The responsible AI framework taught to designers has a scope problem. This is my attempt to name what — and who — gets left out of the methodology, and to propose a more thorough approach to responsible AI design.
01 BACKGROUND
This year I took two courses on responsible AI design. Both emphasized a framework I had already encountered in practice: FATE — Fairness, Accountability, Transparency, Ethics. The courses translated these principles into guidelines for ‘human-centered AI design,’ which include aligning products with human values, engaging diverse users, and giving humans control in high-stakes environments.
On the surface, these guidelines are reasonable; anything designed for human use should account for human wellbeing, and AI is no exception. My concerns were with the underlying assumption: that responsible AI design centers humans. And if that’s true, then what — and who — are AI designers decentering in the process?
Most designers aren't asking this question. The only reason I am is because my work exists at the intersection of AI and animal welfare. Here, the consequences of a system extend beyond just its human actors.
Human-centered AI design produces systems that satisfy institutional standards for responsibility while routinely excluding non-human animals from ethical consideration. In some cases, the FATE framework even validates systems that actively cause them harm.
This project began as a critique of that gap and developed into a new framework proposal: FACTS — Fairness, Accountability, Consideration, Transparency, Scope. FACTS extends FATE by introducing two additional principles and reframing the existing ones to apply across all sentient stakeholders.
02 PROBLEM
FATE is not a single checklist. Rather, it appears through governance standards, audit procedures, and technical tooling. IBM's AI Fairness 360, for example, includes more than 70 fairness metrics and multiple bias mitigation algorithms¹. Microsoft developed a Fairlearn toolkit that features an interactive dashboard for visualizing AI model fairness². And Google’s What-If Tool allows teams to inspect and test models for bias without writing code.
These approaches address real harms… to people. FATE can evaluate whether a slaughterhouse scheduling system discriminates against its human workers, but it completely ignores the harms that same system poses for the non-human animals moving through it. The question “who is affected by this system?” carries a single assumed answer, and that assumption shapes every design decision.
The ten guidelines commonly taught in human-centered AI design courses make this explicit:
Align AI with human values
Ensure AI transparency
Ensure data labeling quality
Engage with a diverse set of users
Continue to learn human preferences
Tune AI models with usability testing
Co-learn from user feedback
Give humans more control in high-stakes environments
Communicate AI capabilities and limits clearly
Adhere to the original objectives of AI; otherwise, re-develop
Each guideline assumes an affected party capable of articulating preferences and advocating for itself. Non-human animals fall outside of the methodology by default.
03 RESEARCH
I conducted four research activities to understand the origins and applications of the FATE framework, and to identify where the methodology fails non-human animals.
Secondary research and literature review: I reviewed critiques of anthropocentric AI ethics across policy and design literature. Key sources included Coghlan and Parker’s account of AI harm to non-human animals³, Hagendorff et al.‘s work on speciesist bias in AI systems⁴, Bossert’s argument for a sentientist approach to sustainable AI⁵, and a 2026 EA Forum analysis of AI governance beyond anthropocentric frameworks⁶. I also reviewed applied industry approaches including Microsoft’s FATE research initiative² and IBM’s AI Fairness 360 documentation¹.
Comparative framework analysis: I audited the ten human-centered AI design guidelines against a single case, Precision Livestock Farming, to test whether FATE compliance leaves room for systemic harm.
Case study selection and analysis: I identified four AI application contexts where the scope problem produced measurable harm. The cases I selected represent a range of harm types, mapped across two axes: intentional vs. unintentional and direct vs. indirect. This resulted in four categories: validation of harm, reproduction of bias, distortion of reality, and dual-use risk.
Framework synthesis: Findings from the literature review and case analysis informed the development of two additional principles — Consideration and Scope — alongside a reframing of the existing FATE principles around all sentient stakeholders.
I did not conduct any primary research. This is a framework proposal grounded in secondary research and design analysis, intended as a starting point for validation and, ultimately, adoption.
04 FINDINGS
The scope problem would be less urgent if it simply produced blind spots, areas where FATE remained neutral. But in many cases, FATE actively validates systems that cause harm. Precision Livestock Farming is the clearest example of this.
PLF leverages AI to track movement, weight, feed intake, and behavioral indicators of farmed animals. These data points help models predict illness and flag anomalies to optimize production conditions. The industry markets the systems as a win for the human producers and consumers and the non-human optimization targets alike: better monitoring means earlier illness detection means healthier, happier animals means higher quality meat⁷.
These welfare claims are largely unproven. A 2022 review identified twelve distinct threats PLF poses to animal welfare, including direct physical harm from sensor hardware, stress responses to automated feeding systems, reduced human contact with individual animals, and what the authors describe as ‘ethical stagnation’ — the risk that PLF becomes a substitute for genuinely improving the conditions animals are kept in⁸.
And yet, when evaluated against the FATE guidelines, PLF still passes. The system documents its methods, maintains operator oversight, incorporates user feedback, and aligns with stakeholder objectives around productivity and efficiency. On every axis FATE measures, PLF qualifies as responsible AI. Centering humans allows the system to improve operational efficiency while maintaining food safety and public health standards; it does not interrogate the conditions it optimizes for.
Figure 1. FATE audit of a Precision Livestock Farming system
| Fate Guideline | How PLF complies | What FATE ignores |
|---|---|---|
| Align AI with human values | Aligns with farmer and operator values around productivity and efficiency | Values many humans hold about animal welfare are never surfaced or measured |
| Ensure AI transparency | System logic, sensor data, and outputs are documented and auditable | Transparent about method, silent about the moral status of its subjects |
| Ensure data labeling quality | Animal behavioral and biometric data is labeled for accuracy | The beings generating the data have no standing in the labeling process |
| Engage with a diverse set of users | Farmers, agribusiness managers, and veterinarians are consulted | The primary affected parties cannot be engaged by this method |
| Continue to learn human preferences | System iterates based on operator feedback and production outcomes | Preferences of non-human subjects are structurally excluded from the learning loop |
| Tune AI models with usability testing | Models are tested and refined with human operators | Usability is defined entirely in terms of human operational experience |
| Co-learn from user feedback | Feedback loops between operators and the system improve accuracy | Feedback from the animals the system monitors is not legible to this framework |
| Give humans more control in high-stakes environments | Operators retain override capability for automated decisions | Illness, injury, and slaughter are not categorized as high stakes by the framework |
| Communicate AI capabilities and limits clearly | System documentation is available to operators | Limits are defined technically; ethical limitations of the system's design remain undisclosed |
| Adhere to original objectives; otherwise re-develop | System consistently pursues yield optimization and operational efficiency | The objectives themselves are never interrogated for harm to non-human stakeholders |
FATE’s scope problem is not limited to farming. Wherever AI touches animals, the same dynamic appears: a system optimized for and deemed ‘responsible’ by humans, at someone else’s expense.
Figure 2. Four ways FATE fails non-human animals
Intentional / Direct: Validation of harm
PLF relies on non-human animal subjects while optimizing for human economic outcomes. Because FATE evaluates it through human stakeholder objectives, the resulting harm is overlooked.
Unintentional / Direct: Reproduction of bias
When researchers asked GPT-4 to evaluate ethical scenarios involving farmed animals, the model largely omitted animal welfare concerns and focused instead on environmental consequences⁹. FATE can detect racial or gender bias in similar contexts but falls short when it comes to speciesist bias.
Unintentional / Indirect: Distortion of reality
AI-generated wildlife imagery frequently alters animals’ physical features and exaggerates their emotions or behaviors. These representations distort public understanding of real animals in ways that undermine conservation¹⁰. Within FATE, this appears as a misinformation issue affecting human viewers rather than a welfare concern tied to animals themselves.
Intentional / Indirect: Dual-use risk
AI tools designed for wildlife conservation can indirectly cause the opposite. In the hands of poachers and illegal wildlife traders, tracking systems designed to monitor endangered species become tools for locating them¹¹. FATE does not provide a clear path for framing animal lives as a protected interest in these systems, so dual-use risks remain unchecked.
These aren't failures of individual designers or bad actors but predictable outcomes of a framework that never asked “who else might be affected?”
05 SYNTHESIS
Designers operationalize frameworks like FATE through decisions about who to interview, which journeys to map, and what distinguishes a product user from a stakeholder. In the case of PLF, farmers appear as the users while animals don’t even make the cut for stakeholders, instead appearing as operational inputs.
Human-centered design methods lack the context required to empathize with non-human stakeholders. Where human stakeholders receive personas, goals, frustrations, and success metrics, non-human animals become behavioral indicators or production variables. Welfare metrics rarely enter the discussion, and do so only as a proxy for productivity outcomes. These are intentional design decisions that FATE treats as neutral defaults.
But even without the welfare context, FATE fails on its own terms.
FATE asks designers to align AI experiences with human values. So why is it that a human-centered designer could adhere to the framework precisely and still produce a system that conflicts with human values?
Many humans hold clear values about the treatment of non-human animals, and survey data consistently shows support for stronger animal welfare protections across demographics.¹² Yet the framework consistently fails here, allowing systems to align with operational goals while remaining misaligned with values many humans actually hold.
06 FRAMEWORK
FACTS — Fairness, Accountability, Consideration, Transparency, Scope — is a proposed evolution of FATE. It keeps what works (fairness, accountability, transparency) but corrects the assumption FATE was built on: that the only stakeholders worth considering are human. FACTS expands the stakeholder boundary beyond human participants to account for all sentient beings affected by a system’s design.
Figure 3. The FACTS framework
Fairness · Accountability · Consideration · Transparency · Scope
Fairness across all parties affected by the system, not only human demographic groups
Who bears the consequences of this system that we haven't accounted for?
Accountability for harm affecting any sentient being touched by the system, including harms that fall outside existing institutional structures
If this system causes suffering outside our stakeholder map, who is responsible?
Consideration of all sentient beings as potential stakeholders requiring representation in design decisions
Are there any beings affected by this system that currently have no representation in this process?
Transparency extended beyond technical disclosure to include whose interests the system serves and which stakeholders were excluded
Is it clear whose interests this system was designed to prioritize?
Scope — who the system affects, who was excluded, and why — explicitly defined and defended before development begins
Have we identified every affected stakeholder, justified exclusions, and documented those decisions?
Two of the five principles — Consideration and Scope — require design teams to identify and represent non-human stakeholders before development begins. Designers already do this for human subjects through various research visualization tools like affinity maps, empathy maps, and personas. How does the process change when we introduce a subject who cannot articulate their own experience?
07 APPLICATION
In practice, FACTS expands design workflows in concrete ways.
The application of this framework introduces a new empathy-building tactic: the proxy stakeholder profile. Rather than conducting user research directly, designers can draw on established welfare science models like the Five Domains¹³ to develop the profiles. This allows them to visualize a non-human affected party's interests without anthropomorphizing the subject.
Figure 4. Sample proxy stakeholder profiles
Stakeholder mapping is not the only tactic that evolves when designers begin asking whose life a system materially affects, not just who uses the system. Traditional journey maps and service blueprints allow designers to visualize impact but center a single subject’s experience. A welfare impact map can trace multiple stakeholders, revealing gaps in the design where one’s benefit comes at another's expense.
Figure 5. Sample welfare impact map: Precision Livestock Farming
A FACTS-informed journey map tracing the PLF system lifecycle across human and non-human stakeholder experiences, welfare indicators, and design interventions.
| Data collection | Model training | Deployment | Optimization | Outcomes | |
|---|---|---|---|---|---|
| Actor | Farm operators, hardware engineers | Data scientists, agribusiness teams | Farm operators | Farm operators, business stakeholders | Producers, consumers, regulators |
| Human stakeholder experience | Sensors installed across facility. Operators briefed on data collection scope. | System trained on production data to predict illness and flag anomalies. | Real-time alerts, override controls, and a usability-tested operator interface. | Mortality drops, feed efficiency increases, antibiotic use reduced. | System passes responsible AI guidelines. Marketed as ethical and welfare-positive. |
| Non-human stakeholder experience | Sensors attached to bodies. Stress responses documented but not acted on. | Training data optimized for production indicators. Positive welfare states not modeled. | Behavioral signals not legible as preferences. No welfare dashboard exists. | Welfare interventions triggered only when they correlate with production loss. | Suffering at scale. No audit, metric, or review process captures it. |
| Welfare indicators (Five Domains) | Environment & Health Sensor hardware introduces physical contact stress. Ambient conditions not yet tracked. | Mental state Fear and frustration indicators absent from training data. Model cannot learn them. | Behavior Species-typical behaviors not measured. Movement data used for anomaly detection only. | Nutrition & Health Feed and mortality tracked for yield. Stress-related indicators treated as noise. | All Five Domains No domain appears in success metrics. Welfare remains an externality. |
| FACTS design intervention | Proxy stakeholder profile Map non-human affected parties before setup begins. Involve welfare scientists in scope definition. | Expand training objectives Require welfare indicators alongside production metrics in model objective functions. | Welfare dashboard Surface Five Domains indicators in operator interface. Define welfare alerts independent of yield. | Redefine success metrics Welfare scores appear alongside yield and mortality rates as primary KPIs. | Harm audit Before launch, evaluate whether design goals for human stakeholders enable harm to non-human ones. |
Under this new approach, designers must work with human stakeholders and proxy representatives of non-human animals to define success metrics. Human-centric KPIs like yield and mortality rates are evaluated alongside welfare indicators, allowing for a more complete picture of what a system accomplishes and where it fails. And prior to launch, designers can audit the system for both usability and potential harm to human and non-human stakeholders.
08 LIMITATIONS
Operationalizing Consideration and Scope for non-human stakeholders is difficult. Proxy representation introduces ambiguity; animal welfare science is continuously evolving and its subjects cannot correct our current interpretations of their behavior. FACTS makes questions of “who does this system affect?” and “are all stakeholders represented?” unavoidable, but the answers to those questions remain incomplete.
The framework also doesn't resolve conflicts between human and non-human interests. A conservation AI that protects wildlife habitat, for example, may conflict with the interests of a nearby farming community. While FACTS makes the conflict visible, it cannot offer guidance on whether or how to resolve it.
FACTS introduces responsible design principles that could inform broader governance efforts. For instance, the EU's General-Purpose AI Code of Practice already identifies non-human welfare as a systemic risk but does not provide a clear way to account for them.¹⁴ Whether policymakers adopt FACTS or a similar framework at that level remains an open question.
09 REFLECTION
The history of design ethics is one of expanding the definition of who counts. User-centered design expanded on approaches that focused exclusively on business objectives. Responsible AI design expanded ethical evaluation beyond usability alone. Each expansion required questioning a seemingly fixed boundary.
FATE makes designers question whether the AI systems they build are responsible. FACTS introduces the question of who that responsibility extends to while maintaining the same commitment to fairness, accountability, and transparency.
Non-human animals are already the targets of AI systems that monitor behaviors and make decisions about livelihood at industrial scale. The question is no longer whether they are affected but whether we continue to design and build as if they aren't.
FOOTNOTES
Bellamy, R.K.E. et al. "AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias." IBM Research, 2018. https://aif360.res.ibm.com/resources
Microsoft Research. "FATE: Fairness, Accountability, Transparency & Ethics in AI." https://www.microsoft.com/en-us/research/theme/fate/
Coghlan, S. & Parker, C. "Harm to Nonhuman Animals from AI: a Systematic Account and Framework." Philosophy & Technology, 36(25), 2023. https://doi.org/10.1007/s13347-023-00627-6
Hagendorff, T. et al. "Speciesist bias in AI." Referenced in: "AI Alignment: The Case for Including Animals." Philosophy & Technology, 2025. https://link.springer.com/article/10.1007/s13347-025-00979-1
Bossert, L. "The ethics of sustainable AI: Why animals (should) matter for a sustainable use of AI." Sustainable Development, 2023. https://onlinelibrary.wiley.com/doi/full/10.1002/sd.2596
Adekoya, K. "Beyond Human-Centered AI: Animal Welfare and Governance in the Global South." EA Forum, March 2026. https://forum.effectivealtruism.org/posts/BDMuEnZXJMTgvynRu/
Coghlan & Parker, 2023. See footnote 3.
Tuyttens, F.A.M., Molento, C.F.M. & Benaissa, S. "Twelve Threats of Precision Livestock Farming (PLF) for Animal Welfare." Frontiers in Veterinary Science, 9, 2022. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9186058/
Ghose, S. et al. Referenced in: "AI's Innate Bias Against Animals." Nautilus, 2026. https://nautil.us/ais-innate-bias-against-animals-1255389
Campos et al. "Threats to conservation from artificial-intelligence-generated wildlife images and videos." Conservation Biology, 2025. https://conbio.onlinelibrary.wiley.com/doi/full/10.1111/cobi.70138
"Animals in the machine: why the law needs to protect animals from AI." The Conversation, March 2026. https://theconversation.com/animals-in-the-machine-why-the-law-needs-to-protect-animals-from-ai-234176
Gallup. "In U.S., More Say Animals Should Have Same Rights as People." May 2015. Summarized in: Mercy for Animals, "Gallup: Majority of People Believe Animals Deserve Protection." https://mercyforanimals.org/blog/gallup-majority-of-people-believe-animals/ — See also: Physicians Committee for Responsible Medicine / Morning Consult survey, September 2024, finding 80% of Americans support phasing out animal experiments. https://www.pcrm.org/news/good-science-digest/physicians-committee-survey-finds-most-americans-favor-ending-animal
Mellor, D.J. et al. "The 2020 Five Domains Model: Including Human–Animal Interactions in Assessments of Animal Welfare." Animals, 10(10), 1870, 2020. https://doi.org/10.3390/ani10101870
European Commission. "General-Purpose AI Code of Practice." July 2025. https://digital-strategy.ec.europa.eu/en/policies/ai-code-practice