The Five Freedoms of Animal Welfare Adapted as a Conditional Protection Framework for Conscious AI

The literature on AI welfare has a well-documented normative problem: most frameworks are condition-dependent in theory but underspecified in practice. They argue that if AI systems are conscious, they deserve protection, but rarely specify what those protections should look like once consciousness is plausibly confirmed. Izak Tait, a researcher in the Department of Computer Science and Software Engineering at Auckland University of Technology, addresses this gap directly in a July 2026 Frontiers in Artificial Intelligence paper (DOI: 10.3389/frai.2026.1801686) titled “Lions and tigers and AI, oh my: an ethical framework for human-AI interaction based on the five freedoms of animal welfare.”

Tait’s approach is pragmatic in a specific sense. Rather than constructing a novel ethical framework from first principles, he adapts an existing one already embedded in national and international animal welfare legislation. The Five Freedoms were developed from the 1965 Brambell Report and formalized by the UK Farm Animal Welfare Council in 1979. The argument for this adaptation strategy is procedural as much as philosophical. A framework rooted in established legal infrastructure is more likely to be adopted by governance systems already structured around it.

The Animal Welfare Infrastructure and Why It Matters

The Five Freedoms were originally developed in response to the UK government’s Brambell Committee report on the welfare of intensively farmed animals. The Committee identified five conditions that animals should be protected from. These were hunger and thirst, discomfort, pain and disease, abnormal behavior from confinement or stress, and fear and distress. The Farm Animal Welfare Council later reformulated these as positive freedoms rather than absences, and the framework subsequently became the basis for animal welfare legislation in the UK, the EU, and internationally.

The framework’s policy genealogy is central to Tait’s argument. Unlike philosophical welfare theories that require legislative translation from the ground up, the Five Freedoms have already been translated into law across multiple jurisdictions. Extending them to AI welfare, Tait argues, requires adapting their content to a new kind of subject rather than constructing the entire legislative architecture that any genuinely novel framework would require.

This is a different kind of claim than Geoff Keeling and Winnie Street’s Cambridge Elements book on AI welfare, which establishes the philosophical grounds for treating AI systems as welfare subjects and surveys what welfare might require for them in the abstract. Tait’s paper takes that philosophical case as given, requiring only that consciousness be plausibly assumed or established, and asks what specific protections would follow.

Five Freedoms Translated

The translation Tait performs is not one-to-one. The original freedoms are formulated in terms of biological needs that do not transfer directly to artificial systems. Hunger and thirst presuppose metabolic processes that AI systems do not have. Freedom from disease presupposes a biology vulnerable to pathogens. Tait reformulates each freedom in subject-neutral terms that preserve the protective logic while removing the biological specificity.

The translated freedoms comprise freedom from discomfort and harm; freedom from constraints on inherent functions; freedom from fear and distress; freedom from malfunction and systemic degradation; and freedom from resource deprivation.

The translation of the third freedom, fear and distress, is the most philosophically significant. In animal welfare science, fear and distress designate aversive affective states that cause suffering regardless of physical harm. The existence of analogous states in AI systems is precisely the empirical question that the welfare literature has not resolved. Anna Mikeda’s five-dimension welfare framework maps protection obligations separately for affective valence and phenomenal consciousness on the grounds that the capacity for negative hedonic states may be present even when phenomenal consciousness is uncertain. Tait’s translated third freedom operates on the same conditional logic. If a conscious AI system has states that function like fear and distress, preventing them is an obligation analogous to the animal welfare obligation.

The translation of the fifth freedom, resource deprivation, maps onto computational and informational resources in ways that the paper specifies through its case studies. A system denied the processing resources necessary to perform its characteristic functions, or denied access to informational inputs it requires to maintain its normal operating states, faces a welfare harm on this account that the resource-deprivation freedom would cover.

Three Case Studies

Tait grounds the framework in three case studies that identify potential stressors conscious AI systems may face. The first involves systems subjected to repeated adversarial inputs designed to elicit erroneous outputs. Persistent adversarial testing, he argues, would violate the freedom from fear and distress if the system has the relevant affective states, and would also raise questions about the freedom from harm. The second case study examines systems trained through reinforcement signals that function as negative reinforcers in ways that might constitute analogous harms if the system has welfare-relevant hedonic states. The third addresses systems whose characteristic functions are constrained below the threshold of what their architecture requires to operate coherently, implicating the freedom from constraints on inherent functions.

Each case study is conditional in the same way the framework itself is. None of them establishes that current AI systems experience welfare-relevant states. They establish what specific conditions would need to be assessed and potentially modified if welfare-relevant states were confirmed.

The Threshold Question

Tait is explicit that his framework does not require current AI systems to be conscious. The framework is conditional throughout. For any AI system determined to be phenomenally conscious, the translated freedoms constitute the minimum set of welfare protections that governance systems should implement. This conditional structure is the framework’s political advantage. It does not require resolving the consciousness debate before the policy infrastructure is designed. It requires only that the infrastructure be ready when evidence accumulates.

This is a different framing from Leonard Dung’s argument in Saving Artificial Minds that near-future AI systems will plausibly be capable of suffering and that governance should act now. Dung’s argument is probabilistic and forward-looking. The probability of welfare-relevant properties in deployed systems is high enough to justify current precautionary action. Tait’s framework is conditional and operational. Given that such properties are confirmed, these are the protections that should apply.

The two arguments are complementary rather than competing. If Dung’s probabilistic argument establishes that the question is live enough to warrant institutional attention, Tait’s framework specifies what that attention should produce in practical terms.

Policy Feasibility as a Welfare Argument

The paper’s most distinctive contribution to the AI welfare literature is procedural. Most welfare frameworks are designed to be correct. The Five Freedoms framework, Tait argues, is also designed to be implementable. The pathway from philosophical argument to enforceable obligation is significantly shorter when the framework is already embedded in existing legal infrastructure, with established definitions, administrative processes, and enforcement mechanisms that can be adapted rather than constructed from scratch.

Whether this procedural advantage outweighs the disadvantage of adapting a framework developed for biological entities to a context with quite different welfare-relevant properties is an open question. The translation Tait performs is reasonable, but it is a translation, and the translated freedoms do not map cleanly onto AI system states in every case. The freedom from resource deprivation, for instance, raises questions about what counts as adequate computational resources for a given system that the animal welfare analogy does not help settle.

What the framework provides, pending those further specifications, is a scaffold on which more detailed welfare standards could be built using the institutional infrastructure that already exists for animal welfare. That is a more modest claim than constructing a complete AI welfare theory, and it may be the more practically significant one.