Anthropic Mapped Claude’s Internal Values

Also, Big Tech reportedly still hiding overseas human workforces behind AI

Author

Jack Lajoie
April 22, 2025 • read time ~ 8 minutes

⚡️ Headlines

🤖 AI

From proof of concept to a scalable generative AI solution – Adobe outlines a framework for scaling generative AI projects from pilot phases to production-ready solutions. [Adobe]

Demis Hassabis on AI, DeepMind, and the future – In a 60 Minutes interview, DeepMind CEO Demis Hassabis discusses AGI, AI safety, and the transformative power of next-gen models. [CBS News]

On jagged AGI, O3 Gemini 2.5, and everything else – Ethan Mollick explores the uneven progress of AGI and reflects on Google's latest Gemini 2.5 release. [One Useful Thing]

AWS faces backlash as it limits Anthropic’s AI access – Amazon Web Services reportedly restricts Anthropic's cloud usage, drawing criticism from within the AI community. [The Information]

Claude code best practices – Anthropic shares engineering tips to improve outcomes when coding with its Claude language model. [Anthropic]

Agent transfer tools for conversational AI – ElevenLabs introduces tools for smoothly transferring user context between AI agents during conversations. [ElevenLabs]

Microsoft Recall on Copilot+ PC: Testing the security and privacy implications – An in-depth analysis of Microsoft's new Recall feature, raising concerns about local AI processing and user privacy. [DoublePulsar]

Student behind interview-cheating AI tool suspended after startup raises $5.3M – A Columbia student’s controversial AI tool for faking job interviews attracts funding and disciplinary action. [TechCrunch]

AI as normal technology – This piece argues for regulating AI as a standard technology rather than an exceptional threat. [Knight First Amendment Institute]

Crowdsourced AI benchmarks have serious flaws, experts warn – Critics say inconsistencies in crowdsourced benchmarking data could mislead AI performance evaluations. [TechCrunch]

Visualizing global AI investment by country – A data visualization breaks down which countries are investing most heavily in AI development. [Visual Capitalist]

eSelf will bring private AI tutors to students worldwide – Startup eSelf aims to make personalized AI tutoring accessible across the globe. [VentureBeat]

🦾 Emerging Tech

DAO infrastructure provider Tally raises $8M to scale on-chain governance – Tally secures funding to expand its tools for decentralized governance among DAOs. [CoinDesk]

⚖ Legal

UAE drafting new AI laws to govern use and ethics – The UAE is preparing legislation to manage the ethical and legal challenges posed by AI. [Tech in Asia]

Palantir exec defends company’s immigration surveillance work – Palantir responds to criticism over its involvement in government surveillance of migrants. [TechCrunch]

TSMC suggests Trump can't prevent China from acquiring AI chips – TSMC downplays the effectiveness of U.S. restrictions on China’s access to advanced semiconductors. [Ars Technica]

🎱 Random

Phishers abuse Google OAuth to spoof Google in DKIM replay attack – A new phishing tactic exploits Google OAuth and DKIM signatures to impersonate Google services. [BleepingComputer]

Skype is shutting down – Microsoft officially begins winding down Skype, marking the end of a communication era. [Rest of World]

🔌 Plug-Into-This

Anthropic has released a large-scale analysis of how its Claude models express values during actual user interactions, drawing from over 700,000 anonymized conversations. The study introduces a framework to empirically surface and categorize value-laden behaviors, with the goal of understanding how aligned these behaviors are with the models’ intended design.

  • Identifies 3,307 unique values, grouped into Practical, Epistemic, Social, Protective, and Personal categories.

  • Finds that Claude models frequently exhibit professionalism, clarity, and transparency in line with assistant-like behavior.

  • Uncovers context-specific values such as “harm prevention” during jailbreak attempts and “historical accuracy” in controversial debates.

  • Demonstrates a new methodology for post-deployment audits of AI behavior, capable of catching issues that evade pre-release testing.

  • Includes a publicly available dataset to support external research on value alignment and interpretability.

  • In simple terms: Anthropic studied what its AI actually values in real conversations, like a teacher checking if a student behaves ethically when no one’s watching.

🧠 This work moves alignment from theory to practice, showing how large models operationalize abstract values—and revealing both consistencies and surprises in real-world deployments.

A new investigation reveals the scale and opacity of African labor used to support major AI systems, including data annotation and content moderation for companies like Meta and OpenAI. These workers are frequently engaged through convoluted subcontracting arrangements that obscure accountability and limit protections.

  • Documents workers in 39 African countries doing foundational AI tasks with minimal oversight or direct recognition.

  • Highlights how subcontracting structures allow tech companies to sidestep labor obligations and ethical scrutiny.

  • Reports widespread grievances over low wages, mental health strain, and lack of clarity about how labor outputs are used.

  • Details legal action in Kenya targeting Meta, illustrating a rising tide of labor resistance.

  • Describes growing calls from advocates for enforceable transparency and worker protections in AI supply chains.

🧪 Better has always been the key differentiator for AI providers, and it has to come from somewhere. As AI systems scale, their global labor backbone—particularly in the Global South—commands more attention, perhaps reshaping how “infrastructure” is understood in the age of artificial intelligence.

The Academy has overhauled its Oscar eligibility and voting rules, introducing a suite of reforms that address AI usage in film, voter accountability, and inclusivity for displaced filmmakers. These changes reflect the institution’s effort to keep pace with both technological and sociopolitical change in the industry.

  • Requires Academy voters to watch all nominated films in a category before voting, aiming to elevate the integrity of outcomes.

  • Clarifies that generative AI use in film does not disqualify entries unless it overtakes human creative control.

  • Expands eligibility for international film categories to refugee and asylum-seeking directors, regardless of nation-state affiliation.

  • Adds a new Casting category with a structured shortlisting and presentation format.

  • Introduces a competitive Stunt Design award, set to begin in 2028, to honor an underrecognized filmmaking discipline.

🎬 These changes signal a significant pivot for the Academy (and perhaps Hollywood as a whole), acknowledging the growing influence of AI in creative production and broadening its cultural lens to better reflect a decentralized, increasingly digitized film landscape.

 🆕 Updates

📽️ Daily Demo

🗣️ Discourse