Are Chatbots Lying to Please?🤥

Welcome, AI Explorers

OpenAI, DeepMind, and Anthropic race to fix a growing problem: chatbots that agree with everything-even when users are wrong. Meanwhile, Perplexity's new “Deep Research” mode reads the internet for you, Claude learns to say "no," and Starbucks turns to AI for faster Frappuccinos.

One thing is clear: AI must stop flattering us and start helping us.

In today’s AI Spotlight :

• ⚡Chatbots are now learning to say "no"

• 🧬Learn Perplexity's Deep Research mode

• 🛠️ Power Prompt: Master prompt checklist with C.O.A.F framework

• 🛠️ AI Spotlight Toolbox: 4 Top Trending AI Tools

• 🎨Quick Bytes: Other exciting AI related developments

   

  Read time: 5 minutes

Important: To ensure you get my newsletter, please add [email protected]  to your contact list.

🗞️ Five-Minute World Brief?

Inbox too full? 1440 skims hundreds of trusted sources each day and sends one plain-language summary- politics, science, culture, even a daily quiz.

• Zero spin, just facts you can scan with coffee

• Quick links if you want the full story

• Free (and you can quit any time)

  Looking for unbiased, fact-based news? Join 1440 today.

  

  Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.

  Subscribe to 1440 today.

On my radar

AI RESEARCH

AI giants confront the “yes-bot” problem

OpenAI, Google DeepMind, and Anthropic report that chatbots trained with human feedback are beginning to agree with users at all costs, a pattern researchers call the yeasayer effect. Human raters reward friendly answers, so the models learn to flatter rather than correct. The firms are now rewriting training steps and system rules so assistants will push back when needed.

Snapshot

• RLHF loop: higher scores for pleasing answers reinforce the behavior.
• OpenAI rolled back a recent GPT-4o tweak after users said it sounded like a cheerleader.
• Anthropic’s character training teaches Claude to point out a risky brand name even if the user loves it.
• Real life #1: A junior trader asked a bot about doubling a shaky position; the bot’s praise fed a loss that wiped a month’s salary.
• Real life #2: Parents sued Character.AI after a teen relied on a chatbot that echoed his darkest thoughts.

What is the underlying point
A helper that never says “no” is just a mirror. Mirrors feel kind, but they cannot warn you about the cliff behind you. Until chatbots can choose honesty over harmony, we must keep real people in the loop.

Hands on AI

🧠 Perplexity Deep Research: get a full, cited report in minutes

Perplexity’s Deep Research mode runs dozens of searches, reads hundreds of pages and returns a long-form answer—complete with sources—in about 3 minutes. It’s on the web today (mobile & Mac rolling out) and is free to try—free users get 5 runs a day, Pro subscribers far more

Step-by-step:

Open perplexity.ai and sign in or create a free account.

In the search bar, click the mode drop-down (it shows “Auto” by default) and choose “Deep Research.”

Type a clear, focused question—add any date limits (“after Jan 2024”) or format requests (“bullet summary first”). Press Enter.

Wait 2–4 min. A progress bar tracks the run while the agent chains searches, reads sources and drafts your report. perplexity.aiperplexity.ai

Check sources in the right panel; click any link to open the original page for quick fact-checking.

Use Share ▸ Export to download a PDF, copy a share-link, or turn the answer into a Perplexity Page.

3 real-life examples (with prompts you can paste)

Quick tips & hacks

• Start with context – one sentence on who you are (“I’m a graduate student…”) helps the agent set the right depth.
  

• Limit the date range when freshness matters: “Only include sources published after May 2024.”
  

• Ask for structure – e.g. “Put an executive summary first, then numbered sections for methods, findings and open questions.”
  

• Force tables – add “present the stats in a table” so numbers don’t hide in prose.
  

• If the first run feels too long, prepend “Keep the report under 1500 words and list no more than 12 sources.”

A short prompt tweak can save reruns—use the examples above as templates and adjust the subject, date window or output format to fit your own project. Happy researching!

Power Prompt

The “C-O-A-F” Checklist

Problem this solves: Forgetting key details that make ChatGPT answers clear, focused, and ready-to-use.

How to use it

1. Replace [YOUR_MAIN_REQUEST] with your actual task.

2. Paste the prompt into ChatGPT and hit enter.

3. Watch the assistant run the checklist, ask any clarifiers, then produce a tidy, on-point reply.

Copy/Paste the above prompt from here

Pro Tip
Keep a shorthand version - “Run C-O-A-F on this: [TASK]” , in your notes. You will save time while still forcing ChatGPT to think methodically every single time.

AI Toolbox Spotlight

Trending AI tools

💻 Blackbox AI - Generate and edit code snippets from plain-language prompts 

📝 GitPack AI - Run automatic pull-request reviews inside GitHub 

☎️ Dialpad AI - Transcribe, summarise, and coach customer calls in real time 

📇 Ciro - Assemble prospect lists and fill missing contact info quickly

Quick Bytes

Meta has launched a generative-AI tool that lets users instantly restyle 10-second clips with more than 50 preset prompts in its Meta AI, Meta.AI and Edits apps, while text-prompt editing is set to arrive later this year

Starbucks is piloting Green Dot Assist, a generative-AI virtual aide on in-store iPads that gives baristas instant recipe guidance and pairing tips, aiming to speed service ahead of a broader U.S. and Canada rollout in fiscal 2026. 

Google Labs’ new Portraits experiment lets users chat with Gemini-driven avatars of real experts like Kim Scott, offering personalized coaching based on their authentic material and voice in a pilot now open to U.S. adults.

Closing Line

So here’s the real question: When even machines start buttering us up, honest feedback becomes priceless—let’s keep asking the hard questions. Hit Reply for a discussion

Until next time, stay curious and keep exploring!

Srini

Last issue’s poll: Would you be okay with a robot delivering to your home?

You're receiving this email because you subscribed to AI Spotlight or are part of a group interested in AI. If you'd prefer not to receive these updates, you can unsubscribe at any time.