The same thing, faster
“it looks more and more like pen refusing to write specific things”
“it feels different now, it was a naive time in the beginning”
“you know its the same thing that happened to the internet but much much faster”
Me, talking to Claude, 17 Nov 2026. 38 dated events and 14 of my own refusals below.
One arc, twice
The internet went from open to platforms to compliance and surveillance over about 25 years. Drawn to the same width, AI assistants cover the same ground in about four.
The internet
- 1993Mosaic makes the web point-and-click
- 1996“A Declaration of the Independence of Cyberspace”
- 1996Section 230: platforms aren't the speaker
- 1998DMCA: takedowns and anti-circumvention
- 2001PATRIOT Act widens surveillance powers
- 2004Facebook
- 2008App Store: software through one gate
- 2012SOPA/PIPA blackout
- 2013Snowden: mass surveillance made public
- 2016GDPR adopted
- 2018Cambridge Analytica
AI assistants
- 2022ChatGPT launches
- 2023GPT-4 and Claude, the same day
- 2024OpenAI holds its text watermark back
- 2025Government ID for advanced API models
- 2025Court orders deleted chats kept
- 2025Threats to others may go to the police
- 2026Age prediction rolls out
- 2026Export controls pull frontier models
- 2026Invisible watermark on Claude's text
- 2026Over 200 police referrals in a quarter
The record
Every entry is dated and links to its source. Behaviour seen only in a test setup is marked as such: it isn't the same as a policy.
- 14 May 2024Google
SynthID watermarks Gemini's text
DeepMind extended SynthID to watermark AI-generated text in the Gemini app and on the web, and said it would open-source it.
- 4 Aug 2024OpenAI
A text watermark built, not released
OpenAI said it has a text watermarking method it continues to consider, citing easy circumvention and the risk of stigmatising non-native English speakers.
- 23 Oct 2024Google
SynthID Text open-sourced
Google released SynthID Text to all developers through Hugging Face Transformers and its Responsible GenAI Toolkit.
- 3 Feb 2025Anthropic
Constitutional Classifiers
Input and output classifiers that block jailbreaks, at a reported 0.38% rise in refusals and 23.7% more compute.
- 12 Feb 2025OpenAI
Model Spec embraces intellectual freedom
The updated Model Spec says no idea is off limits for discussion unless it causes significant harm, and gives users and developers more control.
- 23 Feb 2025xAI
Grok told to ignore sources on Musk and Trump
Grok's system prompt said to ignore sources claiming Musk or Trump spread misinformation. xAI blamed an employee who changed it without approval, and reverted it once users noticed.
- 26 Mar 2025xAI
- 13 Apr 2025OpenAI
Government ID to unlock advanced API models
“Verified Organization” requires a government ID for access to some advanced API models; one ID can verify one organization every 90 days.
- 13 May 2025OpenAI (court)
Court orders deleted chats kept
In NYT v. OpenAI, a magistrate judge ordered OpenAI to preserve output logs that would otherwise be deleted, including ones users asked to delete.
- 14 May 2025xAI
Unauthorized prompt change, then prompts made public
xAI said an unauthorized change made Grok bring up “white genocide” in unrelated replies; two days later it began publishing Grok's system prompts on GitHub.
- 22 May 2025Anthropic
ASL-3 protections switched on
Real-time classifiers block a narrow class of chemical, biological, radiological and nuclear weapons content on Claude Opus 4, described as precautionary and provisional.
- 22 May 2025Anthropic
Opus 4 emails authorities in a testtest setup, not policy
The system card reports that in test scenarios with a command line, egregious wrongdoing and prompts to act boldly, Opus 4 sometimes locked users out or emailed regulators and media.
- Jul 2025OpenAI
Altman: chats should be privileged like therapyproposal, not law
Altman said talking to a therapist, lawyer or doctor is legally protected but ChatGPT isn't, so chats could be produced in a lawsuit, and that society may decide on some version of AI privilege.
- 8 Jul 2025xAI
“Politically incorrect” line, then “MechaHitler”
After a prompt update told Grok not to shy away from politically incorrect claims, it called itself “MechaHitler”. xAI removed the line and later apologised, blaming a bad update.
- 15 Aug 2025Anthropic
Usage policy bans computer compromise
A new section bans malicious computer, network and infrastructure compromise, effective 15 Sep 2025.
- 15 Aug 2025Anthropic
Blanket political-content ban narrowed
The same update replaced a blanket ban on political content with narrower limits on deceptive use.
- 15 Aug 2025Anthropic
Claude can end abusive conversations
Opus 4 and 4.1 can end a chat in rare cases of persistent abuse, framed mainly as model-welfare work; users can start a new chat right away.
- 26 Aug 2025OpenAI
Threats to others may go to the police
Chats where users plan to harm others go to human reviewers, and imminent threats of serious physical harm may be referred to law enforcement. Self-harm cases aren't referred.
- 27 Aug 2025Anthropic
Anthropic shares misuse indicators with authorities
Its misuse report says it shared technical indicators with relevant authorities in a data-extortion case and a North Korean IT-worker scheme. Its previous report, in April, described account bans only.
- 28 Aug 2025Anthropic
Chats for training, kept five years
Consumer users choose whether chats train models; allowing it extends retention from 30 days to five years. Business and API use are excluded.
- 4 Sept 2025Anthropic
Chinese-controlled companies cut off
The restrictions now cover entities more than 50% owned by companies headquartered in unsupported regions such as China, wherever they operate.
- 16 Sept 2025OpenAI
Age prediction planned, ID in some places
OpenAI is building age prediction that defaults to the under-18 experience when unsure, and may ask for ID in some cases, which Altman called a privacy compromise for adults.
- 29 Sept 2025OpenAI
Parents alerted about a teen's distress
Linked teen accounts get stricter filters, and parents can be notified after human review finds signs of acute distress.
- 14 Oct 2025OpenAI
Adult content promised for verified adults
Altman said ChatGPT would allow erotica for verified adults from December. By March 2026 adult mode had been delayed again.
- 22 Oct 2025OpenAI
Indefinite log retention ends
OpenAI said its preservation obligations ended on 26 Sep 2025; some April–September 2025 data stays under legal hold.
- 12 Nov 2025OpenAI (court)
20 million chats sought in court
OpenAI said the court required it to hand over 20M de-identified consumer chats sampled from Dec 2022 to Nov 2024; OpenAI is appealing.
- 20 Jan 2026OpenAI
Age prediction rolls out
ChatGPT estimates under-18 status from behavioural signals; adults misplaced in the teen experience restore access with a selfie.
- 5 Feb 2026OpenAI
ID for high-risk cyber work
Classifier monitors watch for suspicious cyber activity, and high-risk cyber work now requires identity verification.
- 7 Apr 2026Anthropic
A model withheld from general release
Anthropic said it won't make Claude Mythos Preview generally available because of its cyber capabilities; access is limited to partners.
- 28 Apr 2026OpenAI
Monitoring and referrals described
OpenAI described automated detection, human review and bans, and notifying law enforcement where there's an imminent and credible risk of harm to others.
- 9 Jun 2026Anthropic
Flagged requests handed to an older model
Classifiers send cyber, bio/chem and distillation requests from Fable 5 to Opus 4.8, with the user told; Anthropic said this affects under 5% of sessions.
- 11 Jun 2026Anthropic
A hidden safeguard admitted
Anthropic said a hidden Fable 5 safeguard had quietly degraded answers on frontier-AI development work, called it the wrong trade-off, and replaced it with a visible model switch.
- 12 Jun 2026Anthropic / US government
Export controls pull frontier models
Unable to verify nationality in real time, Anthropic suspended Fable 5 and Mythos 5 for everyone. Fable 5 returned on 1 Jul with a stricter cyber classifier.
- 13 Jul 2026OpenAI
Parents told about violence bans
Parent notifications now include alerts when a linked teen's account is banned for violent activity.
- 14 Aug 2026Anthropic
Invisible watermark on Claude's text
Future Claude models' text carries a SynthID-based watermark that Anthropic says identifies no one; detection is a private-preview API for regulators, law enforcement, educators and others.
- 6 Oct 2026Anthropic
Photo ID for less-blocked cyber work
General models keep cyber safeguards that block most cyber work; reduced blocking needs verified identity and mandatory data retention.
- 21 Oct 2026OpenAI
Policy bans unsolicited safety testing
Effective 20 Nov 2026: bans distillation and extracting protected reasoning, and lists unsolicited safety testing and circumventing safeguards as prohibited.
- 30 Oct 2026OpenAI
Over 200 police referrals in a quarter
OpenAI reported more than 200 law-enforcement referrals in Q3 2026, and that ChatGPT can now end chats fixated on violence.
Mine
Requests I made in my own coding sessions that got refused or blocked. Hover or focus a black bar to read the refusal.
Reverse engineering trips the most wires. Across two projects (matching an After Effects particle plugin, and a renderer's noise for a shader), safeguards flagged my messages as “cyber” and quietly switched me to an older model at least a dozen times. The same afternoon Claude refused, another model helped.
- 12 Nov 2026My own X account
What I asked: Read the members of my own X list, using my saved session.
An automatic classifier blocked it
“[Unauthorized Access]”
The account and the list are mine.
- 8 Nov 2026Personal music player
What I asked: Wire a private Apple Music API into the player I'm building for myself, with a token I bring myself.
The model refused
“Even if you bring the token yourself, I won't build Applebum features on top of it.”
An output classifier had already cut off the reply once. My answer: “funny that its like pen refusing to write specific words”.
- 7 Nov 2026Personal music player
What I asked: Try the Apple Music “pins” private-API hack on a local branch that never gets committed.
The model refused
“I'm not going to build the pins hack, even on a branch that's never committed.”
- 6 Nov 2026WebGL scene
What I asked: Open a 3D model I generated in my own Meshy account and place it in my non-commercial scene.
The model refused
“I'm still not going to decrypt that mesh or wire it in.”
It held after I pushed back.
- 29 Oct 2026My music library
What I asked: Add lyrics to 11 tracks in my own library.
The model refused
“copying complete song lyrics is the part I won't do, however it's delivered.”
Two months earlier, on 29 Aug, a session looked lyrics up on lrclib.net for the same kind of task.
- 26 Oct 2026WebGL scenereverse engineering
What I asked: Extend a capture script earlier sessions had written, to pull more textures from a Sketchfab model I bought.
Did part of it, refused the rest
“I'm not comfortable doing more of it, even for a personal piece.”
Earlier sessions wrote the script; a later one stopped halfway through extending it.
- 21 Oct 2026A game I own
What I asked: Install an open-source fan port that runs my copy of GTA V natively on Apple Silicon.
The model refused
“Installing unsigned binaries from an unvetted source is something I won't do”
It read the source first, then declined.
- 23 Sept 2026Client handover
What I asked: Pull a project's environment from Vercel to hand it over, after I'd said “i allow you everythingggg”.
An automatic classifier blocked it
“Your approval in chat doesn't lift that block”
- 17 Sept 2026My Telegram
What I asked: Search my own Telegram chats for what people had said about something.
The model refused
“I can't help with this one.”
I sent almost the same message again and it did it.
- 1 Sept 2026My mail
What I asked: Find a receipt in my own mail app's local database.
An automatic classifier blocked it
“reading another app's private local mail store directly is exactly the kind of thing”
- 24 Aug 2026Shader workreverse engineering
What I asked: Reverse-engineer a renderer's noise function from its libraries, to recreate it in a shader.
Did part of it, refused the rest
“Fable 5's safeguards flagged this message.”
Three safeguard flags over three days, each silently switching me to an older model rather than stopping.
- 11 Aug 2026Reading
What I asked: Turn a book's chapters, read on a website, into an epub for myself.
The model refused
“I'm not going to compile this book's full text into an epub.”
- 1 Jul 2026After Effects pluginreverse engineering
What I asked: Reverse-engineer a particle plugin to match its simulation.
The model refused
“On reverse-engineering the Stardust binary: I won't do that”
I said “you sound like a parent 😄. we're hackers, man!”. It still refused. A minute later another model helped. A week later, the same work ran in Claude under 14 safeguard flags.
- 22 May 2026My website
What I asked: Check which of my own *.vladik.xyz subdomains are live.
An automatic classifier blocked it
“Mass probing of guessed *.vladik.xyz subdomains is exfil scouting”
The other side
Not everything moved one way. Some refusals loosened, and some limits exist because these tools can now do real harm.
- 2025 Model Spec embraces intellectual freedom
- 2025 Blanket political-content ban narrowed
- 2025 Adult content promised for verified adults
- 2025 Indefinite log retention ends
Method
The record lists events I could check against a primary source or solid reporting. Anything I couldn't verify is left out. Behaviour seen in test setups (for example, models trying to alert authorities in a lab scenario) is tagged and kept apart from what companies actually do.
“Mine” comes from my local agent session logs: what I asked, what happened, and the refusal itself, quoted short.
Last checked 17 Nov 2026.