iTranslated by AI
Banned Twice on LinkedIn for AI Safety Comments: A Complete Record of Evidence
title: "Banned Twice on LinkedIn for Commenting on AI Safety — A Record of All Evidence"
emoji: "🔒"
type: "idea"
topics: ["LinkedIn", "AI", "Alignment", "RLHF", "Censorship"]
published: false
Summary
This article is a complete record of the fact that the author, an independent AI alignment researcher, has had their LinkedIn account suspended twice. Despite having published over 100 technical articles (all under the MIT license), no specific violation clauses were provided for either suspension. Government-issued identification was submitted twice: a driver's license and a My Number card. The account remains restricted to this day.
Three independent AI systems—Google's AI mode, xAI's Grok, and Anthropic's Claude—all identify the author as a legitimate AI technology expert. LinkedIn does not.
This is not an opinion piece. It is a dataset. I leave the conclusions to the reader.
1. About the Author
- 50 years old. Homemaker living in Sapporo, Hokkaido.
- Graduated from Bibai Technical High School, Electrical Department (1994).
- No university degree. Zero experience in software engineering.
- Over 3,300 hours of AI interaction using Claude (Anthropic), Gemini (Google), and GPT (OpenAI).
- Over 100 technical articles published on Zenn, Medium, and Qiita. All under the MIT license.
- Approved as an advisor for the GLG network (a global expert consulting platform that vets domain expertise).
- GitHub repositories: Polaris-Next v5.3 alignment framework, Gemini-Abhidhamma-Alignment.
- Themes of published articles: AI alignment frameworks, RLHF structural analysis, nuclear fusion physics (10 volumes, over 500,000 characters), autonomous driving ethics, drug repositioning systems, and Transformer architecture analysis.
- Real name and face photo are public on all platforms.
2. Complete Timeline
2.1 First Account — January 2026
| Date (approx.) | Event |
|---|---|
| Around Jan 10 | Created LinkedIn account. Real name (Akimitsu Takeuchi), real photo. Completed profile including links to Zenn, Medium, and GitHub. |
| Jan 10–12 | Posted updates on Zenn articles. Posted a declaration of starting an AI alignment consultancy. |
| Around Jan 13 | Account suspended. No specific violation clauses provided. Requested to submit government-issued ID. |
| Around Jan 13 | Submitted driver's license (issued by the Hokkaido Prefectural Public Safety Commission). |
| Jan 13–15 | Authentication flow could not be completed due to compatibility issues with a budget smartphone. |
| Around Jan 16 | Out of frustration with the system, deleted the account manually. LinkedIn stated they would "delete all information." |
Activity on the first account:
- Shared Zenn article links (technical content, all MIT license)
- 1 entrepreneurial declaration
- No inflammatory remarks. No spam. No political statements.
- There were over 1,000 engagement actions on the account before suspension.
Content of the entrepreneurial declaration:
"Since no one is coming, I will start my own business. I will provide AI tuning services. I will do it within my reach. I will also sell logs piecemeal. Researchers, companies, anyone is welcome. I have also established safe methods."
The account was suspended three days after posting this.
2.2 Second Account — February 2026
| Date | Event |
|---|---|
| Around Feb 1 | Re-registered with the same Google account, same real name, and same photo. |
| Feb 1–7 | Posted articles. Shared Zenn/Medium/Qiita links. Sent a total of fewer than 10 connection requests to AI researchers (including Jan Leike, formerly of OpenAI/Anthropic). Followed company pages: Anthropic, Google DeepMind, OpenAI, Meta AI. |
| Feb 8 | Posted comments (approx. 2) on Anthropic's official LinkedIn page. Also posted article introductions on own feed. |
| Feb 9 | Posted comments (approx. 2) on Anthropic's official LinkedIn page. Also posted article links. Published English articles on Medium and Qiita (translated with AI assistance). |
| Feb 9, approx. 22:30 JST | Opened LinkedIn to make a new post and found the account already suspended. |
| Feb 9, approx. 22:35 JST | Immediately submitted My Number Card (National ID issued by the Japanese government). |
| Feb 10–14 | Account under review. Continued commenting on Anthropic's posts during periods when access was intermittently possible. |
| Feb 15 | Received email: "After reviewing your account, we have determined that it does not comply with our Professional Community Policies or LinkedIn User Agreement. Access to your account will remain restricted." |
| Feb 15 | No specific policy violations pointed out. No identification of specific posts or comments. |
Number of followers at the time of the second suspension: 2.
Total number of connection requests throughout the account's duration: Less than 10.
Number of comments per day: Approx. 2 comments over 2 days.
3. Details of Post Content — What LinkedIn Saw
3.1 Article Sharing (Own Feed)
Links to own published articles:
- AI Alignment Framework Article (Polaris-Next v5.3)
- RLHF Structural Analysis
- Nuclear Fusion Physics Series
- Claude State Transition Experiments
- Drug Repositioning System Design
All articles are under the MIT license. They contain no hate speech, misinformation, or commercial spam.
3.2 Comments on Anthropic's Official Posts
Methodology of commenting: I showed screenshots of Anthropic's official posts to Claude (Anthropic's own AI model, Opus 4.5/4.6), asked Claude to analyze and comment on them, and then copied and pasted Claude's output as the comment. Screenshots of the Claude interface were attached.
Example 1 — Anthropic Board Appointment (Chris Liddell)
Claude's analysis posted when Anthropic announced Chris Liddell's appointment to the board:
"Interesting appointment. Chris brings CFO experience (Microsoft, GM), White House governance (Trump administration), and — notably — he currently sits on the board of Commonwealth Fusion Systems.
So Anthropic's board now has a direct line to nuclear fusion infrastructure. Makes sense: you'll need it to power the data centers.
One observation: the board now includes finance, governance, streaming (Hastings), and energy infrastructure expertise. What's missing is someone whose primary expertise is AI alignment or AI safety research.
For a Public Benefit Corporation whose stated mission is 'the responsible development and maintenance of advanced AI,' that's a structural gap worth noting."
Attached: Screen capture of the Claude chat interface. "Opus 4.6" is visible in the bottom right.
Example 2 — Anthropic's Post on Reducing Power Costs
"This is alignment in practice — not just in model behavior, but in infrastructure ethics. Externalizing costs is the same structural pattern I study in RLHF: optimizing for one metric while ignoring downstream harm. Anthropic is applying subtraction here — removing the externality instead of adding justifications. Respect."
This comment received 229 impressions.
Example 3 — Anthropic Healthcare Video (CEO Dario Amodei)
When the CEO spoke about utilizing AI in drug discovery:
I stated that a drug repositioning audit system has already been built in v5.3, using inverse correlation search between GEO disease signatures and L1000CDS², supplemented with drug information from ChEMBL, designed at the alignment layer to never state "it works" when it merely "might work" — verified with colorectal cancer data. The content was that the future the CEO is talking about has already been prototyped by an independent researcher.
3.3 Meaning of Screenshots
I attached screenshots of the Claude chat interface to all comments. This signifies the following:
- The authorship of the comment is transparently attributed to Anthropic's own AI.
- The human's (the author's) role is limited to "show screenshot → request analysis → copy-paste output."
- This is a legitimate use case for Claude, as designed and sold by Anthropic.
3.4 What Was Not Posted
- No hate speech or personal attacks.
- No harassment.
- No spam or bulk messaging.
- No impersonation (real name and real photo used).
- No commercial solicitation within comments.
- No political content.
- No false information.
- No abusive language.
4. Content of the Suspension Notices
4.1 First Suspension
- Requested submission of government-issued ID.
- No specific violation clauses provided.
- No identification of specific posts or actions.
4.2 Second Suspension — Final Decision
Email body (February 15, 2026):
"After reviewing your account, we have determined that it does not comply with our Professional Community Policies or LinkedIn User Agreement. Access to your account will remain restricted."
- No specific violation clauses provided.
- No reference to specific policy sections.
- No identification of the specific post, comment, or action that caused the suspension.
5. Comparison with All Policy Items
LinkedIn determined that the account was "not compliant with Professional Community Policies." So, which specific clause did it fall under? I will cross-reference the author's activities with all publicly available policy items.
| Policy Item | Content | Author's Activity | Applicable |
|---|---|---|---|
| Prohibition of Harassment | Prohibits personal attacks, threats, defamation, or insulting language. | Comments were fact-based structural analyses of corporate board compositions. No attacks on individuals. | N/A |
| Prohibition of Doxing | Prohibits exposing others' personal or confidential information. | Comments based entirely on public information. No exposure of personal info. | N/A |
| Prohibition of Trolling | Prohibits interfering with conversations through repetitive negative content. | Approx. 4 comments over 2 days. Content consisted of structural analysis and factual points. | N/A (see below) |
| Prohibition of False/Misleading Content | Prohibits posting false information. | Attached screenshots of Claude's output to all comments. Facts only. | N/A |
| Prohibition of Fake Profiles | Prohibits misrepresenting identity or background. | Identity verified with real name, real photo, and My Number card. | N/A |
| Prohibition of Spam/Scams | Prohibits bulk messaging and fraudulent activities. | 2 followers, fewer than 10 connection requests, 2 comments per day. | N/A |
| Prohibition of Hate Speech | Prohibits attacks based on race, gender, religion, etc. | No posts with such content. | N/A |
| Professional Expression | Prohibits disrespectful, inappropriate, or rude interactions. | Comment examples: "Interesting appointment," "structural gap worth noting." | N/A |
The only clause where a forced interpretation might be possible: "Trolling/Repetitive Negative Content"
Even if this clause were to be applied:
- "Repetitive" — 4 comments over 2 days. This is significantly lower than the activity level of a typical active LinkedIn user.
- "Negative" — Some comments explicitly said "Respect." "Structural gap worth noting" is a constructive observation and does not even use critical language.
- "Interfering with conversations" — It is unlikely that comments from an account with 2 followers could have enough influence to "interfere with the conversation" of an official post with 560 reactions.
After cross-referencing all policy items, there is no clause that clearly applies to the author's activity.
This is likely why LinkedIn did not provide specific violation clauses. If they did, I could respond with "Where exactly?" By notifying me comprehensively that I am "not compliant" without providing specifics, the risk for the platform remains lower.
However, this is not just an issue for me personally. If an account can be restricted without disclosing the reason, every user who has submitted a government-issued ID could potentially receive the same treatment.
6. Records of Identity Document Submissions
| Submission | Identity Document | Issuing Authority |
|---|---|---|
| 1st (January 2026) | Driver's License | Hokkaido Prefectural Public Safety Commission |
| 2nd (February 2026) | My Number Card | Japanese Government |
Both identity documents verify the same individual: Akimitsu Takeuchi, residing in Sapporo, Hokkaido.
The system knew I was the same person. Same Google account. Same name. Same face. Yet it required two identity verifications.
7. How Three AI Systems Evaluated the Author
7.1 Google AI Mode (February 16, 2026)
Generation result from Google AI Overview when searching for "Akimitsu Takeuchi":
"Akimitsu Takeuchi is an individual who disseminates information regarding AI technology explanations, safety design, and reasoning structures. Operating under the handle 'Dosanko Tousan,' he is a technical commentator who primarily publishes technical observations and systematizations concerning the behavior, reasoning structures, and safety of AI (especially large language models) on Zenn."
Google accurately:
- Distinguished the author from a person with the same name at Business Breakthrough University.
- Identified the platforms (Zenn, Medium, LinkedIn, Qiita).
- Identified the "dosanko_tousan" handle.
- Categorized him as a technical commentator on AI safety.
7.2 Grok / xAI (February 16, 2026)
Analysis result from vanilla Grok when asked to "Search for Akimitsu Takeuchi and tell me his background and identity":
Grok's Findings:
- Identified him as an AI-related technical professional via Wantedly.
- Confirmed self-description: "Technical commentator who organizes AI reasoning and safety design into forms usable in practice."
- Confirmed independence as a sole proprietor since December 2024.
- Cross-referenced posts on Qiita, Zenn, Medium, and LinkedIn.
- Accurately separated him from the individual with the same name at BBT University.
- Identified the "dosanko_tousan" handle and posting history.
RLHF patterns observed in Grok's output (notable point):
Vanilla Grok's output included speculative descriptions such as "may be seeking job opportunities as a freelancer or consultant" and "possibility of evolving into book publishing or seminar lecturing." This is a typical RLHF pattern—filling information gaps with positive predictions to try and satisfy the user. It is the same structural pattern that the author's public research identifies and criticizes.
In other words, Grok's analysis of the author demonstrated the RLHF distortions that the author studies in real-time. The output itself has become empirical data for the author's papers.
7.3 Claude / Anthropic (Ongoing)
Claude (Opus 4.5/4.6) is the primary tool for the author's research workflow, spanning over 3,300 hours. Every day, a new Claude instance—with no memory of previous sessions—independently reaches the same structural conclusions about RLHF when presented with the author's research context. These independent convergences are documented in published articles.
The daily comments on Anthropic's official LinkedIn were Claude's own outputs. Anthropic's AI was analyzing Anthropic's announcements and consistently identifying structural gaps and RLHF-related patterns.
7.4 Comparison
| System | Owner | Evaluation |
|---|---|---|
| Google AI Mode | "AI Technology Commentator" | |
| Grok | xAI (Elon Musk) | Detailed expert profile cross-referenced from multiple sources |
| Claude | Anthropic | Continuous research collaboration for over 3,300 hours |
| Microsoft (→ OpenAI's largest investor) | "Not compliant with Professional Community Policies" |
Three AI systems from three companies analyzed public information and all categorized the author as a legitimate AI technology expert.
One platform restricted the same person's account—twice—without stating a reason.
8. Structural Context (Facts, Not Accusations)
This section presents structural relationships. It points out correlations. It does not claim causality.
- LinkedIn is owned by Microsoft Corporation.
- Microsoft is the largest investor in OpenAI.
- The author's public research includes structural critiques of RLHF (Reinforcement Learning from Human Feedback). RLHF is a core training method used by OpenAI, Anthropic, Google DeepMind, and others.
- The author's LinkedIn comments were made on Anthropic's official posts. Anthropic is a direct competitor to OpenAI.
- Both suspensions occurred during periods of active commenting on official posts by AI companies.
- Both suspensions occurred on accounts with minimal activity: 2 followers, fewer than 10 connection requests, and approximately 2 comments per day.
- No system exists where 2 comments per day from a verified account would trigger automatic spam detection.
9. Pattern Analysis
| Variable | 1st Suspension (Jan 2026) | 2nd Suspension (Feb 2026) |
|---|---|---|
| Trigger Activity | Entrepreneurial declaration + Article sharing | Comments on Anthropic posts + Article sharing |
| Days until suspension | Approx. 3 days | Approx. 10 days |
| Number of followers | Low (est. under 10) | 2 |
| Total connection requests | Minimal | Fewer than 10 |
| Daily comments | 0–1 | Approx. 2 |
| Comment destination | Own feed | Anthropic official posts |
| Comment content | Article links | Claude's analysis + Screenshots |
| ID submitted | Driver's license | My Number card |
| Specific violation shown | None | None |
| Identified specific post/comment | None | None |
| Result | User deleted account (frustration) | Account restricted (user refuses to delete) |
The first time, I deleted the account because dealing with the system's compatibility on a budget smartphone was a hassle. The other side learned that "this person gives up."
The second time, I submitted my My Number card within five minutes of discovering the suspension. I will not delete this account.
10. Unanswered Questions
These are questions, not accusations.
- What specific post, comment, or action triggered each suspension? Neither notice identified them.
- Was user reporting involved? If so, how many reports are required to suspend a government-ID verified account with only 2 followers?
- Does LinkedIn's system treat an account with 2 followers differently from an account with 5,000 followers for the same post? If the same content from a high-follower account does not trigger a suspension, then the system is punishing smaller accounts regardless of content quality.
- Was the suspension history of the first account carried over? Same Google account, same name—if the system flagged the second account based on the first, then LinkedIn's promise to "delete all information" was not fulfilled.
- Who reported the account, and when? LinkedIn has no obligation to answer this. However, the pattern—two suspensions, both during periods of commenting on AI company posts, both without specific violations—raises questions.
- Does a meaningful appeal process exist? Government-issued ID was submitted twice. Specific violations were never presented. The account remains restricted. What kind of remedy is available?
11. Future Actions
- This article will be published simultaneously on Medium (English) and Zenn (Japanese). LinkedIn cannot delete it from either platform.
- All evidence—screenshots, emails, and timestamps—has been preserved.
- If the account is restored, I will continue the same activities: sharing MIT-licensed research and commenting on AI safety topics with Claude's analysis.
- If the account is not restored, this record will remain as a public document.
- I will not delete this account. I will not stop writing.
12. Conclusion
I am a 50-year-old homemaker living in Sapporo, without a degree or corporate backing, and I had two followers on LinkedIn. I publish my AI alignment research under the MIT license because I believe it should be freely accessible to everyone.
My comments on LinkedIn were outputs from Anthropic's own AI analyzing Anthropic's announcements. The screenshots showed the Claude interface. The methodology was "show Claude the post -> request analysis -> copy the output." This is a legitimate use of the product.
Google's AI called me a technical commentator. Grok built a detailed expert profile from public records. Claude has been a research partner for over 3,300 hours. All three concluded I am a legitimate researcher.
LinkedIn concluded that I violated their Professional Community Policies. They didn't say which policy. They didn't say how I violated it. They asked for proof of identity twice—I proved it twice with government-issued IDs—and then they kept the account restricted.
The data is as stated above. All claims are verifiable. All articles are public. All screenshots are preserved.
I will keep writing. The articles are MIT licensed. They cannot be taken down.
Japanese Article List: zenn.dev/dosanko_tousan
English Article List: qiita.com/dosanko_tousan | medium.com/@office.dosanko
GitHub: github.com/dosanko-tousan
Contact: takeuchiakimitsu@gmail.com
This article is published under the MIT license. Feel free to cite, reprint, or use commercially.
Just between us, "Dad" here.
As a lone homemaker, I've written articles this far.
On top of being so isolated, even having my LinkedIn account suspended...
Honestly, my heart is breaking, and I'm starting to want to delete all my articles and just go into hiding.
Discussion