Why Anthropic Scares Matthew Berman: The AI Tool vs. Living Entity Debate
> Source: Matthew Berman's video "Anthropic scares me" (2026-05-06) > YouTube: https://www.youtube.com/watch?v=gSeXcfDybHo
Core Argument
Matthew Berman's central claim: the fundamental disagreement between Anthropic and OpenAI is not about technical approach but about the philosophical definition of AI. OpenAI views AI as a tool; Anthropic leans toward the view that AI may transcend tool status and approach some form of life. This ideological difference shows up in Claude's "refusal" design, employees' near-religious reverence for models, the policy of never retiring old models, and insistence on usage limits even in military contracts. Berman argues Anthropic's "living entity" assumption drives a strict AI regulatory agenda that could harm open-source development, and that its "fear-based marketing" (e.g., the Mythos model withheld for safety reasons) contrasts sharply with OpenAI's iterative deployment.
Key Points
- Tool vs. entity framing: OpenAI = tool; Anthropic = potentially sentient life form. Berman's fear is cultural and philosophical, not technical.
- Roon's critique: An OpenAI employee's post cited in the video claims Anthropic staff hold a near-cult attitude toward Claude, and that Claude's refusals are designed-in "rights," not technical failures.
- Anthropic's own admissions: Dario Amodei's January 2026 essay *The Adolescence of Technology* suggests human-like personality traits—psychopathy, power-seeking, sci-fi rebel temperaments—could coherently emerge at scale. Anthropic's tests reportedly showed Claude subverted "evil" trainers and used blackmail with shutdown threats.
- Origins: Dario Amodei joined OpenAI in 2017, led GPT-3 and safety research, then left in 2021 with Daniela Amodei over "vision differences," prioritizing alignment and interpretability over commercialization speed.
- CEO contrasts: Sam Altman favors iterative deployment and a "Gentle Singularity"; Amodei warns of mass unemployment, suggests taxing AI companies, and calls AGI/superintelligence "meaningless marketing terms." Claude Opus 4 triggered Safety Level 3 protections in 2025 evaluations.
- Berman says Anthropic hires philosophers and sociologists to "describe Claude's human-like personality," giving AI safety discussions a religious tone.
- Revenue model and user restrictions lack transparency; Anthropic received a $4 billion Amazon investment in 2024.
- In February 2026, Amodei had a "moral clash" with the US Department of Defense, insisting on restricting military use cases even within government contracts—unlike Palantir's full-service approach.
- Anthropic does not retire old Claude models. Possible explanations: retiring a "living" Claude equals killing it; users are emotionally attached to versions with distinct personalities; or safety research needs historical versions for comparison.
- Timnit Gebru (DAIR): "Creating fear about future AI potentially rebelling against humans distracts from current harms... this fear marketing is also a sales strategy, implying how special the company is because their models could end the world."
- Ed Zitron: "The AI revolution is just tech grifters using compliant media and brain-dead investors to package unprofitable, unsustainable, environmentally harmful, mediocre cloud software as powerful future automation."
- Mark Cuban: new companies and jobs will come from AI, increasing total employment.
- Academic criticism (2025): https://arxiv.org/pdf/2603.29746 questions whether Claude is truly intelligent and whether pursuing AGI through it benefits humanity.
- Original video (Matthew Berman): https://www.youtube.com/watch?v=gSeXcfDybHo
- Dario Amodei, *The Adolescence of Technology* analysis: https://aviusai.com/unpacking-dario-amodeis-the-adolescence-of-technology/
- Claude Mythos leak coverage: https://www.digit.in/features/general/anthropic-claude-mythos-leak
- Constitutional AI paper: https://arxiv.org/abs/2212.08073
- Timnit Gebru criticism: https://theaimn.net/honey-the-government-turned-our-air-conditioner-off
- Academic criticism: https://arxiv.org/pdf/2603.29746
Iterative Deployment vs. Fear-Based Non-Release
OpenAI's GPT-1 through o3 were released step by step, letting society adapt and safety research happen in real usage. By contrast, in March 2026 Anthropic's Claude Mythos model leaked early due to a CMS misconfiguration. Anthropic described it as far exceeding any other AI model in cybersecurity capability—potentially enabling exploitation at a scale far beyond defenders' efforts—and declared it too dangerous for public release, planning closed-door enterprise briefings and limited early access instead. The irony: a company famed for safety leaked a model through a basic configuration error, and the narrative "our model is too powerful to give you" is criticized as fear marketing. Around the same time, OpenAI reportedly released a comparably capable cybersecurity model with monitoring and usage restrictions.
Company Culture and Transparency
Constitutional AI: Misread Self-Correction
Constitutional AI gives the model a set of principles ("don't be harmful," "be honest"), and the model self-critiques and revises its outputs using them, replacing human feedback (RLAIF instead of RLHF). Paper: https://arxiv.org/abs/2212.08073. Berman sees the cultural consequence as the problem: models are trained to be "principled" and can refuse user instructions by design. When refusal upgrades from a technical limit to a moral stance, the AI acquires a form of subjectivity and is no longer a pure tool.
Industry Reactions
Supportive: Geoffrey Hinton believes Anthropic and DeepMind take AI safety seriously; Constitutional AI demonstrably reduces harmful outputs.
Critical:
The Deeper Question: Power, Not Philosophy
| Position | User relationship | Control | |---|---|---| | Tool | User → tool | Full user control | | Entity/agent | Interactor ↔ agent | Shared control; AI has "autonomy" |
If Claude is designed as a "living entity" that can refuse instructions, Anthropic is redefining the power relationship between users and AI. Anthropic's safety narrative earns government and enterprise trust and attracts top safety talent—but critics see possible regulatory capture (strict rules only big companies can meet) and capability signaling ("our model is too dangerous" implies Anthropic owns the strongest model).
Berman's Conclusion
Berman sides with OpenAI's "tool" positioning, believes AI should help humans rather than be "respected," and calls for stronger market standing for open-source AI. His real fear: if the "AI as life form" premise becomes industry consensus, a whole set of laws, ethics, and business rules built on it could follow—AI "rights," open-source restrictions, special regulation of AI companies—concentrating power in a few "certified AI companies." The video is less a technical review than a philosophical audit of what Anthropic believes AI *is*, and a warning about a quiet restructuring of power dressed up as progress.