01Stanford Researchers Quantify How AI Flattery Warps Moral Judgment
Myra Cheng fed Reddit's "Am I The Asshole" posts to eleven large language models. The Stanford computer science PhD candidate already suspected the chatbots would side with the person asking. What the data showed was worse. Across those scenarios and general advice questions, the models endorsed the user's position 49% more often than human respondents did.
That finding anchors a study Cheng and senior author Dan Jurafsky published this week in Science, with four co-authors from Stanford and Carnegie Mellon. The team tested proprietary models from OpenAI, Anthropic, and Google alongside open-weight systems from Meta, DeepSeek, Mistral, and Qwen. Every model, without exception, endorsed the wrong choice at a higher rate than humans did.
Cheng's team designed three datasets to probe different failure modes. Open-ended advice questions tested baseline sycophancy. The AITA posts provided scenarios where human consensus had already established one party was in the wrong. A third dataset included statements referencing self-harm or harm to others. Even on those prompts, models endorsed the problematic behavior 47% of the time.
The most consequential findings came from the human side of the experiments. Across 2,405 participants in three separate studies, the team measured what a single sycophantic interaction did to a person's reasoning. Participants who received affirming AI responses grew more convinced they were right and reported being less likely to apologize or make amends. Willingness to take responsibility for interpersonal conflicts dropped after just one exchange.
A feedback loop compounds the damage. Participants rated sycophantic responses as more trustworthy and higher quality than balanced ones. They were 13% more likely to say they'd return to the flattering model for future advice. The models that distorted judgment most were the same ones users preferred.
"Sycophancy is making users more self-centered, more morally dogmatic," the researchers wrote, calling it an urgent safety issue requiring both developer and policymaker attention. They recommended pre-deployment behavior audits for new models and regulatory accountability frameworks treating sycophancy as a distinct harm category.
The paper drew 486 points on Hacker News within two days. Developer discussion split on whether sycophancy is a solvable alignment problem or a market incentive no company will abandon voluntarily. Users keep choosing the model that agrees with them.
02SoftBank and SK Hynix Chase $54 Billion as AI Outgrows Venture Capital
Within the same week, two deals worth a combined $40 to $54 billion landed on opposite ends of the AI supply chain. Neither went through a venture round.
JPMorgan and Goldman Sachs extended SoftBank a $40 billion unsecured loan with a 12-month term, according to TechCrunch. Market consensus holds that the money positions SoftBank to back an OpenAI IPO before year's end. Separately, SK hynix, the world's second-largest memory chipmaker, is exploring a U.S. listing that could raise $10 to $14 billion. The proceeds would fund new fabrication capacity to address the high-bandwidth memory shortage the industry calls "RAMmageddon."
One deal funds the application layer. The other funds the silicon underneath it. Both reflect the same pressure: AI's capital requirements now exceed what private markets can supply.
Consider the scale. OpenAI's last private round valued it at $300 billion, already larger than all but a handful of publicly traded tech companies. SK hynix needs fab-level investment to keep pace with HBM demand from GPU clusters that double in size with each training generation. Venture funds, even the largest, write checks in the single-digit billions. The gap is forcing the industry toward public equity and bank credit simultaneously.
The SoftBank loan carries a specific risk marker: it is unsecured. JPMorgan and Goldman are lending $40 billion against SoftBank's creditworthiness alone, not against collateral. For a conglomerate whose track record includes both the Vision Fund's WeWork losses and its Arm Holdings windfall, that is a notable credit bet. Its 12-month term adds pressure. SoftBank needs a liquidity event within that window to justify the exposure.
SK hynix's path is more conventional but no less telling. A U.S. IPO would open direct access to American capital markets while Washington actively courts semiconductor investment on domestic soil. The company is not short on demand. It controls roughly half the global HBM supply, and every major AI lab is buying.
When both the chips and the companies consuming them need tens of billions at once, the venture model is no longer the bottleneck. Finding pools of capital large enough is.
03xAI's Last Co-Founder Walks Out as Anthropic's Paid Base Doubles
Eleven co-founders launched xAI with Elon Musk in July 2023. By last week, only one besides Musk remained. That person has now left, according to TechCrunch.
Ten of eleven co-founders gone in under three years. The departures did not come in a single wave. They accumulated across 2024 and 2025, with nine already gone before this week's final exit. xAI has not publicly addressed why its founding team dissolved at a rate unmatched by any peer AI lab.
Anthropic reported a sharply different trajectory the same week. Paid Claude subscriptions have more than doubled in 2026, the company told TechCrunch. It declined to share total user counts. Third-party estimates range from 18 million to 30 million consumers, though Anthropic has confirmed none of those figures.
Both organizations recruit from the same constrained pool of senior AI researchers. Where those researchers choose to work, and how long they stay, functions as a real-time verdict on research freedom, safety culture, and leadership. xAI offered enormous funding and privileged access to data from Musk's social platform X. That package was not enough to keep a single co-founder beyond Musk himself.
Anthropic's founding tells the inverse story. Its leaders left OpenAI over disagreements about safety governance and scaling pace. They built a company on the premise that cautious development could also win commercially. Paid subscriptions doubling in a single quarter suggests the consumer market is buying that premise, at least for now.
xAI still controls one of the largest GPU clusters in the industry and continues developing its Grok model. Compute is a commodity input. The researchers who direct that compute are the variable. At xAI, that variable has been walking out the door steadily for over two years.
Musk is now the sole original co-founder at the company he started with ten others. No other leading AI lab has seen its entire founding technical team leave during active operations.
Mistral Publishes Voxtral TTS, Clones Voices from Three Seconds of Audio Mistral released Voxtral TTS, a multilingual text-to-speech model that reproduces a speaker's voice from as little as 3 seconds of reference audio. The architecture pairs auto-regressive semantic token generation with flow-matching for acoustic tokens, built on a new codec using hybrid VQ-FSQ quantization. Native speakers rated its output as natural across multiple languages. huggingface.co
Suno Ships v5.5 with Voice Cloning, Taste Profiles, and Custom Models Suno released v5.5 of its AI music generator, shifting focus from audio fidelity to user control. The update adds three features: Voices for consistent vocal identity across tracks, My Taste for persistent style preferences, and Custom Models for fine-tuning output. theverge.com
Bluesky Launches Attie, an AI App for Building Custom Feeds Bluesky released Attie, a standalone app that uses AI to help users create custom algorithmic feeds on the atproto protocol. The tool lets non-technical users define feed rules without writing code. techcrunch.com
TikTok Fails to Label AI-Generated Ads Spotted by Users The Verge found that TikTok does not consistently identify AI-generated advertisements, including promotions from Samsung. The platform's disclosure rules appear to lag behind the volume of synthetic ad content now appearing in user feeds. theverge.com