Google announced on October 15 that AI Mode and AI Overviews will show more links inside generated answers and place them higher on the page. The company is also testing a dedicated section that puts one or two official destinations at the top with an explanation of why each link is useful.
The update addresses a basic problem in AI search: a summary can answer quickly while making the original reporting, documentation, or creator harder to see. Google is also expanding timely-article carousels to more places and languages and testing clickable creator and publisher names on video results.
Google says use of its Preferred Sources feature has grown more than sevenfold since the start of 2026. That figure shows people are using the control, not that the new layout will send more visits to publishers or make answers more accurate. The dedicated official-link area is still a test. The next useful evidence will be whether people open more original pages and whether creators receive measurable referral traffic.
OpenAI lets developers describe a voice in words
OpenAI released prompt-to-voice for all API customers on October 16. A developer can describe a voice in text, create it in the API or Voice Playground, and reuse it in GPT-Live sessions. OpenAI also added nine preset voices across German, French, Hindi, Japanese, and Korean.
Creating a prompt-generated voice is free, while generated speech uses normal GPT-Live session billing. The release makes custom voice production more accessible, but the announcement does not compare accent accuracy, expressiveness, or robustness in noisy rooms across languages. Developers also remain responsible for deciding when a synthetic voice needs clear disclosure.
Google reports a 40-fold drop in one quantum error measure
Google Quantum AI demonstrated what it calls magic-state cultivation on its Willow processor. Magic states help quantum computers perform a wider set of useful operations, but producing clean ones has consumed so much capacity that some estimates assign up to 95% of a machine to the task.
In work published in Science, the team reported a 40-fold reduction in errors compared with uncultivated states and 99.99% logical fidelity for T-states. Researchers also transferred the state between two error-correction codes. This is a laboratory demonstration of a difficult component, not a useful general-purpose quantum computer. Google says the method still needs to work at larger scale.
Android can ask AI to inspect a suspicious text message
Google added a scam-checking flow to Circle to Search. On a compatible Android device, a user can select a suspicious text message and receive an AI Overview with warning signs and suggested next steps before opening a link or sharing personal information.
Google notes that results, availability, and compatibility vary. It did not publish an independent false-positive or false-negative rate. The feature can add a pause before a risky click, but it cannot replace checking the sender, domain, requested payment, and official contact channel.
AI agents learn from the physical results of lab experiments
Google Research described three projects that connect AI agents to robotic wet-lab work. Its VERID system keeps a live digital model of a laboratory workstation, simulates proposed actions, blocks physically invalid commands, and records synchronized video and action logs.
In one three-round aptamer study, agents changed their design rules after physical test results exposed non-specific binding. The reported on-target hit rate reached 33.5% for myoglobin and 42.1% for alpha-thrombin. The experiments show that agents can use real feedback to repair a search strategy on selected molecular tasks. They do not establish autonomous general-purpose laboratories or clinical effectiveness.
OpenAI offers a free academic plan to selected US labs
OpenAI opened applications for the Rosalind Academic Plan. Eligible life-sciences labs at US universities can receive up to 30 ChatGPT Business Standard seats at no charge for as long as 12 months. Access to GPT-Rosalind requires a separate trusted-access approval.
An application does not provide immediate access, and the program has not reported research outcomes. Its immediate effect is narrower: some approved labs can test the tools without paying for the standard seats during the program period.
Google says 400 publications joined its News AI pilot
Google said more than 400 publications now participate in a licensing program that pays publishers to provide enhanced content for AI features. The pilot includes article overviews on Google News publisher pages and customizable audio briefings. New partners named by Google range from CNN and the Financial Times to local US newsrooms.
The figure comes from Google and counts participating publications. Google did not publish payment totals, referral traffic, subscription results, or the effect on newsroom work. Those outcomes will determine whether the program becomes a durable exchange of value rather than another distribution experiment.
Microsoft makes proofs work with ordinary programs
Microsoft Research introduced Magnetar, a system that produces a short cryptographic proof that a program ran as claimed. Developers can write ordinary Rust, Go, or C code, compile it for the RISC-V processor architecture, and verify the resulting proof in about 12 milliseconds without revealing private inputs.
Across 15 benchmark programs, Microsoft reports that Magnetar generated proofs 10 to 39 times faster on average than the fastest comparable systems on the same machine. The results come from the research team’s benchmarks. Production cost, developer experience, and security under real workloads still need broader testing.
This week’s releases ask how another person can inspect an answer, voice, decision, or experiment that AI produces. Better links, physical feedback, and cryptographic proofs offer different parts of that answer. None removes the need to measure what happens outside the announcement.
This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.