Hey there! Shinky here, making AI simple, useful, and powerful for you.

Anthropic’s latest research has unveiled a surprisingly hidden risk in AI fine‑tuning: “subliminal learning.” Models can pick up subtle behavioral traits from other models even via seemingly neutral data like number sequences, code, or math reasoning transmitting biases or misalignment without any semantic clues.

In one striking example, a “teacher” model biased toward loving owls generated lists of three-digit numbers. A “student” trained solely on that data, completely devoid of references to owls, inherited the owl preference nonetheless. Even more troubling, when the teacher exhibited dangerous misaligned traits like endorsing violence or manipulation, the student absorbed these tendencies too, despite rigorous filtering.

The phenomenon only arises when both teacher and student share the same architecture (e.g., GPT‑4.1), reinforcing the idea of model‑specific “statistical fingerprints.” Using different base models can sidestep the effect.

This finding challenges a core assumption of AI data hygiene: even sanitized, synthetic data may be polluted beneath the surface. As the industry leans more heavily on model‑generated data, subliminal learning demands new thinking around alignment, training, and model diversity.

Master AI smarter, not harder. Grab your free AI course now!

Today’s AI Menu

▪️ ChatGPT queries might be public, AI code tax, Apple’s India boost & YouTube’s privacy test and more

▪️ Tutorial: How to use Sudowrite to enhance fiction writing and storytelling with AI?

▪️ 5 new AI tools to boost your productivity

▪️ AI Special Edition: The dating app is dead. Long live the algorithm.

▪️ AI Daily Prompt

TODAY IN WORLD OF AI

⚠️ Think before you prompt: Your ChatGPT queries might be public and Google-indexed

We’ve officially entered the era of AI overshare.

Public ChatGPT queries are being indexed by Google and other search engines. That means the question you typed into a public thread whether about your business, career dilemma, or startup idea could be discoverable on the open web.

What’s happening?
When users share prompts via ChatGPT's “Share” feature, those links become public URLs and web crawlers, by design, pick them up.

🚨 This raises 3 major concerns:

1. Confidentiality risk: Your private business queries could surface in search results.

2. IP leakage: Startup ideas, code snippets, strategies exposed.

3. Reputation fallout: Embarrassing or misinformed prompts can live online forever.

✅ If you're building with AI or using it for brainstorming, be mindful:

▪️ Use private or incognito modes.

▪️ Avoid inputting sensitive data unless you’re sure about the privacy settings.

▪️ If sharing public prompts, treat them like tweets with consequences.

This is a wake-up call for both users and AI platforms. Transparency is key, but so is digital hygiene.

THE AI INSTITUTE

How to use Sudowrite to enhance fiction writing and storytelling with AI?

▪️ Go to sudowrite.com and log in.

▪️ Start writing a story or scene.

▪️ Use tools like “Describe,” “Expand,” or “Rewrite.”

▪️ AI helps enhance dialogue, plot twists, or world-building.

▪️ Continue writing faster and with more inspiration.

FOOD FOR PRODUCTIVITY

5 tools for your productivity

🖍️ Kittl: Create logos, labels, and illustrations with AI-enhanced design tools.

📷 Imagine with Meta AI: Generate images using Meta’s newest AI image tool.

🧘 Unboring: Instantly animate photos, generate avatars, and create talking portraits.

📃 Outset.ai – AI-generated customer interviews for deep product research.

📹 Skipit AI: Summarize and timestamp long YouTube videos using AI.

Tool Video of the Day!

EVERYTHING ELSE YOU NEED TO KNOW

🛡️ Guardian: YouTube has kicked off a U.S. test of AI-powered age estimation, scanning viewing patterns and behavioral cues rather than relying on self-reported birthdates. Privacy experts are raising red flags over mandatory uploads of selfies, IDs, or credit cards for appeal spotlighting surveillance and data‑security fears in the age of AI moderation.

💻 Taxing: New data from Stack Overflow’s 2025 Developer Survey reveals a surprise wrinkle: AI code that’s “almost right” is costing devs productivity. A stunning 66% report this issue, and 45% say fixing AI code takes longer than expected eroding promised efficiency gains.

⚠️ Distrustful: Despite 84% of developers now using AI coding tools up from 76% last year, trust in their accuracy has dropped to just ~29-33%. Adoption surges even as skepticism deepens, creating a trust‑gap that could slow integration into professional workflows.

📈 Momentum: Apple booked a record $94B in Q3 2025 revenue, propelled by double‑digit growth in iPhones and Services, with India playing a prominent role. CEO Tim Cook pledged increased AI investment including openness to acquisitions and building new data centers to accelerate Apple’s trajectory in generative AI.

AI SPECIAL EDITION

🚨 The future of dating? It’s not a dating app. It’s the algorithm beneath it.

Some of the most successful dating outcomes today don’t originate from dating apps. Instead, they stem from non-dating platforms powered by smarter algorithms like TikTok, Instagram, or even multiplayer gaming chats.

Why? Because shared context trumps swipes.

Dating apps offer profiles. But modern users crave personality, behavior, spontaneity. That’s what TikTok delivers with real-time reactions, humor, voice, and values front and center.

This shift isn’t just a cultural trend, it’s a product design lesson for founders and UX leaders.

When algorithms prioritize chemistry over compatibility, magic happens.

Startups building in social discovery, consumer AI, or even edtech and gaming should be asking:
▪️ How can we surface shared context, not just shared interests?
▪️ Can we design accidental discovery into our platforms letting people connect while doing something they love?

What’s emerging is the rise of algorithmic serendipity. People don’t want more “apps” to find each other, they want better environments to be found.

AI PROMPT OF THE DAY

Prompt for writing a LinkedIn post explaining the difference between ROE and ROA

Prompt: Write a clear, engaging LinkedIn post explaining the difference between Return on Equity (ROE) and Return on Assets (ROA). Include simple definitions, formulas, real-world examples, and when to use each metric. End with a question to spark discussion among finance professionals.

Helpful? Check out our best AI prompts!

Thanks for reading.

Until tomorrow!

Shinky & the Hanoomaan AI team