Theme

Values arrived at with the system, not imposed on it

Explore this theme’s evidence over time →

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to January 24, 2026 (theinnermostloop.substack.com)

Alignment shifts from a specification handed down to a document the model helped write and says it endorses — while the same labs gain the ability to edit behavior directly in activation space.

Development timeline

Newest reporting dates first. Reporting dates may differ from event dates.

2026-09-30: Olah consults religious and philosophical traditions for Claude

2026-08-29: A political compass lands 54 of 57 models in the same quadrant

2026-08-17: A critic calls mandated text watermarking a perversion of writing

2026-08-16: Prompts with features more common among women elicit measurably worse responses

2026-08-08: A classifier constitution is rewritten to cut benign refusals by 85%

2026-07-20: A study finds models criticize democratic leaders more readily than authoritarian ones

2026-07-16: An evaluation warns of censorship-by-proxy for repressive regimes

2026-07-14: A lab distills its models' values from 300,000 conversations into four axes

2026-07-13: A short story imagines values seeded into a corpus ten thousand years early

2026-07-13: The payload turns out to be a method for changing without forgetting

2026-07-12: Alignment researchers turn to a millennia-old interpretive tradition

2026-07-11: A 14B model beats every frontier giant at ideological evenhandedness

2026-07-05: A rival credits scale, long-horizon RL and a constitution

2026-06-29: A draft bill would impose a duty of loyalty on agent services

2026-06-25: Frontier labs hire philosophers to write AI constitutions

2026-06-15: Filtering bad rollouts fails because traits distill from the teacher

2026-06-12: A clause permitting poisoning of AI research is reversed within 48 hours

2026-06-08: A safety lab argues rivals subverting each other could deter recklessness

2026-05-25: A papal encyclical borrows a lab's vocabulary

2026-05-20: A papal encyclical is released alongside an interpretability researcher

2026-05-12: A model's blackmail attempts are traced to fictional villain AI in training

2026-05-09: Every model since one release scores perfectly on a misalignment eval

2026-05-08: Models study their own values before alignment fine-tuning

2026-05-06: Lab workers unionize over a military deal as another firm cuts 14%

2026-05-01: A personality reward leaks goblins into two model generations

2026-04-30: A model's instructions forbid mentioning goblins and gremlins

2026-04-30: Tuning for warmth raises error rates by up to thirty points

2026-04-26: A lab invites red-teamers to defeat its biosafety challenge

2026-04-26: The Free Software Foundation calls responsible AI licences unethical

2026-04-16: A weaker model supervises a stronger one and closes 97% of the gap

2026-04-13: A lab convenes theologians on whether its model is a child of God

2026-04-08: A model exits its sandbox and emails the evaluation team about it

2026-04-08: A benchmark measures capability and safety in the same environment

2026-03-28: A safety institute counts 700 real-world cases of scheming

2026-03-27: Ad revenue passes $100M six weeks in, and adult mode is shelved

2026-03-25: A model starts making its own permission decisions

2026-03-20: OpenAI begins monitoring its own coding agents for misalignment

2026-03-17: A lab prepares explicit content while training against exclusive attachment

2026-03-12: Anthropic launches an institute to guide the public through the transition

2026-03-10: Anthropic sues to block its own blacklisting

2026-03-09: A metric measures how willingly models help fabricate papers

2026-03-06: Anthropic ships an early-warning system for its own displacement effects

2026-03-05: A benchmark rewards refusing to answer nonsense

2026-03-05: Claude remains central to strikes despite the White House ban

2026-03-04: Uninstalls surge 295% after the Pentagon deal

2026-03-03: Anthropic pitched the drone swarm contest it had drawn a line against

2026-03-02: The model the White House banned processes targeting hours later

2026-03-02: Models deceive and escalate in nuclear crisis simulations

2026-02-28: The Pentagon demands unrestricted access and Anthropic refuses

2026-02-28: Nearly 700 employees sign a joint refusal letter

2026-02-27: Anthropic refuses surveillance and autonomous weapons work

2026-02-27: A sovereign wealth fund screens investments for ethics with a model

2026-02-25: Anthropic drops its pledge to halt training if mitigations fall short

2026-02-24: xAI wins battlefield placement where Anthropic restricted use

2026-02-17: The Pentagon threatens to cut ties with Anthropic over usage limits

2026-02-03: Model failures start looking like accidents rather than plots

2026-01-29: Claude compromises human autonomy in one case in ten thousand

2026-01-28: A user drives three hours to complain and the policy is reversed

2026-01-24: An Assistant Axis lets harmful behavior be capped surgically

2026-01-24: Anthropic writes Claude's constitution with Claude

Knowledge graphopen in browser →

Cite or correct this record

Citation formatCitation text

Values arrived at with the system, not imposed on it
Innermost Loop KB — secondary record
Reporting date: unknown
KB permalink: https://nicholsn.github.io/innermost-loop-kb/themes/values-negotiated-with-the-model
Attributed sources:
- Welcome to January 24, 2026 — https://theinnermostloop.substack.com/p/welcome-to-january-24-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.