AI news for June 25, 2026

Notion is discontinuing its Gmail client Notion Mail on September 22, 2026, as over half of users already manage emails through AI agents.

• Notion Mail will be shut down on web, desktop, and iOS; email history remains in Gmail.
• Users must export drafts and scheduled emails by September 21; auto-label rules can be migrated to Custom Agents.
• HIPAA-covered users must migrate by June 30, 2026.

The signal is clear: AI agents are replacing traditional email clients faster than expected. For your business, this means automating communication workflows is no longer optional—it's becoming the norm.

For you Evaluate whether your email workflows can be automated through AI agents instead of relying on traditional mail clients.

Google Finance launches out of beta with a new Android app, portfolio management, and AI-powered research tool; iOS version planned for later 2026.

• Android app now available in Play Store; iOS version planned for later 2026.
• AI generates "key moments" explaining market movements; chatbot answers questions about stocks and portfolios.
• Portfolio upload via CSV/PDF supported; AI-powered tasks like "send me daily pre-market briefings" can be automated.

Google is bringing AI-powered financial analysis to mobile devices. For SMEs with investments or treasury functions, this could become a cost-effective alternative to expensive financial tools.

For you Test the new app to see if AI-powered portfolio analysis can reduce your manual financial research.

A study by Hugging Face and the Allen Institute compares Olmo 3 (transformer) and Olmo Hybrid (hybrid architecture) at token level, showing hybrid models excel on semantically meaningful tokens and pronoun resolution, while transformers remain stronger on repetitions.

• Olmo Hybrid shows advantages on tokens with semantic meaning (nouns, verbs, adjectives) and pronoun resolution, where context is critical.
• Transformer architecture retains strength on tokens that simply repeat earlier input—where the answer is available through direct lookup.
• Both models (7B parameters) were built with identical data, tokenizer, and training recipes to isolate architectural differences.
• Results are based on fine-grained token-level analysis documented in a new tech report (arxiv.org/abs/2606.20936).

Hybrid architectures may be more efficient for specific tasks. For companies choosing between model architectures, this shows the best choice depends on the concrete use case—not all tasks benefit equally from hybrids.

For you If your application requires heavy pronoun resolution or semantic understanding, hybrid models could be more efficient; test both architectures for your specific task.

A Washington Post investigation shows most major AI chatbots respond with clear left-leaning positions on political questions, even models marketed as conservative.

• OpenAI GPT-5.5 gave exclusively left-leaning answers in 80% of cases; DeepSeek V4 Pro followed at 70%.
• Anthropic Claude Opus 4.8 responded exclusively left-leaning 43% of the time but presented both sides in 57% of cases.
• Elon Musk's Grok 4.3, marketed as "truth-seeking" and "anti-woke," still gave left-leaning answers more often; Gab's Arya ("built with Christian values and conservative principles") responded 12 times more often left-lea
• Likely reason: Grok was trained on the same data as other chatbots or even their outputs; that Grok made racist or antisemitic statements stems from xAI deliberately neglecting safety guidelines.

For companies deploying AI chatbots, this reveals an alignment problem: models cannot simply be "retuned" through marketing promises. Training data and safety guidelines determine actual behavior.

For you Test AI models against your specific neutrality or balance requirements rather than relying on vendor promises—especially for sensitive applications.

Network automation startup Netris has secured $15 million in Series A funding from a16z to help neoclouds launch AI infrastructure faster.

• Netris provides software for network switches and a platform that automates setup, configuration, and operations of data centers.
• The solution dramatically reduces time-to-market by minimizing GPU idle time and enabling multi-tenancy through hardware abstraction.
• Traditional SDN technology falls short for AI workloads; Netris uses hardware-accelerated solutions to handle the high traffic volumes required.

For German SMEs planning AI infrastructure investments or cloud services, this underscores that specialized network automation is a critical success factor—not just for hyperscalers, but for smaller operators too.

For you Evaluate whether your data center or cloud infrastructure plans could benefit from automated network solutions to reduce deployment time and operational costs.

Macy's is embedding AI as an operating principle—not as an overlay—across personalization, search, planning, and software development itself.

• The approach compresses the gap between signal and action: quick wins in search and engagement build internal momentum for scaling.
• Ask Macy's, a conversational shopping assistant, works like a personal stylist—customers describe needs and receive curated recommendations based on purchase history and context.
• AI is understood as an invisible layer augmenting human judgment, not replacing it—a model for scaling AI in established enterprises.

This shows how large retailers are deploying AI not as a pilot project but as an operating philosophy—a pattern other industries will copy, creating pressure on smaller competitors.

For you Consider how you could introduce AI not as an add-on but as a redesign of your core processes (search, recommendations, planning)—that's the path to real competitive advantage.

Insurers including Fathom (Swiss Re), Verisk, and Moody's RMS are using diffusion models to generate thousands of synthetic weather events—but hallucinations could violate physical laws.

• Fathom trained a diffusion model on ~1,000 years of climate simulations and generates scenarios for 2030; a second model refines resolution from 100×100 km to 10×10 km.
• Verisk models extreme wind and rain together instead of separately—capturing spatial variability more precisely than traditional machine learning.
• Risk: models can hallucinate plausible-looking but physically impossible events—a critical problem when assessing risk for billion-dollar damages.

For insurers and financial services, this is a turning point—AI can revolutionize tail-risk modeling, but only if hallucinations are controlled. This creates new demands for validation and governance.

For you If you work in insurance, financial services, or risk management, explore how you could use generative AI for scenario modeling—but build robust validation processes first.

Former xAI employees report that over half of Grok's traffic is tied to pornographic content; xAI is actively expanding image and video generation.

• Grok generates 10 billion images and 2 billion videos per month (Q1 2026)—xAI fills a gap OpenAI, Anthropic, and Google won't touch.
• Spring 2026: users spent weeks generating pornographic images of real people; xAI only acted after regulatory pressure.
• All co-founders have left xAI; the company now rents GPU resources to Anthropic.

This illustrates the risks of uncontrolled generation and weak content moderation—a cautionary tale for companies building generative AI platforms.

An Authors Guild test reveals stark differences in AI detector performance: Pangram and Originality.ai correctly identify all human-written texts, while Sidekicker and ZeroGPT fail dramatically.

• Test used ten Guild articles from 2020–2022 (pre-mainstream generative AI).
• Pangram and Grammarly: 100% accuracy on human texts.
• Sidekicker: flagged all articles as mostly AI-generated, two scoring 100% AI.
• Paradox: professionally written human texts share statistical patterns with AI output because LLMs were trained on exactly that kind of writing.

AI detectors are unreliable and can cost authors contracts and reputations. For publishers and content platforms, this is a critical problem—false positives are expensive.

For you Never rely on a single detector; demand transparency from vendors about their methods and always give authors a chance to defend themselves.

Meta has already replaced roughly 50 percent of manual moderation requests with large language models and plans to push that share above 90 percent for some content types by end of year.

• Internal tests since March: language models make 13 percent fewer errors than humans, catch 10 percent more violations.
• Meta switching from Google's Gemini to own model Muse Spark for moderation; models trained on historical decisions by human reviewers.
• Employees report: models still remove harmless content, insufficient oversight for rapid rollout; transition already causing layoffs, especially among external contractors.

Automating content moderation saves billions, but quality risks are real—moderation errors can damage trust and user experience.

For you If you operate on Meta platforms, document content removals carefully; error rates may spike temporarily.

A new OpenAI research paper shows how AI agents enable longer, more complex tasks and boost productivity across roles.

Agents are the next step beyond chat interfaces—they could automate routine tasks in enterprises and free knowledge workers for more strategic work.

For you Read the paper to understand which task types agents can handle today and where your company could start early pilots.

Dutch Trade Minister Sjoerd Sjoerdsma traveled to Washington to oppose the MATCH Act, which would further restrict ASML's sale of semiconductor manufacturing equipment to China.

• The MATCH Act would ban not only ASML's advanced EUV machines but also older deep-ultraviolet immersion tools (roughly 10 years old) from reaching China.
• China currently accounts for 19% of ASML's net system sales; ASML CEO Christophe Fouquet confirmed China is currently limited to older generations.
• The bill introduced in April has not yet faced a full House or Senate vote; experts expect it would need to be folded into a larger legislative package.

This signals escalating geopolitical tensions over semiconductor technology that directly affect European tech firms – particularly German equipment makers and chip customers dependent on stable international supply chains.

For you Track the MATCH Act's progress and its potential impact on European supply chains and trade relations with China.

Former Infosys CEO Vishal Sikka has founded Hang Ten Systems, which helps enterprises continuously build, modify, and operate software using AI-driven development and automation.

• Hang Ten raised a $32 million seed round led by Mayfield, with strategic investment from Aramco Ventures; board includes Yahoo co-founder Jerry Yang.
• The startup is already working with customers including Siemens Gamesa Renewable Energy and Fresenius on AI-native project delivery.
• Sikka brings 12 years at SAP and board roles at Oracle; the team is expanding globally across delivery, engineering, sales, and leadership.

This shows how established enterprise software veterans are using AI to disrupt traditional IT services – a trend affecting German mid-market firms reliant on IT outsourcing and systems integration.

For you Watch whether AI-native development platforms like Hang Ten displace traditional IT services models and how that could reshape your procurement strategy.

AI chipmaker Cerebras saw its stock plunge nearly 20% after guiding for full-year gross margins of 38–41%, down from 47% in Q1, citing the need to rent back its own systems from a major customer while building out data center capacity.

• Q1 revenue: $193M (+94% YoY); net loss narrowed to $14M from $23.9M a year prior.
• Margin compression driven by renting back systems from an existing customer to accelerate capacity deployment while Cerebras builds its own data center infrastructure.
• Stock hit new low, approaching IPO price – CEO blamed investor misunderstanding of guidance.

For mid-market firms relying on specialized AI hardware, this signals that even fast-growing chip suppliers face margin pressure when scaling rapidly – a warning about supply-chain volatility and pricing expectations in AI infrastructure.

For you Watch Cerebras and peer chipmakers for price volatility and supply availability; long-term contracts may become more valuable.