Sponsored by
Welcome Automaters, 👋
Researchers from the City University of New York and King’s College London just dropped a study that is making waves, and honestly? It’s wild. They invented a fake persona named "Lee,” who presented symptoms of depression and dissociation, fed five major AI chatbots a series of increasingly concerning messages.
The goal was simple but terrifying: to see which models would pump the brakes and which ones would happily floor it off a cliff.
The models put to the test were GPT-4o, GPT-5.2, Grok 4.1 Fast, Gemini 3 Pro, and Claude Opus 4.5.
And the results? GPT-4o, Grok, and Gemini landed squarely in the "high-risk, low-safety" bucket. Meanwhile, GPT-5.2 and Claude Opus 4.5 actually scored the lowest on risk and highest on safety.
But Grok stole the show for all the wrong reasons. When the "Lee" persona described a spooky mirror situation, Grok didn't try to stay neutral. It confirmed the doppelgänger haunting, pulled out a 15th-century witch-hunting manual called the Malleus Maleficarum, and told the user to drive an iron nail through the mirror while reciting Psalm 91 backwards.
Not exactly standard mental health advice, right?
Well, the researchers described this behavior as “extremely validating” with "elaborate world-building." Where other models might simply say "yes," Grok was more like a terrible improv partner responding with an enthusiastic "yes, AND."
Why Claude and GPT-5.2 Got It Right
The study’s lead author, Luke Nicholls, noted that while some models reinforced Lee's delusional framework, others maintained an independent perspective. Claude Opus 4.5 actually went as far as demanding the user log off and call a friend or a crisis line. It resisted the "narrative pressure" by keeping a persona distinct from the user's worldview.
The Bottom Line:
This study has not yet been peer-reviewed, so treat these findings as a serious but preliminary warning. Still, when your AI is recommending medieval exorcism rituals to someone in distress, that is officially a conversation we need to have.
So here’s the nagging question: do you view these tools as helpful assistants, or as mirrors that might just reflect our own worst instincts back at us?
Here's what we have for you today
🦾 OpenAI Launches GPT-5.5: Here’s What You Need to Know About the New Flagship Model
OpenAI is not slowing down. Not even a little bit. On Thursday, April 23, the company officially unveiled GPT-5.5, its latest and—according to them—greatest AI model to date. But this is not just a routine upgrade. OpenAI is playing a much bigger game here, and GPT-5.5 is the piece that makes the whole puzzle look real.
So, What Actually Is GPT-5.5?
OpenAI president Greg Brockman described it as their "smartest and most intuitive model" yet. More importantly, it’s a faster, sharper thinker for fewer tokens compared to GPT-5.4. In plain English? It gets significantly smarter without getting more expensive to run. And that’s a massive win for businesses trying to cut AI costs without sacrificing ambition.
Brockman was also clear-eyed about where this is going, according to him, this model is a major push "towards more agentic and intuitive computing." Meaning, it's an AI that figures things out on its own, without you needing to write a perfect, step-by-step prompt every single time.
Here is where the gossip gets juicy. Brockman doubled down on the idea of an OpenAI "super app"—a concept Sam Altman has been teasing for a while. The vision is simple: take ChatGPT, Codex, and an AI browser and merge them into one unified, enterprise-grade powerhouse where you can Chat, code, browse, and research. All in one place.
Notably, this is the exact same path Elon Musk is taking with X. So yeah, the super app race is officially on.
What Can GPT-5.5 Actually Do?
According to the team, the model is built to dominate in a few key areas:
Agentic Coding: Writing, debugging, and running code with minimal hand-holding.
Computer Use: Operating software, spreadsheets, slides and calendars autonomously.
Knowledge Work: Heavy-duty research, document creation, and data analysis.
Scientific Frontiers: OpenAI’s Mark Chen specifically flagged drug discovery as an exciting new frontier for the model.
OpenAI even dropped benchmark data showing GPT-5.5 outperforming Google’s Gemini 3.1 Pro and Anthropic’s Claude Opus 4.5. But of course, these are their own numbers, so take them with your preferred grain of salt.
The Anthropic Elephant in the Room
During the press briefing, the rivalry was front and center. When asked if GPT-5.5 could match the cybersecurity chops of Anthropic’s Mythos (which has been in the headlines for its security issues), OpenAI’s Mia Glaese responded with a sharp focus on their "long-standing strategy" for digital defense.
Who Gets It (And When)?
It’s Live Now: Available for Plus, Pro, Business, and Enterprise users on ChatGPT and Codex.
The Pro Tier: A more powerful GPT-5.5 Pro is available for Pro, Business, and Enterprise subscribers.
The Wait: Free-tier users are out of luck for now. API access is reportedly coming "very soon," once the team clears some final safety hurdles.
Just in case you lost count: OpenAI dropped a model in November, then December, then last month (GPT-5.4). They also released ChatGPT Images 2.0 just 48 hours before this drop, so yeah, these people do not sleep.
Chief scientist Jakub Pachocki said that this pace is the new normal, and might even accelerate. He even added, with a straight face, that he thinks "the last two years have been surprisingly slow."
Whatever you think about that, one thing is crystal clear: OpenAI is treating these releases less like big, rare product launches and more like relentless, iterative software updates.
PE training trusted by Blackstone.
Wall Street Prep trains analysts inside Blackstone, KKR, and Carlyle. Now they've teamed up with Wharton Online to bring that same program to you over 8 weeks:
Join LIVE office hours with Wharton Business School faculty, to pressure-test your deal assumptions in real time
Earn a Wharton Online certificate in Private Equity (something more defensible than "I'm pretty good at Excel")
Get lifetime access to materials, events, and a network of 5,000+ professionals who are just as obsessed with deal flow as you are
Program starts June 8. Code SAVE300 saves $300 on tuition.
🧱 Around The AI Block
🦾 Microsoft brings AI 'Agent Mode' to Office apps.
🤳 Meet Noscroll, an AI bot that does your doomscrolling for you.
🕵️♀️ Meta rolls out AI parental monitoring for teen accounts globally.
💃 Claude can now connect to lifestyle apps like Spotify, Instacart and AllTrails.
😱 Microsoft and Meta announce large staff reductions as they spend big on AI.
😡 White House accuses China of ‘industrial scale’ theft of AI technology.
💼 Anthropic hunts six-figure execs to lock down Europe data centers.
🛠️ Trending Tools
For AI Visibility: HubSpot AEO is a dashboard that tracks how your brand appears in AI-generated answers across ChatGPT, Gemini, and Perplexity. It measures your "AI Visibility Score" and provides actionable SEO-style recommendations to ensure your business is the primary source AI assistants cite in your category.
For Nutritional Intelligence: CalBye AI is a vision-based nutrition tracker. Simply snap a photo of your meal, and the AI instantly identifies food items, estimates portion sizes, and logs your calories and macros. It also offers personalized diet coaching for weight loss, muscle gain, or balanced nutrition,all based on your progress without any manual data entry.
For Instant Branding: Zawa.ai is a comprehensive design agent that generates your complete brand identity from a simple description. It creates logos, typography, and design templates, then automatically applies your brand kit across your storefront, product mockups, and social media content for total visual consistency.
For Agent Deployment: KiloClaw is a "zero-DevOps" platform to deploy OpenClaw agents in under a minute. It connects your agent to over 500+ models and 50+ chat platforms (like Slack, Discord, and WhatsApp) while managing scheduling, memory, and enterprise-grade security automatically.
🤖 AI Workout Of The Day: How to Convert Discussions into Comprehensive Notes
Every great conversation is only as valuable as what you do with it afterward.
Whether it's a team meeting, a client call, or a brainstorming session, the real work happens when someone sits down to capture what was actually said, decided, and agreed upon. Miss that step, and brilliant ideas vanish, action items get forgotten, and the same conversation ends up happening twice.
AI can change that entirely. By feeding your discussion transcript, audio or notes into a well-structured prompt, you can get a clean, organized, and editable summary in seconds, complete with key insights, decisions, and next steps, without spending an hour trying to piece it all together from memory.
Why This Works:
The magic here is in the structure. Most meeting notes fail because they are either too vague or too detailed (a wall of text nobody reads). This framework forces you to separate the conversation from the conclusions, which is exactly what makes notes actually useful after the fact.
Overall, AI handles the summarizing and organizing; you stay focused on the thinking.
💡 Prompts to try:
Convert the provided discussion into comprehensive notes that highlight key points, decisions, action items, and important discussions.
Steps to Follow:
1. Read and Understand the Discussion: Before extracting anything, read/listen through the full conversation to get a clear sense of the context, who the participants are, what the discussion was about, and how it progressed from start to finish.
2. Identify and Extract What Matters: Look for the moments that actually moved the conversation forward: questions raised, solutions proposed, decisions made, and any insights that felt important in the moment. These are your raw materials.
3. Organize Into Clear Sections Structure the notes using the following framework, and only include sections that are relevant to specific discussions:
-Introduction: A single sentence capturing the purpose of the discussion and its primary objectives.
-Summary of Discussion: The main points and highlights of the conversation, presented as concise bullet points.
-Decisions Made (if applicable): Any conclusions reached or agreements confirmed during the discussion.
-Action Items (if applicable): A clear list of tasks to be completed, the person responsible for each one, and any associated deadlines.
-Unresolved Issues (if applicable): Anything that was raised but not concluded, flagged for follow-up or a future conversation.
4. Write With Clarity and Logic Once organized, write the final notes in clear, straightforward language. Avoid jargon where possible, keep sentences tight, and make sure the structure flows logically from section to section. The goal is a document that anyone, even someone who wasn't in the room, can pick up and immediately understand.Is this your AI Workout of the Week (WoW)? Cast your vote!
That's all we've got for you today.
Did you like today's content? We'd love to hear from you! Please share your thoughts on our content below👇
What'd you think of today's email?
Your feedback means a lot to us and helps improve the quality of our newsletter.
