmap front and centre this week, with the Dialogues stage hosting candid discussions about where its most important products are heading. The recap, published directly on the Google blog, covers sessions on Gemini’s integration into everyday tools, the state of the company’s quantum computing programme, and its thinking on AI agents operating with greater autonomy.

Senior Google leaders used the Dialogues format to address questions about capability limits, deployment timelines, and responsible AI practices, giving a clearer picture of what Google Workspace and Search users should expect over the coming 12 months. For UK users, Gemini is being woven more deeply into Docs, Sheets, and Gmail, with agents expected to handle multi-step tasks on a user’s behalf. For context on how AI systems are tested before being given this kind of autonomy, our explainer on how AI evaluation and safety testing works is worth reading alongside the I/O announcements.

Microsoft has started cancelling internal Claude Code licences for its staff, according to reports surfaced on Hacker News, signalling a notable shift in how large technology companies are managing their AI developer tooling. Claude Code is Anthropic’s command-line coding assistant, competing directly with GitHub Copilot, which Microsoft owns. Pulling access suggests Microsoft is consolidating its AI developer stack around its own products rather than funding third-party subscriptions.

The decision matters beyond Microsoft itself. Large enterprise buyers are increasingly auditing which AI subscriptions they are paying for, and if the company that owns Copilot concludes it does not need a rival tool for its own workforce, others may follow the same logic. For developers currently using Claude Code through a company account, it is worth checking whether your employer is reviewing its AI procurement policy. This is a story about enterprise consolidation rather than a quality verdict on Claude Code itself, which remains one of the more capable coding assistants available.

IBM and Scuderia Ferrari are using AI to transform the experience of watching Formula 1, offering a practical preview of where personalised sports content is heading. IBM’s watsonx platform analyses race telemetry, historical data, and driver performance to generate tailored commentary and predictive insights for fans on Ferrari’s digital channels. The goal is to deepen engagement by giving casual viewers the kind of layered understanding usually reserved for analysts.

For UK sports fans, this is the direction Premier League clubs, Sky Sports, and governing bodies across cricket and tennis are moving toward. IBM is positioning watsonx as the AI layer for any content-heavy business delivering personalisation at scale. If you want a broader look at which AI products are delivering real-world value right now, our guide to the AI tools worth paying for covers the consumer landscape in detail.

Researchers used AI to reconstruct the voices of pilots killed in accidents, working from spectrogram images of cockpit recordings rather than the protected raw audio files, and the results were serious enough that the US National Transportation Safety Board temporarily took its public docket system offline. Spectrograms are visual representations of sound, and by feeding these images into an image-to-audio model, researchers produced reconstructed audio that bypassed the legal protections around cockpit voice recorder content.

In the US, cockpit recording content is legally protected specifically to encourage honest and complete reporting during aviation safety investigations. AI-generated versions, even imperfect ones, create substantial distress for families of those involved in fatal accidents and risk undermining the integrity of safety reviews. The case is a clear example of AI capability outpacing protective frameworks, and UK aviation authorities will be watching what response the NTSB puts in place.

Typing the word “disregard” into Google Search now breaks the AI-powered interface, and the issue appears connected to how the AI layer processes prompt-injection keywords. The word is commonly used in adversarial prompts designed to override AI model instructions, and Google’s Search AI appears to be treating it as a control signal rather than an ordinary query term. Google has not yet commented publicly on the issue.

It is not clear whether the fix requires a simple keyword filter or a deeper architectural change. For now, users of Google AI Mode should be aware that certain common English words can produce unreliable results. This is a small but telling example of what happens when AI systems built for controlled conditions meet the unpredictability of public use at scale, a gap we explore in our piece on model drift and how AI behaviour can change unexpectedly.

Watch Google’s response to the “disregard” issue over the coming days. If the company patches it silently, that is a signal it is treating the issue as a minor edge case. If it publishes a formal update to its AI Mode documentation or acknowledges an adversarial prompt vulnerability, that will be the clearest signal yet that Google is treating prompt injection in consumer products as a serious engineering problem rather than a niche concern.

Here is everything else worth knowing from today’s AI news.

• Elon Musk’s xAI has abandoned solar power ambitions on Earth. While promising a solar-electric economy, xAI is now running entirely on natural gas and SpaceX is pursuing orbital data centres instead. TechCrunch
• Google’s AI glasses are almost ready for consumers. Prototype Android XR spectacles overlaid Gemini-powered translation and navigation into testers’ field of view at I/O, with commercial launch details still to come. TechCrunch
• Nvidia’s Nemotron-Labs has released diffusion language models for faster text generation. The approach uses diffusion rather than autoregression, promising significantly faster inference speeds for text output. Hugging Face
• AI startups are inflating ARR figures and their investors know it. A TechCrunch investigation finds that some AI companies are stretching revenue metrics when talking publicly about growth. TechCrunch
• Anthropic has published its first Project Glasswing update. Glasswing is a cross-industry security initiative bringing together AWS, Apple, Cisco, Google, Microsoft, and Nvidia to harden critical software infrastructure. Anthropic

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s afternoon update on developments in artificial intelligence, published every weekday afternoon.