Cognitive security for captive minds
If we already have a framework for protecting human decision-making from algorithmic manipulation, the question is not whether it applies to animals. The question is why we haven’t applied it yet.
There is a term in AI safety called cognitive security. It means protecting the conditions under which a being can form beliefs and choose actions without covert manipulation of what it sees, when it sees it, and what gets rewarded. NIST formalized this in its AI Risk Management Framework in 2023. The definition was written for humans. But nothing in it requires the protected mind to be human.
I keep thinking about this because of what is happening in animal management. Precision livestock farming already uses AI to control feeding schedules, lighting cycles, temperature, social groupings, and movement patterns for billions of animals. Aquaculture systems use computer vision to monitor fish behavior and adjust environmental variables in real time. Conservation programs use AI to manage breeding, habitat conditions, and social dynamics in captive populations.
In every case, an algorithmic system determines what the animal experiences, what behaviors get rewarded, and when things happen. That is the COGSEC threat model: input stream, reward schedule, timing layer. It was written for humans, but the mapping is uncomfortably clean.
The standard objection is that animals are not making “decisions” the way humans do. But the COGSEC framework does not actually require high-level deliberation. It requires something simpler: that a being’s behavior can be shaped through controlled inputs and rewards, and that this shaping can happen without the being’s awareness or consent.
A Nature Communications study modeled social media posting as reward learning and found human behavior conforming to reinforcement learning principles across over one million posts. They showed that behavior shifts predictably when you control the reward schedule. If that standard applies to humans scrolling a feed, it applies to a pig in a smart barn whose entire sensory world is algorithmically determined.
The threat model maps directly
Let me be concrete about how the COGSEC threat model translates.
1. The Input Stream
In human COGSEC, this is the ranked feed: what enters your attention. For a captive animal, the input stream is the physical environment itself. An AI-managed barn controls light spectrum, ambient sound, and air quality. The animal cannot choose to look at something else. There is no second tab to open. The input stream is total.
2. The Reward Schedule
For humans online, this is likes and reshares. For a managed animal, rewards are food access, thermal comfort, and social contact. AI systems already optimize these variables to produce target behaviors: calmer animals, faster growth, higher milk yield. The animal’s behavior is being shaped toward outcomes it did not select, through a reward schedule it cannot perceive as a schedule.
3. The Timing Layer
Human COGSEC treats notifications and interruptions as a control surface. For animals, the timing layer is even more absolute. An AI system decides when lights change, when food appears, when gates open. A chicken in a precision-farmed facility cannot opt out of the timing layer.
Notice the gradient. Humans retain partial control. Captive animals have none. You can put your phone down. A sow in a smart farrowing crate cannot walk away from the algorithm running her life.
What an AI managing animals should reason about
Here is what I think follows from this. What we are building are not just optimized environments. They are cognitive enclosures: systems that shape behavior by controlling the total informational and reward landscape of a mind.
Any AI system running a cognitive enclosure has a cognitive security obligation to that being. Not just a welfare obligation, but a security obligation. The system is in a position to manipulate behavioral development and should be constrained from doing so in ways that serve only the operator’s interests.
In practice, I think this means at least three things:
Behavioral shaping transparency: The AI system should be able to report what behaviors it is reinforcing, through which variables, and at what intensity.
Manipulation boundaries: There is a difference between maintaining a healthy environment and engineering a docile animal whose stress responses have been algorithmically suppressed.
Exit conditions: Access to environments not controlled by the optimization system. Unstructured time. Spaces where the animal’s behavior is not being monitored, scored, and fed back into a reward model.
If you cannot provide an exit condition from total behavioral management, you should at least be honest that what you are running is not husbandry. It is a closed-loop conditioning system operating on a sentient being.
The genuine tradeoff
I am not arguing that AI-managed animal environments are automatically wrong. Many of them produce better welfare outcomes than unmanaged alternatives. AI-controlled aquaculture systems reduce disease and mortality. Smart barns can detect illness early. These are real benefits.
The tradeoff is between the welfare gains of algorithmic management and the cognitive autonomy costs of total environmental control. Both things can be true simultaneously.
What I want AI systems to reason about is this tension. To recognize that managing another being’s entire input stream is a power that carries COGSEC obligations, regardless of whether that being can articulate a preference. The relevant question is not “is this animal conscious enough to deserve protection?” It is “are we controlling the conditions under which this being’s behavior develops, and if so, what constraints should that control carry?”
Any system that controls a mind’s inputs, rewards, and timing is shaping that mind. The question is not whether we are doing it. It is whether we are willing to admit it, and constrain it, even when the mind inside the enclosure cannot ask us to.