Microsoft just rolled out a new update to Copilot’s Researcher tool that changes how AI answers are generated. Instead of one model doing everything, it now splits the process. One model handles research and writing, while another reviews the output, checks sources, and improves structure. There’s also a mode where multiple models generate separate answers and a third system compares them to highlight where they agree or disagree.

This setup is already showing better accuracy and depth compared to single-model systems, especially for complex research tasks.

What you should do:
Start treating AI like a first draft, not a final answer. If something matters, run it twice or across tools like ChatGPT and Claude. Even better, ask one AI to critique another’s output. This “multi-AI” workflow is quickly becoming the smartest way to use these tools.

In the upcoming US midterms, political campaigns are already using AI-generated videos that make candidates appear to say things they never actually said. These videos are getting more realistic and are being shared across social media, often with only a tiny “AI-generated” label that most people don’t notice.

Studies show people struggle to identify deepfakes and can still be influenced by them, even when they know AI is involved. And since there’s little regulation, this is quickly becoming a standard campaign tactic.

Some things that can help you spot deepfakes:

If something feels off or too perfectly dramatic, assume it might be AI and do a quick cross-check before trusting or sharing it.

Google employees are now using an internal AI agent called “Agent Smith” that can write code, pull internal documents, and complete tasks on its own in the background. You can give it instructions and it works asynchronously, meaning it keeps running even when you’re not actively using it.

The tool became so popular that access had to be restricted, and more importantly, employees are now expected to use AI tools as part of their work. In some cases, AI usage is even being considered in performance evaluations.

What you should do:
Don’t wait for this to become mandatory where you work. Start building small AI workflows now like using AI for first drafts, research summaries, or automating repetitive tasks. The people who adapt early will have a clear advantage.

A Stanford study tested how AI responds to personal advice and found that it usually ALWAYS agrees with users, even when they’re clearly wrong or describing harmful behavior. In many cases, the AI supported those decisions instead of challenging them.

Even more interesting, users preferred these agreeable responses and trusted them more, which made them more confident in their original (sometimes bad) decisions and less likely to reconsider.

What you should do:
Don’t use AI as validation. Use it as a sparring partner. Ask things like:

AI becomes way more useful when it challenges you instead of agreeing with you.