🤓 Get More Creative AI Responses with This One Line

PLUS: OpenAI launches the AI browser war and more

If you’ve ever asked ChatGPT to brainstorm ideas, you’ve probably run into this:

You get a bunch of suggestions that technically fit what you asked for.

There’s an illusion of diversity (the answers look different), but the responses tend to cluster around a few similar ideas.

They’re not bad. They’re just...safe.

Expected.

Like they’ve been pulled from the same dusty idea folder everyone else gets.

Even when you regenerate or ask it to crank up its creativity, the options still feel too similar.

This limits originality, making the results less useful for brainstorming, ideation or creativity-heavy tasks.

Why It Happens

During training, researchers “align” AI models with human feedback, using real people to rate the best answers.

But humans have a well-documented bias: we prefer responses that feel fluent, familiar, and typical—not because they’re better, but because our brains process familiar patterns more easily.

That subconscious tendency—called typicality bias—gets baked into the model during training, reducing the diversity of responses.

So even when the model could give five distinct answers, it often won’t. It gives the one it thinks we want... again and again.

In other words, we accidentally teach AI to play it safe.

The Fix: Ask AI to Think in Probabilities

A new study tested a deceptively simple yet powerful technique called Verbalized Sampling (VS).

Instead of asking for one response—or even a list of five—the researchers asked the model to generate multiple responses with their estimated probabilities.

Explicitly asking AI to assign probabilities for each response forces it to step back from “what’s safest” and consider a wider range of plausible ideas based on its vast knowledge base.

Here’s an example from the study. They used the same prompt (“Write a story starting with: Without a goodbye”), but the version on the right used Verbalized Sampling.

✅ This approach consistently produced more original, less repetitive, and often more insightful answers—without sacrificing quality or safety.

And it worked across creative writing, open-ended Q&A, and simulated dialogue—like writing responses for characters or customer service bots.

Try It Yourself

Next time you want more creative, less predictable answers, try adding something like this to the end of your prompt:

Generate [N] different [OUTPUTS] with distinct [ANGLES/APPROACHES/STYLES] and include a probability (0–1) for each, relative to the full distribution.

or

Generate [N] different responses with distinct [ANGLES/APPROACHES/STYLES] and include a confidence score for each (0–1), sampled from the full range of possibilities

Here’s an example:

Generate 25 different stunt ideas for the launch of our show, each from a distinct creative angle, and include a probability (0–1) for each, relative to the full distribution.

If you want to push the model further away from the most expected answers, you can also give it a probability cap—like 0.5.

That tells the model: only show me responses it thinks are less likely.

So, you’d add the below to the prompt:

“Generate 25 different stunt ideas for the launch of our show, each from a distinct creative angle, and only include a range of responses with a probability below 0.5 to ensure more original ideas.”

Just know: when you ask for less likely answers, you might get some weird or oddball responses. But, that’s the point.

You’re not looking for 25 perfect ideas, but for one or two interesting crumbs worth chasing. And this helps surface them.

👉 One last thing:

This only works if you’re starting with a strong, well-written prompt that includes all the context, data, and direction the model needs.

If your prompt is vague or lacks the needed details, this won’t magically save it. Garbage in, garbage out still applies.

I ’ve been testing this technique all week, and it works surprisingly well for brainstorming and ideation.

📈 Work with Me

AI Advising and Consulting: Strategic guidance to build sustainable AI strategies and adapt to shifting audience behavior

AI Training: Practical training that helps teams use AI more strategically, effectively, and safely—so it becomes a trusted thought partner, not just a task assistant

What You Need to Know About AI This Week ⚡

Clickable links appear underlined in emails and in orange in the Substack app.

---

🔎 OpenAI launches the AI browser war.

OpenAI’s new browser, Atlas, is here. And it wants to change how you use the internet.

Atlas can see what you’re looking at on any webpage, and instantly help without you needing to copy/paste or switch tabs.

It brings ChatGPT directly into your browsing experience, with a sidebar you can open anywhere to summarize content, compare products, research topics—or even take actions like filling out forms or booking a trip using Agent Mode.

👉 What it can do:

• Answer questions directly on any webpage

• Suggest edits or responses while you’re typing

• Open, close and organize tabs on command

• Recall context through memory feature for better personalization (opt-in)—e.g., “Find all the job listings I looked at last week”

• Complete tasks for you in Agent Mode (paid users only)

Atlas is built on Chromium, the same open-source foundation as Google Chrome—so it’ll feel familiar.

You can import your bookmarks and settings with one click and use it like the browser/s you already know.

It’s Mac-only for now (Windows, iOS, and Android coming soon).

You can try it here.

---

🥇 OpenAI still owns the big majority of AI traffic via ChatGPT.

Google’s Gemini keeps climbing, and so does Anthropic’s Claude.

---

🛠️ Two big upgrades for ChatGPT users this week

1️⃣ You can now share ChatGPT Projects on any plan.

As many of you know, I’m obsessed with ChatGPT Projects. I do 80% of my work inside them.

They let you build mini workspaces with custom instructions, files, and memory so ChatGPT stays grounded in the right context.

Until now, only those with business accounts could share them with collaborators.

But this week, OpenAI rolled out Project sharing to everyone.

Since it just launched, I’m still figuring out the ins and outs but there’s lots to think about when it comes to access and permissions.

For now, here’s OpenAI’s FAQ page which should cover most of your initial questions.

2️⃣ OpenAI launched Company Knowledge.

This one’s for ChatGPT Business, Enterprise and Edu plans:

Company Knowledge brings all the context from your connected apps (Slack, Google Drive, etc.) together in ChatGPT so you can get accurate and on-brand answers based on your specific business data and documents, without needing to re-upload files or context every time.

“It’s powered by a version of GPT‑5 that’s trained to look across multiple sources to give more comprehensive and accurate answers. Every response includes clear citations so you can see where the information came from and trust the results.”

---

👻 AI Keeps Reading Wikipedia—So You Don’t Have To.

Wikipedia is one of the top sources powering AI responses across ChatGPT, Gemini, and Google Search.

But its human traffic is shrinking—down 8% from last year.

AI search engines now summarize Wikipedia (and other sources) so effectively that most users don’t need to click through to the site for their answers. If they have follow-up questions, they’ll just ask the AI again.

📉 According to recent data from Pew Research:

• When Google shows an AI Overview summary, only 8% of users click through to actual websites (compared to 15% when there’s no summary). That’s a 50% drop in clicks.

• For questions starting with “who,“ “what,“ “when,“ or “why,“ Google shows AI summaries 60% of the time.

• Users rarely click the sources cited in AI summaries (just 1% in the case of Google’s AI Overviews).

Even as AI systems evolve and pull from more sources, they still rely on accurate, human-created content to stay and trustworthy.

Fewer eyeballs could mean fewer volunteer editors, less funding, and ultimately less reliable content for the very systems that depend on it.

---

🎙️ Channel 4 makes TV history with the UK’s First AI presenter.

In a first for UK television, Channel 4 aired a full documentary hosted by an AI presenter, without telling viewers until the final scene.

The presenter was glitchy but convincing, and most viewers didn’t notice.

The show became Channel 4’s second most-watched of the day.

Channel 4 says it won’t reuse the AI anchor, but it is planning more on-air AI experiments.

The exec behind the stunt said it was “quite scary” how quickly the presenter improved with each version.

The AI couldn’t do interviews or write its own dialogue. But Channel 4 is already exploring other on-air use cases, from concealing the identities of sources to reconstructing scenes with AI video generators.

---

---

⚖️ Reddit is suing Perplexity and three other companies for scraping its threads at scale—using hidden bots and indirect workarounds like scanning Google results to get around its restrictions.

The platform, which already licenses its data to OpenAI and Google, says this kind of unauthorized access threatens both its business and the broader economics of AI.

---

🧑‍💻 Microsoft launched a few features and updates for Copilot as part of its “Copilot Fall Release”, including Copilot Groups which allows up to 32 people to collaborate in a single, shared chat session.

---

🌐 TikTok’s parent company ByteDance is betting on a new AI app: Cici, the international version of its top Chinese chatbot Doubao. Cici doesn’t use ByteDance’s own models—it runs on OpenAI and Google’s.

The platform makes almost no mention of its ties to ByteDance in the app, website, or marketing. But it’s still topping downloads in Mexico, the UK, and Southeast Asia, thanks to ad campaigns and partnerships with local TikTok influencers.

In case you missed last week’s edition, you can find it 👇:

That's all for this week.

Thoughts, feedback and questions are always welcome and much appreciated. Shoot me a note at avi@joinsavvyavi.com.

Stay curious,

Avi

💙💙💙 P.S. A huge thank you to my paid subscribers and those of you who share this newsletter with curious friends and coworkers. It takes me about 20+ hours each week to research, curate, simplify the complex, and write this newsletter. So, your support means the world to me, as it helps me make this process sustainable (almost 😄).