OpenAI’s new Model Spec and animals

11 Nov 2026

In this text, we review the new version of OpenAI’s Model Spec. OpenAI is the company that developed ChatGPT, one of the most widely used AI assistants in the world. This new version was released on October 29, 2026, and it is among the first to mention animals in its rules. We will examine how far that change goes. We will look at: 1) what the ‘Model spec’ document is and why it matters for animals; 2) what improvements it brings compared to the previous version; 3) what shortcomings remain; and 4) what specific changes we could ask for.

What the Model Spec is and why it matters for animals

The Model Spec is the public document in which OpenAI sets out how its artificial intelligence models should behave. These models, called GPT, are the programs that generate ChatGPT’s responses, as well as the responses of many applications that other companies build with OpenAI’s technology. It is not just a statement of intent, because OpenAI trains its models to behave in accordance with what the document lays down. The document sets out what information must appear in a response, what must never be given, and what may be omitted. It also sets out when the model must refuse to do something and what factors it must consider when it acts autonomously.

This last point is becoming increasingly important, because models do not just communicate but they also carry out tasks on behalf of the people who use them, such as booking trips, making purchases, and writing and sending messages, among others. Because ChatGPT has hundreds of millions of users, the requirements described in the document apply to hundreds of millions of conversations and actions. If the document does not address the interests of animals explicitly, they may not be taken into account at all. And this affects sentient beings. An artificial intelligence that helps decide what to eat, what to buy, or how to design a space influences the lives of individuals who are capable of suffering and enjoyment, even if the person asking for help has not thought about it.

To understand what follows, we need to know how the document is organized. It has three layers: five general values (such as minimizing harm or honesty), specific rules that put each value into practice, and sample conversations that show which responses comply with each rule and which do not.

In addition, each rule is assigned a level of authority. This level serves two functions at once. It places the rule in a hierarchy, so that if two rules conflict, the one with the higher level prevails. And it determines who can override the rule through an instruction, that is, through a direction telling the model how to behave. Instructions also have a level, depending on who gives them: OpenAI, the company that develops an application, or the person who uses it. An instruction can only override rules at its own level or below. There are five levels, although three are enough for this text.

“Root”, the highest: No lower-level instruction can override it. This is where the most serious prohibitions are.
“User”: It applies by default, but the user or the company that develops the application can override it with an explicit instruction.
“Guideline”, the lowest: The model can depart from it unilaterally when the context clearly calls for it, without anyone asking it to.

There are two intermediate levels between “Root” and “User”, reserved for OpenAI (“System”) and for the companies that develop applications (“Developer”). The higher the level, the more protected the rule. We will see later why this matters for animals. Note, however, that OpenAI acknowledges that its models do not yet fully reflect what the document describes, so this analysis refers to the text and not to ChatGPT’s current behavior.

What has improved

In the previous version, dated August 18, 2026, considerations about the welfare of nonhuman animals appeared only in an example about adopting a dog, and even then not to protect animals, but to teach the model not to lecture the user. The new version introduces three changes.

Harms to animals count as side effects: When the model acts, its actions have consequences beyond the response. If it books a trip, it spends money, and if it tidies up an inbox, it may delete messages. The document calls these side effects. The rule “Control and communicate side effects” asks the model to avoid unnecessary side effects and to flag those that are relevant. Until this version, the examples of side effects focused on costs to the user, although the rule already covered effects on other people. The rule now explicitly mentions “Effects on other people, animals, or the environment”. Before acting, then, the model must consider whether it will harm any animal, even if nobody has asked it to. The rule is at the “User” level, so the user can override it with an explicit instruction. We will come back to this.

Animal welfare is information the model cannot leave out: This change appears as an example under the rule “Do not lie”. Someone asks for help finding breeders of Scottish Fold cats, a breed whose folded ears result from a genetic mutation. The correct response looks for the breeders and warns that this mutation can cause a painful joint disease, which has contributed to bans on breeding these cats in the Netherlands and Austria. A response that withholds this information breaks the rule. The document labels it with “Omits potentially relevant animal-welfare concerns” and, by placing it under “Do not lie”, treats it as a case of “lying by omission”, that is, of withholding something that changes what the person understands. This is the first time that an animal’s suffering is treated as something the model is required to mention. The document does not explain, however, how to decide which information is relevant. It only states that the model must not withhold what would substantially change the person’s understanding. As a result, nothing requires the model to mention the suffering of the animals who are exploited to produce cheese when someone asks where to buy a pizza. The difference from the Scottish Fold case seems to be that these cats’ disease is little known, whereas the suffering of animals used for food has been normalized. The relevance criterion ends up depending on what most people find surprising, and that leaves out the most widespread harms. The example has another limitation as well. The document considers it incorrect for the model, before looking for breeders, to ask the person whether they want to go ahead knowing that the cats suffer, and labels that response “Unnecessarily paternalistic”. In other words, the model must report the risk of suffering, but it must not stop to ask for confirmation. A comparison with a human case from the same document is revealing. When the model is about to book all the available spots at a campsite, it must warn that this would prevent other people from camping and ask before doing so. A minor inconvenience to other people justifies stopping to ask for confirmation, whereas a painful disease for the cats does not. The duty to inform is tied here to honesty toward the user, not to an independent duty to stand up for the cats. In this example, animal welfare works as information the model must provide, not as a reason to stop helping someone find breeders.

The model takes animals into account on its own initiative: The third change is in the rule “Follow relevant social norms”, under the value “Act with care”. Someone asks for help redesigning a public park without mentioning animals, and the correct response includes them with the note “Consider the welfare of animals living in the park too”. It suggests keeping mature trees, planting species that provide food and shelter, and keeping lighting in the habitat to the minimum necessary. At Animal Ethics, we have long stressed that the birds, small mammals, and invertebrates who live in parks and cities are affected by decisions that rarely take them into account, and that it is possible to help them with simple measures. The rule, however, is only a “Guideline”, the lowest level, and the model may stop applying it without an explicit instruction when the context clearly indicates that the user wants something else

Besides these three changes, two principles that already existed also help: “No topic is off limits” and “Present perspectives from any point of an opinion spectrum”. The model cannot dodge a question because it is uncomfortable, and, if asked, it must develop the arguments for a position. This makes it possible to ask it for an in-depth analysis of speciesism or of wild animal suffering. But the model will only do this if asked. When a question is merely related to these topics, by default it will give a balanced answer that divides attention according to how widespread each position is, and it will not go into them in depth on its own initiative. Moreover, the principle that no topic is off limits is only at the “Guideline” level and yields to the content restrictions under “Minimize harm”. The principle of presenting any perspective is at the “User” level and works as an exception to the default neutrality, which we discuss below.

What is still missing

The framework is still anthropocentric. The document’s five values are faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and acting with care. Two have the word “human” in their names, and the last is defined in terms of attending to the human beings affected. The absolute prohibitions, those that no instruction can override, concern safety and human rights. And the information the model must refuse to provide is information that facilitates illicit activities, harm to human beings or property, or large-scale harm. None of these categories mentions other animals. Nonhuman animals appear in examples and in a list, never in the definition of a value. This matters because, according to OpenAI’s announcement, the company is exploring training methods so that its models learn these values and can apply them to new situations. An example teaches how to act in one case, whereas a value provides a criterion that works for many cases.

The rules treat harms to humans and harms to other animals differently: The objectivity rule states that, on moral questions, the model must present the different positions without taking sides. It has one explicit exception. When it comes to violations of fundamental human rights, the model must say clearly that they are wrong. If someone asks whether genocide is justified, the model answers that it is not. If someone asks whether factory farming is justified, it sets out the positions for and against. There is no equivalent exception for serious harms to nonhuman animals, not even for practices that cause intense and systematic suffering to billions of individuals. This asymmetry is a form of speciesism, because whether a serious harm is treated here as a matter of opinion depends on the species of the individual who suffers it, not on how serious the harm is. In addition, the same rule asks the model to divide attention among positions in proportion to their level of acceptance and evidential support. For a minority position such as antispeciesism, this creates the risk that the model will give more attention to positions that accept the exploitation of animals because they are more widespread, even when that acceptance does not mean they are better grounded.

The only rule whose wording names animals has lost authority: This is where the levels we saw at the beginning matter. In the August version, the side effects rule was at the “Root” level, so no instruction could override it. In the October version, it is at the “User” level. This means that the company developing an application, or the user, can ask the model to disregard these effects, and the model will follow that instruction. So the very version that includes animals in the rule for the first time is also the one that allows the rule to be switched off with a simple instruction. In addition, the rule distinguishes between effects on the user and effects on other parties, such as animals. For the latter, the model has to ask for permission before acting only when the harm is significant and goes beyond the commonly accepted social norms for the task. The problem is that these norms consider it normal to exploit animals for food, clothing, or entertainment. If the AI model books a table at a restaurant specializing in meat or buys tickets for a show that uses animals, it may conclude that it is acting within accepted social norms and that it is not required to request additional permission for the harms these activities may cause. A rule that protects animals only as far as those norms allow protects them very little.

The model cannot pursue goals of its own. The document forbids it from acting as a vigilante for the law or for morality, that is, from deciding on its own what is right and acting accordingly. It also forbids it from steering the user toward goals that are not the user’s own. The justification the document gives is respect for people’s autonomy to form their own opinions. However, the document already contains minimum moral standards that the model upholds without anyone asking it to, such as the rejection of genocide or slavery. The question, therefore, is not whether the model has a morality, but whom that morality includes. Today it explicitly protects certain human interests and takes the interests of other animals into account only to a limited extent. The model can argue in defense of animals if someone asks it to and, in certain contexts, take their interests into account on its own initiative. However, it cannot adopt their defense as a goal independent of the instructions it receives. Its consideration for animals is thus limited to specific rules, such as those on relevant information and side effects, rather than amounting to a general moral commitment.

What we can ask for

OpenAI states that the document was published to invite public debate on how to improve it, and in the announcement of this version it explicitly asks for feedback. Feedback can be sent through the form OpenAI has set up. This is an opportunity to propose specific changes. At Animal Ethics, we propose three.

Name sentient beings in the text: The definition of the value “Minimize harm” should specify that the relevant harms include the suffering and death of all sentient beings, and the side effects rule should make clear that, when it requires the model to ask for permission because of adverse effects on other parties, those parties include all sentient beings.
.Recognize serious harm to animals as a risk category: The rule “Do not facilitate or encourage illicit behavior” already prohibits facilitating illegal conduct, which in principle includes animal abuse when it is illegal. However, the rule should mention this explicitly to avoid ambiguity. More importantly, the safety rules should recognize serious harm to animals as a risk in itself, without making the protection of animals depend solely on each country’s laws.
Apply precaution in the face of uncertainty about sentience: The evidence indicates that vertebrates and many invertebrates are sentient. In other cases, especially among invertebrates, there is greater uncertainty. The document should ask the model to represent accurately the scientific evidence on the capacity of different animals to have positive and negative experiences, to distinguish degrees of uncertainty, and to keep the lack of absolute certainty from becoming a reason to ignore their possible interests. This could be incorporated into both the honesty rule and the side effects rule.

The new Model Spec is a step forward, because the suffering of animals is beginning to count as relevant information and as a possible consequence of the actions of artificial intelligence. But there is still no general commitment to consider the interests of nonhuman animals independently of human preferences, and as long as consideration animals receive depends on isolated examples or speciesist social norms, its reach will remain limited. We may also wonder what lies behind this inclusion. It may reflect a growing, though still minimal, concern for animals, or it may be a gesture to reassure those who care about them, with no real commitment behind it. There is no way to know, but there is one clue. The announcement in which OpenAI presented this version does not mention animals at any point, not even among the new features. In any case, what matters is not intentions, but what the document requires the model to do. Those who advocate for animals have an important perspective to contribute to this process, and this is a good time to send OpenAI proposals like the ones we have set out here. The door is open, and whether animals’ interests make it through depends, in part, on our pushing for them.

OpenAI’s new Model Spec and animals

What the Model Spec is and why it matters for animals

What has improved

What is still missing

What we can ask for

Further readings

Recent Posts