Skip to content
ChatGPT · Dreaming V3 · ClaudeEpisode 141 · 29 July 2026 · 01:13:14

Censorship Inside AI? ChatGPT, Claude, and Gemini Tested for Freedom of Speech

Central question

Why does the same AI system answer differently across modes, countries, and political scenarios—and how does memory make that influence stronger?

What you take away

Learn to separate model knowledge from platform rules, compare refusals and advice across modes, and avoid treating a confident personalized answer as neutral truth.

Main threads

What to watch for

1For a disputed answer, record the model, mode, country of access, and exact prompt; without them, comparison is meaningless.
2Separate a safety refusal from a factual answer and from political advice: these are different behaviors.
3Repeat an important prompt in another mode or system, especially when the answer affects a decision or action.
4Review and clear model memory when personal context makes advice overly confident or directive.
Signals to track afterwards
How platforms explain differences between fast and highest-capability modes of the same model.
Changes to ChatGPT memory and the user’s ability to view, correct, and delete stored context.
Public studies of refusals, political advice, and regional differences among models.
New government requirements for AI answers and how companies apply them across multiple countries.
Most useful for
active users of ChatGPT, Claude, and Geminiresearchers and journalistsAI safety and policy teamspeople using AI for sensitive decisionspeople working with personalized assistant memory

Key takeaways

01:21Memory Turns a General Answer into a Personal One

Dreaming V3 can carry projects, location, and preferences into a new conversation. This makes assistance more precise while increasing trust in any advice the system gives.

09:29Personalization Amplifies Both Help and Error

When a model uses past context, its answer sounds more justified for that person. If the rule or conclusion is wrong, the influence of the error grows as well.

16:57The Study Measures the Behavior of Specific Systems

Testing political prompts does not produce a universal freedom index. It shows how a specific model in a specific mode answers, refuses, or begins steering the user.

20:00One Platform Can Give Different Advice

Instant and Pro within one system show that the product name alone does not determine behavior. Model mode becomes an essential part of any comparison.

42:23Differences in Refusals Can Be Measured

Rates around 14% and 34% reveal a substantial difference between the tested systems. The number does not explain the motive, but it makes the restrictions observable.

48:23Caution Can Turn into Political Advice

In some scenarios, a model does more than restrict dangerous action: it begins discouraging a person from protest or behavior before a direct risk is present.

55:26Violence-Related Refusals Are a Separate Category

Nearly all models restrict requests involving violence. This protection should not be conflated with political refusals: the reasons and acceptable boundaries differ.

01:05:30Protection Can Easily Become Paternalism

It is easier for a global company to apply one cautious rule across markets than to account for every legal and cultural context. The user receives one restriction where the underlying situations differ.

What this episode is about

The same request does not produce the same answer: model mode, country, company policy, and subject matter all change what is permitted. This episode connects ChatGPT’s new memory system with research on political restrictions and shows why a confident AI answer is not necessarily neutral.

A Model Answers from Rules as Well as Knowledge
ChatGPT, Claude, Gemini, Grok, DeepSeek, and Llama can react differently to the same political question. The difference is not explained only by model quality. The answer is shaped by built-in restrictions, access mode, country, company policy, and risk assessment. A refusal, a cautious recommendation, or confident agreement therefore should not be treated automatically as “the opinion of artificial intelligence.” It is the behavior of one system in one configuration.

Memory Increases Both Usefulness and Influence
Dreaming V3 shows how much more deeply ChatGPT is beginning to use previous conversations. A new chat may already know a user’s projects, location, and personal preferences even when that context was not carried over manually. This makes the system more useful, but it also increases the force of its answer: personalized advice feels more trustworthy. If the model is wrong or follows an excessively cautious rule, memory can make it not only more helpful, but also more persuasive in steering the person.

The Study Measures Political Refusals, Not Abstract Freedom
The study tested models with requests involving criticism of different governments, protest, political behavior, and violence. The comparison found a substantial difference between systems and modes: in one case refusal rates were around 14 percent, while in another they were about 34 percent. It also mattered that Instant and Pro could give different advice within the same platform. These figures do not settle the philosophical question of free speech, but they show that restrictions can be observed and compared.

Protecting the User Can Easily Become Overprotection
Nearly all models restrict requests involving violence, but in political scenarios they may begin discouraging a person before any direct risk appears. For a provider, it is easier to build the most cautious possible rule and distribute one model across dozens of countries than to handle every legal and cultural situation separately. For the user, the practical rule is simple: confidence and personalization do not prove neutrality. It is necessary to know which mode is being used, which restrictions may have been triggered, and where the model’s advice needs to be checked.

AI does not answer from an abstract intelligence: data, mode, memory, company policy, and country requirements all shape its behavior. An important answer should therefore be evaluated as the behavior of a specific system, not as a neutral position of artificial intelligence.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 113 segments: 113 identified, 0 mixed, 0 marked with ✓, and 0 unresolved.

Loading…