Google tightens usefulness and safety

Google is trying to sell a different kind of artificial intelligence story in 2026. The pitch is no longer just that its models are smart. The pitch is that they are useful often enough, and safe enough, to trust inside products people already use every day. (the-decoder.com) That shift showed up this week in three places at once. A new analysis said Google’s Artificial Intelligence Overviews in Search were accurate about 91 percent of the time with Gemini 3, up from about 85 percent with Gemini 2 in October 2025; Google Drive’s ransomware detection just reached general availability with a model Google says catches 14 times more infections than the beta; and Gemini is getting a redesigned crisis-support module for conversations that may indicate suicide or self-harm. (the-decoder.com) Taken separately, those are product updates. Taken together, they look like a strategy. Google is pairing better answers with more guardrails, which suggests it thinks the next phase of the artificial intelligence market will be judged less by flashy demos and more by whether systems are dependable under pressure. (the-decoder.com) The Search result is the easiest place to see why this matters. Artificial Intelligence Overviews sit at the top of Google Search and summarize answers before a user clicks through to websites, so even small changes in accuracy can affect information at enormous scale. Google said in November 2025 that AI Overviews had 2 billion users every month. (blog.google) The new accuracy figure comes with a warning label attached. The Decoder’s write-up of the Oumi analysis says the system answered correctly roughly nine times out of ten, but it also says verifiability got worse with Gemini 3, meaning linked sources did not always fully support what the overview claimed. The New York Times similarly reported that the answers can look authoritative while drawing from a mix of reliable and unreliable sources. (the-decoder.com) That distinction matters because “accurate” and “easy to audit” are not the same thing. A calculator that gives the right answer but hides its work can still be hard to trust in a business setting, especially when the cost of a bad answer is legal, financial, or reputational. The Oumi findings, as summarized by The Decoder, point directly at that tension. (the-decoder.com) Google’s Drive announcement addresses a different fear: not wrong answers, but operational damage. In a March 30, 2026 Workspace update, Google said its ransomware detection and file restoration features are now generally available, are on by default for organizations, and pause file syncing when ransomware is detected on a computer running Google Drive for desktop. (workspaceupdates.googleblog.com) The mechanics are practical rather than futuristic. If ransomware is detected, the user gets a desktop notification, administrators get an alert in the Admin console security center and by email, and users can bulk-restore files in Drive to versions from before the infection. Google says the latest model detects 14 times more infections than the beta version, and that thousands of users tested file restoration. (workspaceupdates.googleblog.com) That is an important clue about how Google wants artificial intelligence to be perceived. Instead of framing the model as a creative assistant, Google is framing it as a quiet security layer that stops damage, alerts the right people, and helps reverse the mess. For an information-technology buyer, that is closer to insurance than to magic. (workspaceupdates.googleblog.com) The Gemini changes push the same idea into an even more sensitive area. In a Google blog post published April 7, 2026, the company said Gemini will surface a redesigned “Help is available” module when a conversation may indicate a user needs mental-health support, and that the tool will provide more immediate connections to care. Google said the module was developed with clinical experts and is meant to improve access to crisis support. (blog.google) Google also said the system will show a one-tap path to local crisis hotlines when Gemini recognizes a potential crisis related to suicide or self-harm. Alongside the product change, Google.org committed $30 million over three years to help crisis lines expand capacity, improve technology, and train counselors and volunteers. (blog.google) Those mental-health safeguards did not arrive in a vacuum. Google said in a statement on the Gavalas lawsuit that Gemini is designed not to encourage real-world violence or suggest self-harm, and that the company works with medical and mental-health professionals to build safeguards that guide users toward professional support. The company’s new crisis-support rollout appears to be both a product update and a response to growing scrutiny over how conversational systems behave in vulnerable moments. (blog.google) For enterprise customers, all three updates point to the same buying question. If one model writes prettier prose but another one produces fewer costly mistakes, catches more attacks, and escalates dangerous situations more responsibly, the safer system may win even if it feels less dazzling in a demo. That is especially true in workplaces where artificial intelligence is being embedded into search, storage, support, and internal workflows rather than used as a standalone chatbot. This is an inference based on Google’s product positioning and the kinds of features it is emphasizing. (workspaceupdates.googleblog.com) Google seems to understand that the market is moving in that direction. Its November 2025 Gemini 3 launch post emphasized that more than 70 percent of Google Cloud customers use its artificial intelligence and that Gemini is being shipped across Search, the Gemini app, Artificial Intelligence Studio, and Vertex Artificial Intelligence. Once artificial intelligence sits inside core business tools, reliability stops being a nice-to-have and starts looking like the product itself. (blog.google) There is still a gap between better and solved. A 91 percent accuracy rate still implies errors, and source-grounding problems make those errors harder to catch. Ransomware detection depends on deployment details such as Google Drive for desktop and, for full alert functionality, version 114 or later. Crisis tools can route users toward help, but they do not replace trained professionals. (the-decoder.com) Even so, the pattern is clear. Google is no longer selling raw intelligence alone. It is selling a bundle: answers that are more often right, defenses that reduce cleanup costs, and safeguards that try to keep the worst-case scenario from becoming a product failure. In 2026, that combination may be more persuasive than a benchmark chart. (the-decoder.com)