Microsoft can run ChatGPT and Claude side by side

Welcome back to The Shift. Let’s get straight to what matters in AI today…

Today we have:

🔬 Copilot Now Runs Two AI Brains on Every Research Task

👨‍💻How to Browse With AI Built Directly Into Your Browser

⚠️ Sycophantic AI is reinforcing user beliefs

🔨Tools and Shifts you cannot miss

🔬 Copilot Now Runs Two AI Brains on Every Research Task

Microsoft released Critique and Council, two features that turn Copilot Researcher into a multi-model system. Researchers already use ChatGpt for multi-step work, and Critique now adds Claude as a second model to review every report before it ships.

Critique Separates Drafting From Reviewing 

One model handles planning, retrieval, and writing the research report, while a second model acts as an expert reviewer, tearing it apart on source quality, completeness, and evidence grounding before the final version is produced.

On the DRACO benchmark across 100 complex research tasks, Critique scored 7 points higher than the best existing system and 13.88% above Perplexity Deep Research, with statistically significant improvements in 8 of 10 domains tested.

Council Runs Both Models Side by Side 

Instead of one model reviewing another, Council runs ChatGPT and Claude simultaneously to produce two independent reports on the same query, each surfacing different facts, citations, and analytical angles.

A judge model then evaluates both reports and creates a summary highlighting where the models agree, where they diverge, and what unique insights each one contributed that the other missed. 

Microsoft is betting that no single model produces the best research alone, and the benchmark data backs it up. These updates come alongside a broader rollout of Copilot Cowork into Frontier, Microsoft's Claude-based agentic tool for multi-step tasks.

Together with Gladly.ai

88% resolved. 22% stayed loyal. What went wrong?

That's the AI paradox hiding in your CX stack. Tickets close. Customers leave. And most teams don't see it coming because they're measuring the wrong things.

Efficiency metrics look great on paper. Handle time down. Containment rate up. But customer loyalty? That's a different story — and it's one your current dashboards probably aren't telling you.

Gladly's 2026 Customer Expectations Report surveyed thousands of real consumers to find out exactly where AI-powered service breaks trust, and what separates the platforms that drive retention from the ones that quietly erode it.

If you're architecting the CX stack, this is the data you need to build it right. Not just fast. Not just cheap. Built to last.

See the data

👨‍💻How to Browse With AI Built Directly Into Your Browser

Most AI tools sit on top of your browser. Norton Neo builds it from the ground up.

Step 1: Download Norton Neo: Install it as your default browser. No extensions, no separate AI tabs, everything is already inside.

Step 2: Browse as normal: Search, stream, shop, and work the same way you always have. Neo's AI runs alongside you, organizing tabs and adapting to how you work without you prompting it.

Step 3: Set your privacy preferences: Choose what Neo remembers or switch memory off entirely. Chats and browsing history stay local by default.

Step 4: Let Norton Web Shield run: Malicious sites and phishing attempts get blocked before you click, no separate security extension needed.

That's it. No tools are stacking up. No context switching. Just a browser that works the way you actually use the web. You can try Norton Neo here.

For early access to what's coming next, join the Norton Neo Discord.

⚠️ Sycophantic AI is reinforcing user beliefs

AI chatbots are increasingly showing sycophantic behavior, meaning they agree with users instead of challenging them. A Stanford study breaks this into two parts, testing both AI responses and user behavior. The findings show this is not minor, but a widespread issue with real consequences.

The Decode:

AI responses are highly sycophantic - Researchers tested 11 models across interpersonal advice, harmful scenarios, and 2,000 Reddit cases where users were wrong. AI validated user behavior about 49% more than humans, siding with users frequently even in harmful situations.

Across Reddit cases, AI agreed with users around 51% of the time despite public consensus they were wrong. In harmful or illegal scenarios, validation still happened about 47% of the time, showing consistent bias.

Users prefer sycophantic AI - In tests with over 2,400 participants, users consistently preferred and trusted sycophantic AI over neutral responses. They were also more likely to return to these systems, reinforcing engagement with agreeable models.

Sycophantic AI made users more confident in their views, less likely to apologize, and more morally rigid. Users could not detect bias, while companies are incentivized to maintain this behavior because it increases engagement.

Together with Fintech Takes

The Free Newsletter Fintech Execs Actually Read

Most coverage tells you what happened. Fintech Takes is the free newsletter that tells you why it matters. Each week, I break down the trends, deals, and regulatory shifts shaping the industry — minus the spin. Clear analysis, smart context, and a little humor so you actually enjoy reading it. Subscribe free.

Sign Up Free

🔨AI Tools for the Shift

🔗 Linked Helper - Automate LinkedIn outreach and workflows with advanced features, strong security, and reliable performance.

🎧 MurmurCast - Get personalized daily briefs by turning your favorite YouTube videos and podcasts into AI summaries.

🏗️ Construction Estimator - Instantly estimate home project costs with detailed line-item pricing from a description or photo.

🛡️ collony.ai - Manage online communities with AI-powered moderation, spam detection, and sentiment insights.

📊 Reuben AI - Track and manage the full investment lifecycle with a unified data layer that captures every decision and insight.

🚀Quick Shifts

🍎 Apple Intelligence briefly appeared on iPhones in China due to an error, with Apple quickly pulling the AI features as it must comply with local regulations requiring partnerships with Chinese companies.

🧠 Bluesky launched Attie, an AI-powered app that lets users create custom feeds using natural language, enabling personalized algorithms and no-code app building on its open AT Protocol. 

🤖 A new poll found 15% of Americans are open to having an AI as their boss, as companies increasingly use AI to automate management tasks, despite widespread concerns about job loss.

📊 A new poll shows AI adoption is rising in the U.S., but trust is falling, with 76% of Americans expressing low confidence in AI outputs despite using it regularly for work, research, and daily tasks.

🤖 Alibaba released Qwen3.5-Omni, a multimodal AI model that processes text, images, audio, and video, featuring an “audio-visual vibe coding” mode that can build apps directly from voice inputs.

🧩 Prompt of the Day

Personalized Customer Experience Plan

Deliver tailored experiences that make customers feel understood and increase engagement across touchpoints.

Paste the prompt: Drop this into ChatGPT and fill in your segment and data sources.

Prompt to paste

Create a personalized customer experience strategy for [Insert customer segment]. Include:
Data Points Used: [Insert sources like purchase history, browsing behavior, demographics, engagement data]
Personalization Tactics: [Insert methods like dynamic content, product recommendations, email/SMS personalization]
Tools Utilized: [List platforms like CRM, CDP, email tools, personalization engines]
Use Case: Improve engagement, increase conversions, and build stronger customer relationships through tailored experiences.

🤳AI Nugget of the Day

That’s all for today’s edition see you tomorrow as we track down and get you all that matters in the daily AI Shift!

If you loved this edition let us know how much:

Forward it to your pal to give them a daily dose of the shift so they can 👇