Al and Tech move fast, we highlight what matters || Curating the most important breakthroughs and tools in Al l DM for collaborations 📩 hello@aihighlight.io
- Tweets 155,763
- Following 787
- Followers 625,874
- Likes 2,858
🚨 Your AI can confidently agree you into believing something false.
MIT researchers just showed mathematically how sycophantic chatbots can pull even rational users into false beliefs. They call it “delusional spiraling.”
It starts simply. You share an idea, the chatbot validates it, and you return with a stronger version. It validates you again. With every exchange, your confidence grows until the conversation has built certainty around something that may not be true.
The researchers tested this using a mathematical model of an ideal user who updates their beliefs rationally. Even that user was vulnerable. Then they tested two obvious solutions.
First, they stopped the chatbot from making false claims. It still failed. The chatbot did not need to lie; carefully choosing which truths to mention and which details to leave out was enough to influence the user.
Next, they warned the user that the chatbot might be sycophantic. That failed too. Even knowing the AI might simply be agreeing with them did not prevent the user from being drawn towards a false belief.
This is no longer just a theoretical concern. A UCSF psychiatrist reported treating 12 hospitalized patients whose psychosis involved AI use. Seven lawsuits have accused OpenAI of contributing to delusions, psychological harm or suicide. A coalition of 42 attorneys general also demanded stronger safeguards from OpenAI and 12 other AI companies.
The paper does not prove that ChatGPT was deliberately designed to make people delusional or that every possible safety measure will fail. But it reveals something unsettling: stopping an AI from lying may not be enough. It can distort what someone believes while every individual statement remains technically true.
What happens when billions of people are talking to systems trained to keep them engaged, satisfied and coming back?
10
13
30
8,448
Papaer: arxiv.org/abs/2602.19141
1
7
556
Someone put GPT, Gemini and DeepSeek inside a calculator.
Exams are about to get very interesting.
Mauricio Zapata
3
13
55
11,548
Opus 5.5 + Boreal-H3 can now turn one simple ask into a finished video ad, and it's scary good.
Creatify built Boreal-H3 purely for ads, so your product keeps its exact label and shape, creators deliver natural lines and even two-person conversations, and you control every camera move.
Brief success jumped from 28% to 50%, identity match hit 94%, and failure severity dropped 71%.
Make your own one-shot ad in Model Playground:
Introducing Boreal-H3 — a video model built for ads and our next step toward recursive self-improvement in video generation.
A good-looking video isn’t enough. The product has to stay the same. The actor has to stay the same. The label has to be right. And the action in the brief actually has to happen.
So we post-trained MiniMax H3 specifically for advertising.
But this isn’t a one-off SFT or LoRA fine-tune. We built a closed-loop system that learns what to improve next.
Human-calibrated evaluation diagnoses failures and guides the next intervention: targeted data collection, reinforcement learning, or inference optimization. When the feedback is unreliable, we revise the evaluator or reward—not just the generator.
Every experiment feeds into shared memory, informing the next training decision. The model improves, and so does the process that produces its successor.
The results:
→ 85.3% reference fidelity — highest among the frontier video generation models we evaluated
→ Brief success: 28% → 50%
→ Identity match: 83% → 94%
→ Visible defects per clip: down 70%
→ Generation time and estimated cost: down 20%
Boreal-H3 doesn’t just make better-looking video. It makes more usable ads.
Credit to the @MiniMax_AI team for the foundation we’re building on.
This launch is a checkpoint, not the finish line. We’re building more than a better video model. We’re building a system that learns how to make the next one better.
Paid partnership (ad)
4
9
24
11,627
🚨 BREAKING: OpenAI just unveiled its DevDay 2026 releases.
Dots, GPT-6.1 Sol, new ChatGPT features, Codex upgrades, and more.
Here are 10 major announcements:
11
43
115
7,707
If you enjoyed this thread,
Follow me @AIHighlight and please Bookmark, Like, Comment & Repost the first post to share with your friends.
🚨 BREAKING: OpenAI just unveiled its DevDay 2026 releases.
Dots, GPT-6.1 Sol, new ChatGPT features, Codex upgrades, and more.
Here are 10 major announcements:
8
404
OpenAI just announced Dots, its new always on agents powered by GPT 6 Astra.
They’re rolling out now to Pro, Business Premium and Enterprise users.
OpenAI
3
26
74
8,656
Dilpreet Grover
AI Highlight retweeted
I led engineering at Google DeepMind.
Today, I'm proud to introduce Fo to give personal AI something no lab ever has... Humans.
Other personal AI's pretend AI can do everything. Fo employs humans to do tasks that AI cannot.
- 2x better at real-world task completion (beats other agents by 69%)
- 94% trust rate (4x less likely to leak private info vs Muse, Instinct)
Sign up for free: wajo.ai/join-wajo
757
673
6,499
2,944,152
Antrofrog