On joining humans&
Author: Taylor Sorensen (Research @ Anthropic, PhD from UW) Published: January 20, 2026 Reading time: ~13 minutes
Summary
Taylor Sorensen left academia to join humans&, a startup working on human-AI interaction. He frames his decision around four claims that guide his thinking about humans + AI. Each claim has direct relevance to MSBAi’s design philosophy.
Claim 1: People are different from each other — AI should support that
People are often different from each other. That’s a feature, not a bug, of humanity. Our AI systems should be explicitly built to support that.
Current AI training assumes well-specified correct answers (math+coding focus) and is ill-equipped for the messiness of real-world collective processes. Sorensen argues AI should ingest and support diverse perspectives — acting as an uninterested third party to guide discussion, reach consensus, bridge divides, and aggregate information across thousands of people.
Connection to MSBAi:
- Design Principle 2 (Critical Engagement) builds epistemic integrity — students ask “whose knowledge is represented?” and “what perspectives are missing?” when working with AI outputs. This is Sorensen’s concern made curricular.
- Pre-AI → AI-mediated → Post-AI sequencing forces students to form independent reasoning before AI enters, preventing the homogenization Sorensen warns about.
- The Me:chine distinction (Anderson & Rainie, 2026) deliberately cultivates the “unmachinable self” — judgment, relational presence, ethical reasoning — that AI cannot replicate.
- Cognitive offloading U-curve: Zone 2 (scattered AI use) produces worse outcomes than no AI — empirical proof that unstructured AI doesn’t support diversity, it standardizes mediocrity.
Tension: Sorensen imagines AI as a collective deliberation tool across thousands of people. MSBAi’s AI use is individual (student + AI), not multi-stakeholder consensus. The collective intelligence vision lives in the reorg/SOS platform, not here. MSBAi is about individual epistemic integrity.
Claim 2: Optimize for human flourishing, not myopic metrics
Modern AI is incredibly good at optimization. If we aren’t incredibly careful about what we optimize for, we are likely to be in a bad place. We should optimize for human flourishing.
Goodhart’s law and the alignment problem: optimizing for engagement → polarization; optimizing chatbots for preferences → sycophancy. Sorensen calls for longer-term, people-centric metrics we’d choose upon reflection.
Connection to MSBAi:
- Stanford DEL Canaries data (June 2026): Automation-ratio AI use (replacing human tasks) correlates with employment contraction. Augmentation-ratio AI use (extending human judgment) does not. MSBAi’s “human-centered, augmentation-first framing” is explicitly grounded in this — “not a positioning choice, a measured labor market fact.”
- Success metrics target human flourishing outcomes: “80%+ employed in analytics roles within 6 months,” “80%+ of Practicum sponsors rate graduates AI-ready” — not vanity metrics like “AI courses offered.”
- AIAS (AI Assessment Scale, 0–4) is a constraint on AI use, not an optimization target — a guardrail, not a Goodhart-able objective.
- Oral defense is the anti-Goodhart mechanism: you can’t game a verbal defense with AI-generated polish.
Alignment: MSBAi has operationalized Sorensen’s abstract claim. “Human flourishing” = career pivoters moving from the “Reorganize” segment (jobs contracting at -4.2%/year) into “Grow with AI” segment (jobs growing as AI scales).
Claim 3: Enhance human agency, don’t displace it
AI systems should enhance and augment human agency whenever possible, instead of displacing it.
The current direction is AI doing tasks longer without human supervision (METR evaluations). Sorensen argues people get meaning from agency, and that something important is lost when we outsource thinking itself. LLMs are under-explored as Socratic questioners and ever-patient tutors that guide to understanding.
Connection to MSBAi — the deepest alignment:
- Design Principle 1 (Cognitive Friction by Design): “Students formulate hypotheses, analyze problems, or generate arguments independently before consulting AI. AI extends thinking; it doesn’t replace the productive struggle that builds understanding.”
- Neuroscience grounding: Dopamine neurons respond to prediction errors (gap between expected and actual), not rewards themselves (Schultz et al., 1997). Removing struggle removes the brain’s teaching signal. This is the empirical version of Sorensen’s intuition about meaning and agency.
- The “scenic route” — destination is the same, but the brain has time to build predictions, invest attention, and be surprised — is MSBAi’s version of Sorensen’s “20-minute voice recording vs. 2-sentence prompt.”
- Existential literacy (above L-C-E stack): capacity to understand what kind of entity AI is, recognize when AI is shaping perception and self-concept, navigate AI relationships with calibrated trust. This directly addresses Sorensen’s concern about outsourcing thinking.
- Claude Code agent design doc (in repo): “Agents must be able to work autonomously… But humans should retain control over how their goals are pursued.” Key limitation noted: “amplifies short-term capabilities but offers limited mechanisms that explicitly support long-term human improvement.” MSBAi’s entire existential literacy layer exists to solve this gap.
Tension: Whether 15 months of structured friction builds durable agency habits — or whether students revert to prompt-and-accept after graduation — is an open empirical question.
Claim 4: Creativity from combining unique human contexts with AI
Creativity often comes from people’s unique contexts and approaches. We as a society will be more creative by combining people’s strengths and contexts with AI’s strengths.
Shannon’s insight came from intersection of symbolic logic + electrical engineering. LLMs have been exposed to everything but “country of geniuses in a datacenter” hasn’t manifested. Risk: homogenizing everyone’s context by using the same AI systems in similar ways.
Connection to MSBAi:
- “Build to Learn” approach: students build AI tools (RAG pipelines, agentic workflows), not just use them. A student who builds a LangChain pipeline has a different cognitive context than one who just prompts ChatGPT.
- Portfolio-driven model (10+ portfolio pieces, GitHub): each student’s creative output is traceable to their unique path, not uniform AI output.
- Agentic AI elective + BADM 576 LLMOps: graduates can architect AI systems — this is “combining human strengths with AI strengths.”
- Metacognitive regulation (prompt logs, reflective journals) prevents the “metacognitive laziness” that unstructured AI use creates — Sorensen’s homogenization risk, named and designed against.
Tension: Sorensen says “most creative leaps will still involve humans.” MSBAi’s L-C-E progression ends at “Expertise” = “Design AI-enhanced analytics pipelines; lead adoption initiatives.” The creative leap — using AI to make a novel connection nobody else has made — isn’t explicitly in the learning outcomes. It’s implied by the portfolio and Practicum, but not assessed. The oral defense checks reasoning, not originality.
Overall Synthesis
| Sorensen’s Claim | MSBAi Implementation | Evidence Level |
|---|---|---|
| Support diverse perspectives | Epistemic integrity, Pre-AI sequencing, Me:chine distinction | Design principle (not yet measured) |
| Optimize for human flourishing | Augmentation-first → labor market outcomes, Stanford DEL data | Empirically grounded |
| Enhance human agency | Cognitive friction, oral defense, existential literacy | Design principle + neuroscience |
| Creativity from unique contexts | Build to Learn, portfolio-driven, Agentic AI + LLMOps | Structural (curriculum architecture) |
MSBAi is the most direct institutional implementation of Sorensen’s vision across the Gies repos. Where the reorg repo deliberates about AI strategy and the a2i repo questions whether a new program is viable, the msba-online repo is actually building what Sorensen advocates — with empirical evidence to back it up.
The key difference: Sorensen writes from a research lab about what AI should be. MSBAi is a curriculum about what students will do with AI as it currently is. The program doesn’t wait for better AI — it designs the human-AI interaction protocol (Pre-AI → AI → Post-AI, AIAS levels, oral defense) to make current AI safe for learning. That’s the most practical answer to Sorensen’s question: “What should we optimize for?” — MSBAi optimizes for the student who, after 15 months, can still think without AI.
See also
- AI-Native, Human-Centered Strategy — MSBAi strategic positioning
- Design Principles & Constraints — cognitive friction, critical engagement, Pre-AI sequencing
- Dive into Claude Code: Agent Design Principles — agent governance architecture
- Philosophy Eats AI — Schrage/Kiron framework (complementary philosophical lens)
- Ocasio AI Governance — attention-based view and collective intelligence