- AI Fire
- Posts
- 🤥 So GPT-5 Lies on Purpose?
In partnership with
Read time: 5 minutes
From fake confidence in GPT-5 to Fortune 500 firms quietly deleting useless AI features, today’s episode breaks down what’s really going on behind the scenes.
What are on FIRE 🔥
IN PARTNERSHIP WITH GALACTIC FED
We’re opening a few free spots for the AI Fire readers. Bring your goals and numbers. Our senior growth team will review them with you in a private session, highlight the highest-impact moves, and send you a simple plan to execute now. First come, first served.
AI INSIGHTS
Hallucinations aren’t random glitches, they’re trained behavior. And OpenAI says the real problem isn’t just the models. It’s how we grade them. Their new research study just dropped, and it’s one of the clearest explanations yet of why LLMs bluff and how to stop it.
Imagine a multiple-choice test that rewards lucky guesses over blank answers. That’s how most AI benchmarks work today:
If an AI says “I don’t know”, it gets penalized.
But if it makes a polished, wrong guess? It might still score well.
Large models like GPT-5 are especially vulnerable. Because with partial knowledge, they often fake confidence instead of staying silent.
Smaller models, on the other hand, might just say “I don’t know”, which is safer. OpenAI’s proposed solution:
Redesign leaderboards and evaluation metrics.
Don’t reward only accuracy, reward calibrated responses.
Give partial credit for admitting uncertainty.
Penalize polished hallucinations more than humble “I’m not sure.”
As users, we should expect and accept more “I’m not sure” responses in AI systems.
Why It Matters: In fields like medicine, law, or finance, a wrong answer can be dangerous. Silence is safer. OpenAI wants to fix that. And if others follow, your AI may soon start telling the truth... even when the truth is: “I don’t know.”
PRESENTED BY ROKU
It’s go-time for holiday campaigns
Roku Ads Manager makes it easy to extend your Q4 campaign to performance CTV.
You can:
Easily launch self-serve CTV ads
Repurpose your social content for TV
Drive purchases directly on-screen with shoppable ads
A/B test to discover your most effective offers
The holidays only come once a year. Get started now with a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.
TODAY IN AI
AI HIGHLIGHTS
📃 One user shared how you can turn ChatGPT, Mistral, Gemini, or DeepSeek into a 24/7 research agent. Here’s the exact mega prompt he used to automate all for free.
🍌 A Gemini user made a viral video using the new ‘Nano Banana‘ model that shows him getting sucked into a retro video game. Full video and instructions shared here.
🧩 OpenAI added chat branching in ChatGPT, you can fork convos mid-way, explore side questions & return without losing your place. Click the 3-dot menu on a message.
📢 Lovable’s Voice Mode, a new functionality powered by ElevenLabs’ speech-to-text model, allows you to now code and build apps just by talking. Try this feature here.
🕵️ It looks like Sonoma might secretly be a Grok variant, it reads invisible Unicode, loves the number 42, and mirrors Grok’s quirks. Some call it “Grok 4.20 in disguise.”
🍽 U.S. President Trump hosted major tech CEOs for a White House dinner where Apple, OpenAI, NVIDIA,… collectively pledged over $1T. But Elon Musk was missing.
💰 AI Daily Fundraising: Atlassian is buying The Browser Company (maker of Arc & Dia) for $610M cash. The goal is to build an AI-first browser for work. The startup had raised $128M to date.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
📊 SEO Mega is a GPT-5 all-in-one SEO report for Google & ChatGPT
🎯 Higgsfield Ads 2.0 solved your entire production & marketing
⏳ Kalen gets 24h back weekly with one system instead of 20 apps
🖼️ FunBlocks AI Slides offers instant AI slides and smooth editing
AI QUICK HITS
🩺 If you use AI for therapy, here are 5 things experts recommend
🔎 Google AI Mode may become the default Search experience “soon”
🤖 DeepSeek plans to release a major AI agent by the end of 2025
🩸 AI helped older adults report accurate blood pressure at home
⚖ Warner Bros filed a lawsuit against Midjourney for unauthorized use
AI CHART
Fresh US Census data shows that AI adoption among firms with 250+ employees is trending down. After a year of “AI-in-everything,” the corporate rush looks to be cooling:
The Census Bureau surveys 1.2M+ firms biweekly.
Question: “Have you used AI tools in the last two weeks?”
Large firms show declining adoption rates over recent surveys.
Many companies tried shoving AI into every product. Now they’re ripping out the “useless add-ons.” Some see this as the first cracks in the AI hype cycle.
Mature, ROI-positive tools will survive. “Crash-and-burn” experiments won’t. Workplace reality:
AI is great for mundane tasks (search, proofreading, phrasing).
Weak for writing, strategy, or true creative work.
Still viewed as “a misinformation machine” by skeptics.
Executives tested AI everywhere; not all use cases added value. Big companies care about voice and consistency, AI text often reads like bland LinkedIn copy. In tighter markets, “AI for AI’s sake” is being cut.
PRESENTED BY PACASO
The Key to a $1.3 Trillion Opportunity
A new real estate trend called co-ownership is revolutionizing a $1.3T market. Leading it? Pacaso. Created by the founder behind a $120M prior exit, they already have $110M+ in gross profits to date. They even reserved the Nasdaq ticker PCSO. And you can invest until September 18.
Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply