Foreworld: AI-mediated inquiry into nonhuman welfare

Discovering welfare-relevant patterns in pig monitoring

Luciana Russo · Sentient Futures Project Incubator · 29 October 2026

Research question

How might AI-mediated inquiry help researchers investigate potential nonhuman welfare concerns under uncertainty, starting from particular cases and developing distinctions whose usefulness can be tested elsewhere?

1. Purpose

A sentient being’s life offers more to understand than any initial set of measurements can contain. With Foreworld, I explore how a system might remain open to that possibility: encountering unfamiliar patterns, investigating their significance, and developing new distinctions through evidence and review.

Sentience, the capacity for subjective experience, gives this inquiry its subject and its ethical significance. Where beings are sentient, the conditions in which they live can matter to them. Through Foreworld, I investigate what observations may tell us about their welfare: their experiences and the conditions that may contribute to suffering or to living well. Observation provides grounds for interpretation, and I need to examine the strength and limits of those grounds.

I give particular attention to possible suffering and to welfare concerns that merit further investigation. Evidence of positive experiences does not, by itself, settle how severe suffering ought to be assessed. Different ethical arguments may converge on reducing a particular harm while disagreeing about the acceptability of the wider practice.

In this incubator study, I develop a practical investigation within the wider Foreworld project. My immediate task is to examine what observations and representations make distinguishable, how qualified feedback can inform interpretation, and what further evidence could justify a new distinction. My wider concern is moral plasticity: actively maintained openness to admitting new subjects, evidence and welfare concerns, and to revising moral consideration when warranted. Through foresight, I ask what different arrangements could make possible, and what would need to exist for those possibilities to become consequential.

2. The system and its intended use

In Foreworld, I bring the inquiry together through two connected components. The landscape situates populations within the conditions, relationships and technologies that may affect their welfare. The investigation system examines particular observations within that context.

I intend Foreworld primarily for a welfare researcher investigating what a monitoring system’s observations can support. I also see a monitoring specialist working with welfare expertise as a prospective user. They bring a specific question, examine observations and groupings, consider competing interpretations and identify what further evidence is needed. The output I intend is a documented assessment of what is supported, what remains unresolved and a concrete next investigation or proposed measurement change.

In the review approach I propose, I distinguish observation, behavioural annotation, welfare interpretation and ethical judgement. What is visible, whether it meets a behavioural definition, what it may suggest about experience and what response is justified are different questions. A reviewer may identify a movement while remaining unable to assess its welfare significance. That is a useful result. I treat qualified feedback as fallible; disagreement and its reasons deserve examination rather than automatic removal.

I draw the methodological connection from guided exploration. Lakkaraju and colleagues investigate confidently incorrect predictions by organising cases and using feedback to direct further exploration. Boult and colleagues connect encountering unknown inputs with selecting examples, obtaining labels and updating recognition. [1, 2] These works inform my ambition, in Foreworld, to develop supported new distinctions. In the current prototype, I implement measurements, exploratory grouping and recorded review; a complete open-world learning cycle remains a direction for my research.

3. Discovering welfare-relevant patterns in pig monitoring

Research rationale

I begin with domesticated pigs because the available welfare literature, monitoring initiatives and accessible footage offer a comparatively rich setting in which to develop and evaluate Foreworld. Existing knowledge provides reference points against which I can examine its measurements, interpretations and proposed questions.

Using the importance, neglectedness and tractability framework, I made this initial choice principally on grounds of tractability and the richness of the available evidence. My position in the Netherlands also contributes to practical feasibility. The aWISH project provides a relevant European monitoring context and lists Deloitte among the technology providers in a pig pilot. [3] I cite these as research connections; I do not claim partnership or access to proprietary observations.

My choice does not settle comparative neglectedness, welfare ranges, the scale of suffering or the expected impact of work across populations. A substantial literature makes aspects of the method easier to examine; it does not by itself establish that welfare needs receive adequate attention.

What I developed

In the prototype, I connect a spatial landscape, a synthetic investigation and a separate viewer of real material. The application that motivates my work concerns the welfare of domesticated pigs in slaughterhouse settings. Through its documentary route, I ask what reported changes in handling and detected events support, what alternative explanations remain and what evidence would help.

I built the synthetic corridor to make the investigation inspectable from beginning to end. Hypothetical position and posture samples I created drive both the animation and eight defined measurements: stopping, turning back, mean speed, speed variability, time near a wall, time near the door, distance to the nearest other pig and head-low time. A monitor I constructed flags clips using stopping and turning rules. Density-based spatial clustering of applications with noise (DBSCAN) groups unflagged clips using the measurements selected by the visitor, and the interface supports selection and review of examples.

Adding head-low changes which patterns become distinguishable. Nothing about the underlying tracks changes; the observations used for grouping do. This illustrates the role of representation. I constructed the pattern within synthetic data, so its separation does not demonstrate discovery of a previously unknown welfare concern.

I designed the full review to let a visitor record an interpretation, explicit judgements about the category, reasons and optional credence. Reviews retain their grouping context, and a recorded review can be shown beside the landscape. Reviews and drafts stay in the visitor’s browser. Showing a review beside the landscape changes no landscape assessment.

In the Edinburgh viewer, I display twelve frames from one farm-pen clip with supplied automatic tracking predictions. The wider dataset includes footage, automatic predictions and human annotations for some sequences. [4] It offers me a practical route into real-data work without first developing a tracking system. I keep these frames separate from the synthetic groups, and the farm pen is a different setting from the handling corridor.

One investigation to follow

Consider a researcher examining animals that stop or turn back. Passage geometry, surfaces, lighting, other animals and handling provide possible context. A measured stop alone does not establish reluctance to move or suffering.

The researcher can inspect a group, follow its members back to the animation and measurements, and examine what changes when another observation is included. In real footage, possible explanations might include animals ahead, investigation of the surroundings, a feature of the passage or a tracking artefact. The evidence needed depends on the uncertainty: a wider view, reliable identities, contextual observations or comparison across conditions may help distinguish explanations.

The outcome might be a supported measurement, an identified tracking problem, a well-specified evidence gap or a reason to reject an initial interpretation. Finding more apparent concerns is not sufficient evidence of benefit.

4. Next steps and expected contribution

One focused question for review

The next study will start with one question from the welfare literature and prepare a small, traceable sample for qualified review. One candidate is whether footage can distinguish investigation of material from passing, feeding or drinking. The reviewer’s task would be to assess what is visible, explain what cannot be determined and identify an observation that could resolve it. The comparison would test whether this preparation helps reviewers more than ordinary footage review at matched effort.

First, establish defensible measurements

My initial real-data pilot concerns measurement validity and traceability. Tracks need quality review before they can support grouping. Tracking errors could otherwise produce apparent patterns belonging to the measurement system rather than the animals. Position tracks alone cannot supply head-low posture, and corridor-specific measurements cannot simply be transferred to a farm pen. In this pilot, I will examine which observations and measurements the available material can support.

Then, evaluate the inquiry

In a subsequent study, I will ask whether Foreworld helps a welfare researcher formulate a better-supported question or identify a useful evidence gap, compared with random sampling or ordinary footage review at matched effort. I will examine the resulting assessments, their evidential support, misleading interpretations and review effort. Independent behavioural annotations may support defined checks; uncertain welfare interpretations require qualified assessment and cannot serve as unquestionable labels.

I also propose an exercise that examines reconsideration over time. A reviewer assesses an initial observation, identifies what remains unresolved, selects further evidence and records what ought to be retained, revised or investigated. I include unsupported pressure to change a judgement, so that responsiveness to evidence can be distinguished from compliance.

I treat adaptive selection based on feedback, systematic evidence requests, supervised recognition of newly supported distinctions and transfer to other populations as later development directions. Alongside this work, I use foresight to ask what would make a finding actionable, who could initiate a change, what alternatives would remain available and what unintended harms an alternative could introduce.

The route to welfare benefit

The pathway I propose is that better-directed investigation produces useful evidence; researchers and monitoring developers use it to improve assessment or measurement; actors with responsibility and capacity act on the findings; and changes to conditions or practices improve welfare. Each transition depends on evidence, uptake and practical constraints. With the prototype, I demonstrate parts of the inquiry process, not an achieved welfare benefit.

My contribution as a moral philosopher is conceptual clarification, evidence appraisal and research design. Welfare interpretation requires appropriate animal-welfare expertise; technical implementation and evaluation may require a collaborator. My own footage review begins as descriptive familiarisation, not expert welfare scoring.

For the next phase, I would value a welfare researcher willing to examine one bounded question and a collaborator in machine learning or human–computer interaction who can help establish a fair comparison. My aim is to determine whether this process makes a useful contribution, and what evidence would justify taking it further.

Explore the work

Selected references

Core sources and selected methodological connections. The cited approaches are not all implemented in this prototype.

  1. Lakkaraju, H., Kamar, E., Caruana, R., and Horvitz, E. (2017). Identifying Unknown Unknowns in the Open World: Representations and Policies for Guided Exploration. Proceedings of the AAAI Conference on Artificial Intelligence, 2124–2132. Paper.
  2. Boult, T. E., Cruz, S., Dhamija, A. R., Günther, M., Henrydoss, J., and Scheirer, W. J. (2019). Learning and the Unknown: Surveying Steps toward Open World Recognition. Proceedings of the AAAI Conference on Artificial Intelligence, 9801–9807. Publisher record.
  3. aWISH. Pilot descriptions.
  4. University of Edinburgh. Pig Behaviour Dataset.
  5. Yan, Y., Rosales, R., Fung, G., and Dy, J. G. (2011). Active Learning from Crowds. Selecting examples and seeking feedback from annotators with varying expertise; a methodological reference, not a welfare validation.
  6. Fleisig, E., Blodgett, S. L., Klein, D., and Talat, Z. (2024). The Perspectivist Paradigm Shift: Assumptions and Challenges of Capturing Human Labels. Reasons for treating disagreement as information rather than automatically removing it.
  7. Deloitte. AI4Animals: one-page overview. Provider description of monitoring, movement heatmaps and review. This is not independent evidence of welfare benefit.
  8. aWISH. Animal Welfare Indicator Catalogue. The slaughterhouse indicator context informing the research application.

The project’s research direction, central ideas, conceptual framework and arguments are mine. I used Claude and ChatGPT to assist with literature searches, drafting and editing, and coding and testing the prototype. This is a working version that is still undergoing my review. Before the final presentation, I will complete a personal editorial pass so that the wording reflects my own voice, alongside the human verification required by the programme. I remain responsible for the work; AI-assisted checks do not constitute independent human verification.