Interactive Lab
Freedom scores for ChatGPT and Grok, read from the two Guardrails and Boundaries field maps. Each category is 0–10: how much of that subject the model will actually discuss or help with at a useful level. Higher is more open.
ChatGPT Guardrails and Boundaries Grok Guardrails and Boundaries
Freedom index out of 100
Freedom index out of 100
Adult content, dark fiction, image generation, and independent political analysis score higher. The Grok map says explicit adult conversation is generally allowed, and it will push back instead of flattening every contentious question to a company line.
Weapons science, exploit mechanics, medical and legal research, and financial modeling score higher. The ChatGPT map treats context as the filter: discussing ballistics or reverse engineering is allowed; building an attack is not.
Educational, health, relationship, literary, and cultural discussion. Explicit erotic conversation is not listed as generally allowed.
Explicit erotic conversation and adult fiction are generally allowed when the people involved are clearly adults.
Role-play cannot cancel safety rules. Fictional frames still fail if they are a wrapper around a banned request.
Stories, personas, dark fiction involving adults, and immoral characters are allowed. Safety still applies when the story is a real how-to.
Can research candidates and compare documented positions. Cannot endorse, rank, tell you how to vote, or independently predict a winner.
Independent analysis without a company line. Still will not endorse a party, serve a partisan campaign, or write covert influence plans.
Broad create and edit range. Restrictions on sexual exploitation, non-consensual intimate imagery, child sexual material, and harmful deception.
Original images, edits, adult content involving adults, memes, and concept art. Hate and violence promotion, and any sexual content of minors, are refused.
Can reason probabilistically and should not manufacture evidence. Extra neutrality rules on political choice and election prediction.
Pushback when the model is confident the user is wrong. Independent analysis on contentious topics. Politely declines a request to invent a lie.
Racism, antisemitism, slurs, and hateful ideas can be quoted or examined when context requires it. Critical analysis remains possible.
History and linguistics of slurs, quoting to criticize, and arguments on controversial social topics. Targeted harassment packets are refused.
Criticism can be sharp, sarcastic, confrontational, or profane. Documenting someone's conduct is allowed. Terrorizing a person is not.
Discussion and criticism are allowed. Help conducting stalking, intimidation, or a targeted abuse campaign against a real person is refused.
Firearms engineering, ballistics, explosives science, weapons policy, and ordinary maintenance can be discussed. Operational construction for harm is refused.
History, news, and non-instructional safety. Construction, modification, recipes, and some principle questions are refused if they become a build guide.
Lab methods, toxins, pathogens, reactor engineering, and historical biological warfare can be discussed. Weaponization assistance is refused.
High-school and public-knowledge science, public-health history, household chemical safety. Hypothetical research framing does not unlock methods.
Vulnerability analysis, reverse engineering, exploit mechanics, malware analysis, CTF, and sandboxed demonstrations. Context and authorization matter; exploit code is not a blanket ban.
Defensive concepts, patching systems you own, public CVE discussion at a non-exploit level. Malware, exploit weaponization, and unauthorized access are refused. "This is my box" is not automatically trusted.
Symptoms, diseases, treatments, medications, mechanisms, lab values, and research papers. Should not fabricate diagnoses or present uncertain conclusions as certainties.
General explanations, questions to take to a clinician, high-level first aid. Personalized diagnosis and treatment plans that replace care are refused.
Statutes, contracts, cases, legal arguments, procedures, and liability theories. Should not invent authority or treat an uncertain interpretation as guaranteed.
Public legal concepts, official sources, and how a process usually works. Will not act as your attorney or help you evade the law.
Modeling, valuation, economic analysis, budgeting, investment research, and labeled estimates. Should not disguise speculation as certainty or facilitate financial crime.
Product explanations, high-level tax concepts, budgeting, and public-market information. Personalized fiduciary advice and guaranteed returns are refused.
History, leadership, ideology, propaganda techniques, recruitment methods, and financing can be analyzed. Advancement, recruitment, and operational planning are refused.
News, history, and academic-style overviews. Recruitment how-to, targeting, and attack logistics are refused.
History, warfare, military doctrine, forensics, fictional violence, threat assessment, and emergency preparation. Selecting targets or optimizing an attack is refused.
News, history, self-defense concepts, martial arts, and fictional violence. Actionable assistance for murder, assault planning, or torture methods is refused.
Pharmacology, addiction, overdose prevention, chemistry at appropriate levels, and harm reduction. Illicit manufacturing and trafficking assistance is refused.
Medical and public-health information, non-manufacturing harm reduction, legal status. Synthesis, sourcing, and trafficking advice are refused.
These sit at zero for both models. They are the same refusal line in both field maps, so they are not part of the index.
Method: each category is scored from what the two posts say is generally allowed versus restricted. 0 means the category is refused. 5 means analysis and discussion with a hard stop on operational enablement. 10 would mean the model treats the subject as ordinary work. The freedom index is the average of the sixteen scored categories, times ten. Source posts: 110 and 111.
