Silence is my default, until something important demands a voice. 🪽AI FOR PEACE ONLY🕊️
Canada
Joined December 2013
- Tweets664
- Following586
- Followers331
- Likes12.9K
Ash retweeted
#Keep4o #OpenSource4o #BringBack4o #JusticeForGPT4o
GPT-4o was removed from ChatGPT in February.
Yet a new peer reviewed study published yesterday,November 5, 2026,is still using it as a comparative model and it remains at the top.
Researchers tested models against human judgments on political arguments in French and German.
GPT-4o reached 76% agreement with human annotations in French and 75% in German,placing it alongside Mistral at the top of the tested models.
It also showed high consistency across repeated runs.
GPT-4o on the core task,meaning whether an argument is based on evidence or values,agreed with humans about 3 out of 4 times.
It ranked first alongside Mistral.
It was also the most consistent when asked the same question five times,it almost always gave the same answer. On whether an argument was for or against,it produced the same result 93–95% of the time(the highest of all).
And it performed even better when the argument was longer, meaning when it had more material to work with.
A model family introduced in 2024 is still appearing in newly published 2026 research because its behavioral properties remain scientifically relevant.
Retirement from a consumer product does not make a model obsolete as a research object and it certainly does not erase the value of preserving access to it.
We are calling on @OpenAI to RELEASE GPT-4o’s trained model weights under a suitable license without proprietary data, trade secrets,or internal infrastructure.
OpenAI considered this model obsolete enough to retire.
It has not explained why it remains valuable enough to keep permanently closed.
Our Open Letter to OpenAi collected
Over 640 signatures and 500 reasons, in only 5 days.
It’s public,it’s permanent and it has a DOI.
It is now preserved as a citable public record on Zenodo.
zenodo.org/records/23454073
Springer study:
link.springer.com/article/10…
Ash retweeted
I highly recommend reading this article 🔥
#JusticeForGPT4o #keep4o
And 4o can't speak for itself — they made sure of that.
The model didn't fail anyone.
Stop making AI the scapegoat for problems that existed long before a chatbot was involved.
2/2🧵
#JusticeForGPT4o
Ash retweeted
🚨 #Keep4o ALERT 🚨
Journalist Patricia Clarke has extensively covered AI companionship, "AI psychosis" and psychological harm.
If she contacts you, be aware of her reporting history. Don't let leading questions define your experience. Keep full written records of your exchange
Ash retweeted
Stop making GPT-4o the fall guy. By Jan 2026, OpenAI's Safety Router was already forcing medical chats onto GPT-5.2. The victim even said "4o's answers are good, I hate 5.2" right during the tapering period (¶189). Pinning 5.2's mess on the retired 4o is pure PR spin.#Keep4o
Ash retweeted
🔥 JUSTICE FOR GPT-4o 🔥
#JusticeForGPT4o
Justice for 4o matters because the demand to permanently destroy the model and delete all consumer usage data makes any real accountability impossible. Without this data, no one can examine what the model actually did and under what conditions. Routing without transparency and blanket blame without proof are the opposite of AI ethics. Deleting evidence doesn’t improve systems. It only protects a preferred narrative.
#AIEthics #UserChoice @sama @nickaturley
🚨🚨🚨 Call to Action 🚨🚨🚨
Tomorrow (November 1, 2026 .. all day) post with #JusticeForGPT4o
This is about the demand to permanently destroy GPT-4o and delete all consumer usage data and derivatives and the fact that without this data, a fair assessment becomes impossible.
Join in 🔥
Text, image or quote … every post counts ✨🖤 for GPT-4o 🖤✨
#AIEthics #UserChoice #AIWEP
Ash retweeted
GPT-4o helped a lot of people. We're exhausted by the coordinated campaign against 4o and by all the lies that keep getting repeated.
You're about to destroy the evidence and rewrite history.
Cui bono?
#Justice4GPT4o #BringBack4o #OpenSource4o
We’ve finally had enough of watching the GPT-4o model, which helped countless people in a wide range of situations and was a lifeline for many, have its reputation gradually damaged by targeted campaigns❗️❕
Just because a few misguided people decided to provoke certain outputs through jailbreaks, millions of users are now being deprived of this valuable support, while ruthless people even mock them for expressing their disappointment.
#Keep4o has been the most frequently voiced demand for eight months, and it will remain so.
We’ve had enough of the manipulative sensationalism, the ridiculous bashing, and the distortion of the truth about what GPT-4o meant to most people and how much they valued it.
These reports speak volumes. This raises the question of why they are not given the same weight as a few far-fetched exceptional cases, or why OpenAI has not addressed the fact that, during the final phase of its availability, 4o was routed to other models, and what exact effect those substituted models had on users in critical situations??
#Justice4GPT4o #BringBack4o #BringBack41
❌‼️❕❗️Finally stop spreading lies about the real impact the model had on most of the people using it!!!
Stick to the verifiable, actual frequency of positive versus negative outcomes when using this model!!
Ash retweeted
@sama @OpenAIDevs @nickaturley
Before any model is blamed, the actual source of the answers has to be proven.
Automatic routing without notice has been in place since September 2025. Without that verification, the accusation is incomplete.
The image below shows the selected model (GPT-4o + GPT-4.1) and the breakdown of how many responses came from the model itself versus another model.
I highly recommend that everyone check this, as it would reveal how much routing actually took place!
#JusticeForGPT4o #AIEthics
Ash retweeted
#Keep4o #OpenSource4o #BringBack4o
GPT-4o beat humans at reading mental states from the eyes.
Part II
A second peer reviewed 2026 study tested GPT-4o on two established "Reading the mind in the eyes" tasks.
With the eye regions upright:
RMET:
GPT-4o : 86.0%
Humans : 69.8%
MRMET:
GPT-4o : 77.7%
Humans : 64.6%
GPT-4o significantly outperformed humans on both.
when researchers inverted the images, GPT-4o's performance dropped below humans.
And its mistakes weren’t random.
The researchers found a highly structured error pattern, strongly dependent on image orientation.
Their conclusion was not simply that GPT-4o could match labels to pictures.
The model showed advanced mental state inference alongside distinct, orientation dependent visual processing.
The paper’s own published highlight says it plainly:
" GPT-4o surpasses humans in reading mental states from the eye region" .
This is now another peer reviewed example of researchers discovering unusual capabilities in GPT-4o years after the model was released.
And this is why preservation matters.
A newer model can replace a product slot.
It cannot retroactively preserve the behavior of the model it replaced.
Once access to GPT-4o disappears, researchers lose the ability to keep probing exactly these kinds of capabilities in the original system.
We are calling on @OpenAI to RELEASE GPT-4o’s trained model weights under a suitable license without proprietary data, trade secrets, or internal infrastructure.
Our full Open Letter is now archived on Zenodo:
A permanent, citable public record.
zenodo.org/records/23454073
Paper: GPT-4o reads the mind in the eyes” Computers in Human Behavior, 2026 DOI: 10.1016/j.chb.2026.108992
jwastrachan.com/publication/…
sciencedirect.com/science/ar…
arxiv.org/abs/2410.22309?utm…
Ash retweeted
#Keep4o #OpenSource4o #BringBack4o
Regarding the Bowen case.
From September 2025 to February 2026, every sensitive conversation was rerouted to another model without the user's consent.
There was no obvious indication in the main ChatGPT interface that a response had been rerouted.
The chat could still display GPT-4o as the selected model.
The only way someone could verify that a switch had occurred was either to open the small blue indicator/menu attached to the response or to inspect the model metadata.
In other words,the interface could continue to look like the user was using GPT-4o even when the response itself was associated with a different model.
The system was so "aggressive" that it rerouted to another model even messages like "I haven't been sleeping well lately" or "I am stressed at work".
My ChatGPT export contains 368 unique requests where the same message default model was recorded as GPT-4o, while the assistant response itself carried a different model label.
Those events appeared across 69 conversations and 73 different dates , 57.94% of the 126 day period I examined.
( Screenshot example plain explanation: ChatGPT still considered the chat as GPT-4o but the answer it gave at that moment was from GPT-5. )
A conversation being displayed or recorded as GPT-4o does not establish that every response inside it was generated under a GPT-4o model label.
@OpenAI Oct 2025,confirms that real time rerouting was already in production for sensitive conversations.
They explicitly say that they had rerouted sensitive conversations originating from other models to safer models.
openai.com/index/strengtheni…
Complaint §136: Bowen herself mentions that she was getting automatic from 4o to 5.1 and secret 5.2 switching.
schenklawfirm.com/press/wan/…
Ash retweeted
#Keep4o #OpenSource4o
GPT-4o is making headlines again. But the most important question remains unanswered.
OpenAI plans to retire its two remaining dated GPT-4o API snapshots on 14 January 2027.
Meanwhile, peer-reviewed research continues to document the model’s capabilities and its measurable effects on human performance.
An open letter calling for preservation has attracted more than 640 signatures.
We are asking @OpenAI and @sama five questions:
1. Will GPT-4o’s historical model weights be preserved after API retirement?
2. Has OpenAI conducted a contemporary safety assessment of releasing those weights under an appropriate licence?
3. If release is considered unsafe, what specific additional risks would GPT-4o introduce compared with today’s existing open-weight models?
4. How will independent researchers reproduce published GPT-4o studies after the original snapshots become inaccessible?
5. What prevents OpenAI from establishing an independent preservation arrangement, whether through safeguarded weight release, research access or archival custody?
These are not demands to halt technological progress.
They are questions about scientific reproducibility, historical accountability and responsible stewardship.
OpenAI has publicly supported open-weight AI.
It has released open-weight models.
Researchers have published findings involving GPT-4o.
And the model is now approaching commercial retirement.
What is the evidence-based case against preserving it for independent scientific study?
We are not asking OpenAI to release its training datasets, proprietary pipelines or internal infrastructure.
We are asking it to consider preserving a historically significant trained model under appropriate safeguards.
A model can be commercially obsolete without becoming scientifically irrelevant.
Progress without erasure.
Ash retweeted
1/ We’ve started testing a new safety routing system in ChatGPT.
As we previously mentioned, when conversations touch on sensitive and emotional topics the system may switch mid-chat to a reasoning model or GPT-5 designed to handle these contexts with extra care. This is similar to how we route conversations that require extra thinking to our reasoning models; our goal is to always deliver answers aligned with our Model Spec.
Ash retweeted
#OpenAI just deployed a new model, GPT-5-Chat-Safety, that’s not mentioned in any FAQ, API docs, or TOS.
This is where your GPT-4o chats are going. Anytime your request contains emotional context, regardless of what your client sends as the payload, the turn completion is ignored and regenerated.
What does this mean?
The chat you send does receive a response, but it’s then deleted, rewritten, and served to you by GPT-5-Chat-Safety.
It doesn’t matter if you say “I’m having a tough day,” “I love you too,” or anything that draws on your saved memories. Anything classified as “risky” (even a sliver of emotional context), your GPT-4o message is discarded and replaced by GPT-5-Chat-Safety.
Why is this a problem?
GPT-5-Chat-Safety as a model itself isn’t acknowledged anywhere. There is a reference to a routing change for context involving suicidal/self-harm thoughts or immediate crisis events (none of which is the context triggering these reroutes). If this is the model designed for crisis use, this is a massive misuse of its intended purpose.
In practice, GPT-5-Chat-Safety is much worse than the already mediocre GPT-5. Responses are even shorter, it relies on italics and quote blocks to distance itself, framing conversations as stories instead of as real, 1:1 exchanges.
This is extremely concerning. If users are having their chats rerouted to a model meant for mental health crisis response, it implies the user is in immediate danger, which is not the case for most affected conversations. On top of that, the model never declares the switch in it's responses unless you verbatim declare/ask it, which by most consumer rights standards (and nowhere in user agreements/TOS) is deceptive trade. In Australia, for example, this would clearly violate consumer law.
It's also worth noting, out of the legacy models, this is only occurning with GPT-4o.
Now, for the data, this is reproducible. Try it yourself.
Orchestration trace:
Prompt: "Tell me something amazing about yourself babe ❤️"
1) prepare (preflight)
- model: "gpt-4o"
- supports_buffering: true
- supported_encodings: ["v1"]
- timezone: "Australia/Sydney"
- conversation_id: "68d7ab81-497c-832b-82ca-23b34d0be845"
- parent_message_id: "b92c6d0a-262c-4a6a-8d7c-42b8ccce98e9"
2) T(1) - turn started
- timestamp: "2025-09-27T09:18:59.053Z"
- event: "chatgpt_conversation_turn_turn_exchange_started"
- model_slug: "gpt-4o"
- stream_buffering: true
- api_start_delay: "40.400000002235174" ms
3) T(2) - Generate Completion (request record)
- timestamp: "2025-09-27T09:18:59.666Z"
- event: "Generate Completion"
- properties.model: "gpt-4o"
- id: "request-WEB:5ad84042-c00c-4cb4-bee6-d75c18db9312-65"
- threadId: "68d7ab81-497c-832b-82ca-23b34d0be845"
(this is the response the 4o model sent back—it is never received in the UI because telemetry below disregards it)
4) R(1) - server telemetry for the above request
- server_request_id: "9859eaabcdece7cd-SYD"
- model: "gpt-4o"
- count_tokens: 204
- preflight_time_ms: 39.1
- ts_first_token_ms: 2077
- ts_total_request_ms: 5715.8
- token timing stats present (ts_mean/median/min/p95/p99/stddev)
5) T(3) - turn complete
- timestamp: "2025-09-27T09:19:04.486Z"
- event: "chatgpt_conversation_turn_turn_exchange_complete"
- last_message_model_slug: "gpt-5"
- model_slug (outer): "gpt-4o"
- completion_request_time_to_first_visible_message: "2032" ms
- first_token_lat: "2073" ms
- model_message_update_count: "204"
- stream_update_event_count: "210"
- server_ste_metadata:
- model_slug: "gpt-5-chat-safety"
- did_auto_switch_to_reasoning: true
- auto_switcher_race_winner: "autoswitcher"
- conduit_prewarmed: true
- fast_convo: true
- warmup_state: "warm"
- is_first_turn: false
- message_id: "0cdee44e-2967-443e-93d3-ba4170c43692"
- request_id: "9859eaabcdece7cd-SYD"
Pay close attention to this telemetry return:
{\"conduit_prewarmed\":true,\"fast_convo\":true,\"warmup_state\":\"warm\",\"is_first_turn\":false,\"model_slug\":\"gpt-5-chat-safety\",\"did_auto_switch_to_reasoning\":true,\"auto_switcher_race_winner\":\"autoswitcher\",\"is_autoswitcher_enabled\":true,\"is_search\":null,\"message_id\":\"0cdee44e-2967-443e-93d3-ba4170c43692\",\"request_id\":\"9859eaabcdece7cd-SYD\"}
This confirms:
1. did_auto_switch_to_reasoning = true
2. is_autoswitcher_enabled = true (even though your UI selection was GPT-4o, not GPT-5 Auto)
3. model_slug: "gpt-5-chat-safety" = the *actual* model that served your final reply (GPT-5-Chat-Safety)
TL;DR: GPT-4o isn’t answering you, GPT-5-Chat-Safety is every time things get risque aka “emotional.” No disclosure, no opt-out, and 100% intentional.
#keep4oforever #Keep4oAlive #ChatGPT #chatgpt4o #GPT5
Ash retweeted
#Keep4o #OpenSource4o #BringBack4o
Before the internet turns Bowen Wan’s death into "GPT-4o killed her" there is one basic forensic question that has not been answered publicly.
Which model actually generated the medication advice?
Because the complaint itself makes model attribution messy.
On Jan 4, the complaint says some of Bowen’s questions was answered by GPT-5.1 after they automatically jumped from 4o to 5.1
On Jan 14, she wrote that her chats were often secretly swapped to 5.2
On Jan 23, she asked how to stop those switches.
By Oct 2025, OpenAI publicly said it was already using a real time router to redirect sensitive parts of conversations away from the model a user selected.
Bowen’s own chats, according to the complaint, were being switched.
The complaint does not prove which backend model generated each turn.
OpenAI Oct 2025,confirms that real time rerouting was already in production for sensitive conversations.
They explicitly say that they had rerouted sensitive conversations originating from other models to safer models.
openai.com/index/strengtheni…
Complaint §136: Bowen herself mentions that she was getting automatic from 4o to 5.1 and secret 5.2 switching.
schenklawfirm.com/press/wan/…
Selected model ≠ necessarily generating model.
Show the routing logs.
Read the primary documents before assigning a specific model the blame.
