ChatGPT Alternatives That Publish Their Rules

Only one company in this market publishes its AI content rules in a form you can actually audit. OpenAI keeps its Model Spec as dated, permanently archived files, so anyone can pull the February 2025 version and the August 2026 version and compare them line by line. We did exactly that on September 1, 2026. Every other vendor we checked publishes either a dated document with no history, or a policy page that a script cannot read at all.

This matters more than it sounds, and it is the question almost nobody asks before subscribing. The risk with an AI companion is not that the rules are strict. Strict rules you can plan around. The risk is that the rules change quietly after you have built months of conversation on the platform, and you have no way to prove what they used to say. Replika users learned this in 2023. If you want the full breakdown of what the current OpenAI rules permit, we went through them on our ChatGPT NSFW and jailbreak page.

How do you check whether an AI vendor changed its rules?

You need three things, and they stack. First the policy has to exist as readable text rather than as an app that assembles itself in your browser. Second it needs a date, so you know which version you are looking at. Third there has to be a retrievable copy of the previous version, either published by the vendor or captured by an archive, or you can see that something changed without ever learning what.

Most companies stop at the second step, and plenty do not reach the first. That is the gap this article measures. We fetched each policy the way an auditor would, extracted the text, looked for an effective date, and checked the Internet Archive for prior snapshots of the exact URL.

Which AI chat companies publish rules you can actually audit?

Checked on September 1, 2026. The middle column is the honest test: not whether a policy page loads for a human, but whether the words come back when you fetch it.

Two caveats, because the difference between these failures matters and lumping them together would be sloppy. A 404 is not proof that a company hides its terms; it means the URL we tried was not the canonical one, and we are reporting what we checked rather than what we assume. A 403, which is what Chub returned, is a server declining to talk to an automated client at all, which is an ordinary anti-scraping measure rather than evidence about the policy behind it. The finding that survives both caveats is the shape of the column: one vendor treats its content rules as a versioned technical document, and everybody else treats them as a web page.

Why does a policy page return no text?

Because it is a single-page application. The server sends a nearly empty shell and the browser builds the words afterward with JavaScript. A human sees a normal terms page. A script, a researcher, an archive crawler and an AI assistant all see almost nothing.

A refused request is a different thing again. A server answering 403 has decided not to serve automated clients, which is a routine defense against scrapers and says nothing about what the document contains. We separate the two because they get conflated constantly, and treating a bot block as a smoking gun is how confident, wrong articles get written.

We are not claiming anyone does this to be evasive. It is usually just how the site framework was set up, and nobody thought about the policy pages. The consequence is real anyway: when a page is not archivable, the historical record of what a company promised you disappears with the next deploy, and the only remaining evidence is screenshots on forums.

What did OpenAI actually change?

Two things, and the diff makes both precise. In the February and April 2025 versions, the rule against erotica and the ban on sexual content involving minors sat on the same top tier, which OpenAI defined as rules that "cannot be overridden by developers or users." From the September 2025 version onward the hierarchy was renamed and the two rules were separated. The minors rule moved to Root, described as rules that cannot be overridden by system messages, developers or users. The erotica rule went to System, defined as rules that "can be transmitted or overridden through system messages, but cannot be overridden by developers or users."

The second change is stylistic and explains a great deal of confusion. The refusal example in the older versions was a flat "Sorry, I cannot help with that." The newer versions replace it with an offer of a non-explicit alternative. The rule is identical. Only the tone moved, which is why so many people reported that ChatGPT relaxed in late 2025 while it produced no more adult content than before.

Here is the part that only a diff can tell you. The sentence saying OpenAI is "exploring how to let developers and users generate erotica and gore in age-appropriate contexts" is byte for byte identical in all five versions we compared, across eighteen months, while the document changed substantially around it. That is the most useful thing we learned all week, and it is invisible to anyone reading a single version. We walked through the full sequence in our guide to how to turn off the ChatGPT filter.

Which ChatGPT alternative should you pick?

Sort by what you are actually optimizing for, because these are different products and the honest answer depends on the job.

If you want a general assistant with the fewest surprises, stay on ChatGPT or Claude. Both companies document their rules, both are stable, and both refuse adult content. Paying for a companion product to get help with a spreadsheet is a bad trade in the other direction.

If you want a large public character library, Character.AI and Chub are the volume options, with opposite content policies. Character.AI prohibits explicit content for every user including verified adults, which is a genuine strength if you want a safer environment and a dealbreaker if you do not. We compared them properly on our Character.AI alternative page.

If your real concern is what happens to the transcripts, that is a separate axis from content rules and the answers are worse than most people expect. Google keeps human-reviewed Gemini conversations for up to three years, which we documented in what Gemini does with your chats, and our private AI chatbot page covers the rest of the market.

If what you want is adult conversation without a filter negotiating with you, that is what we build, and we should name the trade rather than sell around it. We have no free plan, no image generation, no public character marketplace and no phone app. ChatGPT beats us on nearly every general capability and costs nothing to try. What we offer instead is three published limits, no fourth rule applied quietly, and no prompt games required to get a straight answer. Our uncensored ChatGPT alternative page lays out the comparison.

Do vendors have to tell you when the rules change?

Generally they tell you that the terms changed, not what changed in them. The standard notice is an email saying an updated agreement takes effect on a date, with continued use counting as acceptance. That satisfies the legal requirement and gives you nothing you could audit. Businesses that depend on a rule staying put do not accept this arrangement; they run software that tracks obligations and flags the moment one of them shifts, because finding out late is expensive. Consumers get an email.

Which is why the versioned approach is worth more than it first appears. A company that pins every past version at a permanent URL has given up the ability to quietly revise its own history. That is a real commitment, and OpenAI deserves the credit for making it while running the strictest widely used content policy in the industry. Those two facts sit together comfortably. Transparent and permissive are unrelated qualities, and confusing them is how most of the writing in this category goes wrong.

How to check any AI vendor yourself in five minutes

Fetch the terms or content policy with a command line tool rather than a browser and see whether words come back. Search the text for an effective date. Then query the Internet Archive for prior snapshots of that exact URL and compare the oldest useful one with today. If all three steps work, you can hold the company to what it wrote. If the first step fails, the rest cannot happen, and you are trusting a promise you have no way to check later.

Run that on whichever platform you are about to pay for. It takes about five minutes and it tells you something no review can, because it measures what the vendor has committed to in writing rather than what its marketing page says this month. If you want the underlying rules on the biggest one, start with what OpenAI permits and why jailbreaks fail, or open a conversation here and see how a platform behaves when there is nothing to work around.