Companion apps have pulled decisively ahead of Tinder and Bumble in total engagement hours — and the gap is still widening, quarter over quarter.
New data from SensorTower shows just how far the balance has tipped. In the first quarter of 2026, Americans logged roughly 705 million hours on AI companion apps, compared with about 280 million hours spent on dating and social-discovery platforms like Tinder, Bumble, and Hinge combined.
The trend isn't new, but it's accelerating. A year earlier, in Q1 2025, companion apps held a narrower lead — around 580 million hours against 330 million for dating apps. By the third quarter of 2025, companion-app hours had already climbed past 715 million while dating-app hours slipped under 300 million. Five consecutive quarters of data now point the same direction.
The shift sits inside a larger pattern: SensorTower's broader State of AI 2026 report finds that overall time spent on generative AI apps is set to more than double year over year. Dating apps haven't necessarily gotten worse — they've simply stopped being where people choose to spend idle time, while companion apps have gotten better at making themselves hard to leave.
A Munich ruling says Google's AI summaries are the company's own statements — not protected third-party content — opening the door to liability for every AI answer engine.
The Regional Court of Munich issued a temporary injunction barring Google from repeating false claims its AI Overviews made about two Munich-based publishers, wrongly linking them to scams and subscription traps that never appeared in any of the underlying sources the tool cited.
The court's reasoning marks a break from how search engines have traditionally been treated. Germany's Federal Court of Justice had long granted search operators only indirect liability, since they merely pointed to outside websites. The Munich court found that doesn't apply here: AI Overviews evaluate, combine, and rewrite information into self-contained answers, which makes them Google's own speech.
Google rejected the central defense that users can always click through and verify sources themselves — research shows barely 1% of users do. Google says it will appeal; the ruling is not yet final, but its logic could extend well beyond one case, and beyond Germany.
An export-control directive citing national security has suspended all access to Fable 5 and Mythos 5 worldwide — even for Anthropic's own foreign-national staff.
Anthropic disabled Fable 5 and Mythos 5 for every customer after receiving a directive from the US government at 5:21pm ET, citing national-security authorities. The order targeted access by any foreign national, inside or outside the US — a scope broad enough that the company says a blanket worldwide shutdown was the only practical way to comply.
Anthropic says the government's letter did not specify the concern in detail, but the company understands it relates to a technique for bypassing Fable 5's safeguards. Anthropic disputes the severity, calling the disclosed jailbreak narrow and non-universal, and noting that comparable capability is reportedly accessible through other publicly available models not subject to the same restriction.
Access to other Claude models, including Opus 4.8, is unaffected. Anthropic says it agrees governments should be able to block unsafe deployments — but argues this directive didn't follow a transparent, fact-grounded process, and that it is working to restore access.
The OpenAI Partner Network arrives three months after Anthropic launched its own, signaling that deployment — not raw model power — is the new enterprise battleground.
OpenAI's first formal global partner program launches with Accenture, BCG, McKinsey, Bain, and PwC as founding members, structured across three tiers — Select, Advanced, and Elite — based on sales performance, technical depth, and deployment experience.
The company framed the move bluntly: model capability is no longer the limiting factor for enterprise AI value — workflow redesign and change management are. Anthropic's Claude Partner Network, launched in March with $100M behind it, had already certified more than 10,000 consultants from over 40,000 applicants by mid-June.
A new "AI Control Roadmap" assumes, for planning purposes, that a capable internal agent could evade oversight, copy itself, or sabotage safety work — and builds layered defenses accordingly.
The roadmap borrows directly from cybersecurity practice, adapting the MITRE ATT&CK framework to map potential agent misbehavior across three risk categories: loss of control, sabotage of safety research, and direct harm. Safeguards escalate through evaluation, active monitoring, and intervention — the ability to restrict or shut down an agent in real time.
DeepMind is careful to note the plan is precautionary: an internal review of roughly one million coding-agent sessions found the vast majority of flagged events stemmed from overeagerness or misinterpretation, not adversarial intent.