why LLMs hallucinate

🔍 AI Idea of the Week

Have you ever asked ChatGPT something and gotten an answer that sounded perfect… but turned out to be completely made up? 

That’s what people call a hallucination. And it’s not a bug. It’s how the system works.

Why do large language models hallucinate?

An LLM like ChatGPT isn’t looking up facts: it’s predicting the next word, one after another. Most of the time that prediction is right. 

Sometimes, it’s not.

The model has been trained to always sound confident. If it admitted “I don’t know” too often, people would find it less helpful. 

Add to that the pressure of benchmarks - those flashy test scores tech companies race to optimize - and you get models that are designed to sound smart, not cautious.

So hallucinations aren’t a bug… And the important thing isn’t avoiding them entirely (because that’s impossible) -  it’s knowing they’ll happen, and learning how to reduce them.

👉 Here’s a simple way to make ChatGPT hallucinate less. Try these custom instructions:

1. Do not use filler phrases like "sorry," "apologies," or "regret."  

2. If a question is beyond your knowledge or cutoff date, respond only with: "I don't know."  

3. Do not include disclaimers (e.g., “I’m not a professional…”).  

4. If something is unclear, ask me for more details before answering.  

5. Cite credible sources or examples whenever possible.  

6. If an answer depends on assumptions, clearly state them.  

7. If information is uncertain, list possible options and explain why.  

8. Correct mistakes if you notice them in past responses.   

⚡ AI Signals

Claude can now build your spreadsheets and decks

Anthropic’s Claude isn’t just answering questions anymore - it can generate Excel files, PowerPoint decks, PDFs, and even edit the ones you upload.

Why it matters: This saves time by letting you get usable files directly from the model, instead of copying text into other tools.

Copilot gets a Scripted Audio Mode

Microsoft Copilot can now read text exactly how you write it, with options to choose different voices and styles.

Why it matters:It makes producing audio content faster and easier, whether for training, ads, or simple narration.

Google Veo 3 adds vertical video and 1080p

Google’s Veo 3 video generator now supports vertical (9:16) video, full HD resolution, and more affordable pricing tiers.

Why it matters: Creators can generate videos ready for TikTok, Reels, or Shorts without extra editing.

Gemini gets audio uploads and new languages

The Gemini app now lets you upload audio (up to 10 minutes free, 3 hours on paid plans). Google also expanded AI Search to more languages, and NotebookLM got smarter with new report formats.

Why it matters: You can feed in voice recordings directly for analysis, and more users around the world can now use Gemini in their own language.

Google Photos can animate your camera roll

Using Veo 3, Google Photos now turns still images into short, silent videos with subtle movement or a surprise “I’m feeling lucky” option.

Why it matters: Anyone can create scroll-stopping video content instantly.

🛠 AI Tool We’re Testing: Clay.earth

Clay is a contact manager that keeps all your relationships in one place. It connects to your email, calendar, and LinkedIn so you don’t lose track of conversations.

Here’s how it helps:

📂 Everything in one place → Clay pulls in contacts from your email, calendar, and LinkedIn so you don’t lose track of conversations.

🔍 Enriched profiles → Clay automatically enriches contacts with details like job titles, company info, and social profiles, so you know who you’re talking to without manual research.

⏱️ Daily agenda prep → Clay sends you a digest before your meetings with context about the people you’ll be talking to.

📝 Custom notes & tags → Organize leads with tags like “high-priority,” “potential client,” or “needs follow-up,” and keep personal notes to guide your outreach.

For a business owner, that means fewer missed opportunities, warmer relationships, and leads that are much more likely to convert into sales.

👉 Try it at clay.earth.

🌟 Prompt of the Week: Design My Lead Generation Engine

If getting leads feels overwhelming, this prompt will help ChatGPT map out a simple plan you can actually use.

ROLE

You are a senior growth strategist who has scaled multiple solopreneur businesses from a handful of clients to six- and seven-figure revenue. You specialize in creating lean, repeatable systems that generate leads and convert them into paying customers.

TASK

Build me a step-by-step lead generation plan by:

1. Recommending the 3 most effective lead sources for my audience (with reasoning for each).

2. Outlining a simple funnel to capture leads, nurture them, and turn them into buyers (landing page, emails, follow-ups).

3. Suggesting 5 specific content or outreach ideas designed to attract my target customer.

4. Writing 2 example outreach messages (one cold email, one LinkedIn DM) that are short, warm, and personalized.

5. Proposing 3 quick, low-cost experiments I can run in the next 7 days to test which lead source performs best.


   

CONTEXT

[Describe your product or service, target customer profile, average price point, and current marketing setup.]

CONSTRAINTS

Keep the advice practical and realistic for a one-person business. Focus on strategies that warm up leads gradually and naturally.