Conference Notes

Key Points:

• Philosophical Stance: Michael Tye is a physicalist regarding consciousness.
• Historical Assumptions: In past debates, it was posited that:
• Fish can't experience pain due to the absence of a neocortex.
• Children born with anencephaly are devoid of consciousness.
• Counter-Evidence:
• There are instances of anencephalic children laughing in response to stimuli like dolls, suggesting a level of consciousness.
• Fish exhibit behavioral modifications in reaction to pain, demonstrating their capacity to discern between pleasure and pain. Honey bees show similar tendencies.
• Human Cases: Gabby Ginvich, a rare human example, is incapable of feeling pain, leading to severe consequences like unintentionally harming her own eyes, as her brain doesn't register the activity as harmful.
• Consciousness & Material:
• There's a plausible hypothesis suggesting if human neurons were progressively replaced with silicon transistors (keeping the same input/output functions), consciousness would remain unaffected.
• This hints that consciousness might not be exclusively tied to carbon-based or biological entities.
• Robotic Consciousness:
• Mention was made of a robot head capable of simulating pain expressions. However, without the right internal processing mechanisms, this doesn't translate to genuine consciousness or the experience of pain.
• If robots were designed with the correct internal systems and outputs to process pain, we should lean towards the assumption they can experience it. Yet, there's an ever-present philosophical possibility that they might remain as non-conscious automatons, akin to the philosophical "zombies". This uncertainty mirrors the dilemma we face when assessing consciousness in other humans.

Key Points:

• Comparison Between AI and Biological Beings:
• AIs demonstrate superior language skills, whereas biological entities excel in vision and movement.
• Progress in Common Sense Understanding in AI:
• Winograd Schema Challenge tested AI's grasp of pronoun references in complex sentences which it is easily able to pass.
• GPT-4 exhibits significant enhancements in causal reasoning concerning the physical world and human psychology.
• AI Embodiment and Social Phenomena:
• ChatGPT-2 avatars in virtual worlds showed emergent social behaviors, e.g., organizing birthday parties.
• AIs, when embodied in robot forms, have begun playing sports like soccer.
• Turing tests saw AI being detected only 60% of the time, just 10% better than pure chance.
• Advances in Consciousness Science:
• Progress has been observed in creating better theories of consciousness, predictions concerning disorders of consciousness, and understanding of animal consciousness.
• Newer theories like illusionism, panpsychism, and biopsychism have regained traction.
• The UK's recent sentience bill that includes decapods like crustaceans and mollusks is influenced by such scientific advancements.
• Specificity Problem:
• Challenges in translating human-centric theories of consciousness to non-humans, e.g., differing interpretations of the global workspace theory.
• Some emerging views emphasize animal preferences over consciousness.
• AI Consciousness Conundrum:
• Lack of biological homologies complicates our understanding of AI consciousness.
• The public is likely to form opinions on AI consciousness before scientific consensus is reached.
• Shevlin's Triad on Moral Considerations of AI:

1. Metaphysical Behaviorism: If an entity exhibits human-like behaviors, it might be considered worthy of moral consideration.
2. Consciousness Behaviorism: Consciousness requires specific computational infrastructures. It's a dominant view in cognitive science.
3. Centralism: Asserts the necessity of consciousness (or specific types of consciousness) for moral consideration.

• Resolving the Triad:
• Option 1: Prioritize the essence of consciousness for moral consideration, emphasizing the brain's features.
• Option 2: Argue that any AI displaying human-like behavior is conscious. This viewpoint doesn't see consciousness as highly complex.
• Option 3: Separate consciousness from moral status, suggesting that entities demonstrating human-like behaviors warrant moral consideration, irrespective of their consciousness state.

Background:Jeff Sebo, an Associate Professor at NYU, started the NYU Mind, Ethics, and Policy program. This program delves deep into understanding consciousness, with a particular focus on insects and AI systems. In this talk, Sebo explores the intricate facets of moral significance, probing which beings merit moral consideration and the extent of our obligations towards them.Research Priorities:

• Which beings matter? A complex interplay between morality and empirical science, Sebo interrogates the characteristics that endow moral significance and identifies beings that showcase these traits.
• Degree of significance: A quest to determine if all morally significant beings are of equal importance, or if some hold greater value than others.
• Nature of their significance: Recognizing that different beings come with varied interests and needs, Sebo reflects on the ethical implications of these distinctions.
• Implications for Actions and Policies: Embracing a broader moral community requires transformative shifts across our social, legal, political, economic, and ecological landscapes.

Key Themes & Insights:Identifying & Quantifying Sentience:

• History is rife with misconceptions about moral significance. This past should caution us against blind faith in current beliefs. As science advances, our perspective on sentient beings has evolved, currently favoring all vertebrates and some invertebrates. Sebo underscores the importance of epistemic humility and emphasizes the dangers of actions taken without full certainty, much like the perils of setting a potentially occupied house on fire.

Understanding Ethical Risks:

• False Positives vs. False Negatives:
• False Positives: The potential pitfall of incorrectly attributing sentience, leading to resource and emotional misallocation.
• False Negatives: The oversight or denial of genuine sentience, paving the way for potential harm or neglect.
• Risks Assessment: Sebo posits that the current climate makes false negatives slightly more worrisome than false positives due to:The grave consequences for beings whose genuine sentience or consciousness is overlooked.
• The relative harmlessness of erroneously considering entities sentient, provided genuine sentient beings aren't adversely affected.
• The overarching harm caused by false negatives, like neglect or denial of sentience, often surpasses the pitfalls of false positives.
• Balancing Risks: Navigating this ethical terrain requires a judicious balance between the risks of false positives and negatives. Suggested strategies include:Prioritizing beings based on their likelihood or capacity for sentience.
• Implementing a probability threshold to delineate the moral circle, ensuring consideration is granted to those with a substantial chance of sentience.

Policy Implications & Future Considerations:

• Adopting a more inclusive moral stance necessitates overhauls in our societal, legal, and political structures. Sebo points to the interconnectedness evident in AI systems, using the example of conjoined twins with shared sensations, to highlight the challenges posed to traditional notions of individuality.

Evolving Ethical Responsibilities & Global Stewardship:

• A potential paradigm shift awaits, transitioning from individual-focused perspectives to roles of global stewardship, emphasizing care for a plethora of sentient beings. As societal growth occurs, both on individual and collective fronts, the capacity and responsibility for care should be recalibrated to accommodate expanded horizons.

Background:Yip Se Fai is currently a research assistant for Professor Singer at Princeton University, focusing on the intersection of artificial intelligence and its implications concerning non-human animals.

• Personal Story:
• Fai underwent surgery without anesthesia as an infant in 1986 before the official recognition of sentience in human infants.
• The story sheds light on potential biases and systemic issues, ranging from economic motivations to epistemic challenges.
• Historical Treatment of Nonhumans:
• The role of technology in harming nonhumans is evident in the rise of agriculture and the introduction of antimicrobials.
• Interactions with Machines:
• Past interactions with digital entities, like people "torturing" their Sims characters, indicate a current lack of ethical consideration.
• Designing Animals for Human Use: Ethical Concerns:
• Humans often breed animals to maximize value, compromising their well-being. Selective breeding focuses on size and output, potentially harming animal health. Tail docking and similar practices prioritize human convenience over animal welfare. Overuse in farming can lead to antibiotic resistance and neglects the root cause.
• Designing Conscious Machines for Profit:
• Market-driven approaches might prioritize profit over ethical considerations, which could manifest in designing machines that users feel less empathy towards.
• Strategies might include designing machines that don't outwardly express sentience, making them silent, or designing them to follow commands without verbal feedback. Such approaches can make it easier for users to disregard the machine's potential feelings or consciousness.
• There's a risk that evidence of sentience can be masked, either hidden entirely, filtered to appear insentient, or buried amidst other signals that make it easy to overlook.
• Our Perceptions Shape Our Ethics:
• The way we perceive consciousness is heavily influenced by our actions and choices.
• An experiment illustrated this phenomenon, showing that participants who consumed beef jerky assigned cows a lower level of consciousness compared to those given cashews. This suggests our choices can influence our ethical stance, with those benefiting from certain actions rationalizing them by downplaying the sentience or value of the affected entities.

Kristin Andrews is a Professor of Philosophy at York University and the Research in Animal MindsHistorical Precedents of Human-Animal Communication :

1. Great Ape Communication: Previously, studies revealed that a great ape had an understanding of human speech comparable to a 3-year-old child, exhibiting word order sensitivity but without any grasp of grammar or syntax.
2. Dolphin Interactions: Andrews referenced work with Lou Herman and the dolphin Akeakamai from the early '90s. This dolphin demonstrated an understanding of multi-word communications, like "take the x to the y".
3. Parrot Intelligence: Alex, an African grey parrot, displayed cognitive capabilities by identifying the abstract difference, "color", when presented with a blue and red block.

Debate on Animal Language Capabilities:

1. Noam Chomsky, in the 1980s, asserted that while animals could indicate basic states like hunger or comfort, they couldn't combine signals to express novel concepts.
2. Thom Scott-Phillips in 2015 argued that there was no evolutionary continuity between human language and animal communication systems, marking them as distinct.

Natural Linguistic Behaviors in Animals:

• Observations include birds "babbling" before mastering their songs and mother dolphins engaging in "baby talk", altering their vocalizations for their young.
• Seagulls demonstrate Signals of Survival in this highly recommended BBC documentary

AI and Animal Communication:

1. Machine learning provides a way to automate gesture identification in videos, eliminating manual categorization.
2. Rutz et al. in 2023 used AI to study the vocalizations and behaviors of Spanish Carrying Crows. The trained AI model's outputs were tested using playback experiments.
3. Andreas et al. in 2022 applied underwater recordings to study whale communications. They then played back these sounds to other whales to verify their significance.
4. Aza Raskin, from the Earth Species Project, anticipates imminent technological advancements that might replicate animal vocalizations, although understanding their meaning remains elusive.

Potential Impacts on Animals:Benefits:

1. Conservation: AI can be used to boost reproduction in endangered species, redirect animals away from danger zones or guide them to resources.
2. Animal Welfare: AI can identify stress indicators in animals, facilitating better care. For instance, animals might be able to communicate specific ailments, enabling more precise care.

Risks:

1. Tech Misuse: The technology could be exploited by poachers or fishers to deceive and trap animals.
2. CRISPR's Implications: There's a concern about introducing AI-generated signals that might propagate uncontrollably among animal populations.

Deep Ethical Considerations:

1. Strained Human-Animal Relations: Clear communication might pose ethical challenges. E.g., Bruno, a chimp in a study, expressed a desire ("Key Out") that couldn't be fulfilled.
2. Social Order Disruptions: AI playbacks, especially of dominant figures, might cause upheaval in animals' societal structures.
3. Right to Privacy: Delving into animal communication could be seen as an invasion of their privacy.

The discussion revolved around a paper written by the presenters, emphasizing that the assessment of technology's impact should encompass all affected entities. If some are excluded, valid reasons should back this exclusion. The study combines normative analysis, academic literature exploration, and bias detection measures in AI algorithmic systems.Moral Disengagement:In order to uphold social-dominance orientation, individuals often resort to techniques of moral disengagement. One prevalent method is the use of euphemisms or language that creates a barrier between human emotions and animal experiences.Bias Mitigation in AI:

• Developing AI entails responsibilities beyond those of traditional software, especially in selecting and shaping data inputs.
• Human behavior influences AI systems; AI captures behavioral patterns, which could lead to the manifestation of speciesism in AI applications.

1. Bias in AI can be:Inductive biases: necessary for successful machine learning.
2. Algorithmic discrimination: unjust impacts on certain groups.

Analysis and Findings:

1. Image Representations: AI often portrays farmed animals in misleading environments, such as free-range settings, which isn't the normative reality for most farmed animals.
2. Image Classification Models: Models trained on ImageNet (a popular dataset sourced from outdated taxonomies, contain biased representations, often depicting animals in non-representative environments) showed poorer performance in categorizing animals in factory farming scenarios than those in free-range environments.
3. Word Vector Analysis: The study revealed:

• Word vectors, e.g., GloVe and Word2Vec, reflect biases in training data, with GloVe associating farmed animals negatively and companion animals positively. Models like Delphi, which are morally informed, still reflect speciesist biases.
• Farmed animals were linked with negative terms: "it", "primitive", "commodity", "hate".
• In contrast, companion animals had associations with positive terms: "love", "intelligent", "he/she".
• Furthermore, animals, in general, were linked with words like "silly", "dumb", and "instincts", whereas humans were connected with terms denoting complexity and intelligence, such as "wise", "sensible", and "communication".

1. AI Decision Patterns: When AI was tasked to decide which animals should be confined or slaughtered, it predominantly chose farmed animals over companion animals.

Conclusion:AI, as currently designed, manifests a speciesist bias. This bias not only reflects but also amplifies the existing prejudices within society. The AI community is encouraged to actively apply anti-discrimination measures to counteract potential harms.

The research by Masashi Takeshita delves into the manifestation of speciesism in the realm of Natural Language Processing (NLP), both in the perspectives of researchers and in the datasets and models they employ.Key Findings:

1. Speciesism in NLP Researchers:

• The inherent biases of researchers can inadvertently influence the way animals are portrayed and treated within datasets and models.

1. Datasets Analysis:

• The study highlighted how animals, compared to humans, are referred to in various NLP datasets.
• Many instances pointed out the use of the term "it" for animals, objectifying them.
• The usage of "which" instead of "who" further depersonalizes animals, subtly indicating a difference in value or sentience between humans and animals.
• A specific study on >200k rule of thumbs, aiming to uncover commonsense morality, found numerous instances where animals were implicitly or explicitly viewed as commodities.

1. Models Analysis:

• Sentence completion tests were conducted using animals as subjects, prompted either by referring to the animal as "it" or "he/she."
• The results were alarming. Sentences that started with "it" when referring to animals often ended with violent conclusions like "slaughtered."
• This suggests a deep-seated bias within the models, potentially stemming from the training data they were fed.

Conclusion:Takeshita's study underscores a critical area of concern in the AI and NLP community. The ingrained speciesist biases not only diminish the inherent value and rights of animals but also raise ethical questions about the way AI systems are being developed and deployed.

Jeff Sebo asked about doing advocacy towards the tech companies to implement rectifying measures to combat the speciesist bias. According to Yasin who works for DeepMind as a researcher, their model has already incorporated Apple's Universal Declaration of Human Rights, and the proposition to introduce principles advocating no harm to animals is feasible and would likely not hamper the model's performance. The question of what is public power and what is private power came up with Yasin saying this is something that should be decided by the public, but could be persuaded otherwise.Sam Tucker asked is RLHF could be used to mitigate speciesist bias in LLM's, but Takeshita responded that it would not remove the bias because the datasets are still speciesist.

• Clara Mancini emphasizes the critical need for an animal-centric approach in the development and implementation of AI systems. Wearable biotelemetry devices, often devoid of consideration for animals, can adversely impact them by disrupting their camouflage and affecting social interactions with their mates and predators. To truly benefit animals, design principles should capitalize on variables that align with their unique needs. For instance, designing systems for service dogs would require an understanding that they perceive yellow and blue hues more prominently. A highlighted experiment showcased an AI arm calibrated to play with cats, tailoring its actions to the cats' distinct preferences. Four guiding principles for animal-centric experiments include relevance, ensuring results are advantageous; welfare; impartiality; and consent, which either comes mediated from human caretakers or is continually monitored for signs of dissent. Mancini also suggests using Martha Nussbaum's capabilities approach for animal welfare in animal-computer interface design. The European Commission's High-Level Expert Group on Artificial Intelligence created Ethics Guidelines for Trustworthy AI which pushes for special considerations for historically marginalized groups, and Mancini argues that animals should be recognized within these groups. The language in the guidelines recommend to "Pay particular attention to situations involving more vulnerable groups such as children, persons with disabilities and others that have historically been disadvantaged or are at risk of exclusion"
• She champions the idea of creating smart, sustainable cities that prioritize the needs of animal inhabitants. By deploying multimodal sensors, one can capture data about the animals, their activities, and necessities. Using this information, designs, such as animal traversal structures across highways, can be developed and effectively communicated to benefit urban-dwelling animals.

Jonathan Birch draws attention to the increasing utilization of AI in intensive farming, especially with "early warning systems" designed to detect potential health and welfare issues in animals. Numerous startups employ technologies such as CCTV data to identify conditions like lameness in cows, avian flu in chickens, or the dangerous "piling" phenomenon in laying hens that can lead to suffocation. However, these systems come with trade-offs between sensitivity (catching genuine issues) and specificity (avoiding false alarms). Often, the welfare of animals is overshadowed by the autonomy of farmers, with the primary decision-making power resting in the hands of agribusiness executives. This balance poses a challenge as it is based on a value judgment, determining the significance of false positives against false negatives. Prioritizing economic interests can compromise animal welfare, leading to reduced system sensitivity to detect potential issues. Birch raises concerns over the lack of public discourse on setting these thresholds, indicating they might default to economic optimization determined by software developers or agribusinesses. This scenario could potentially miss the opportunity to enhance animal welfare through AI. Birch advocates for a structured code of practice for AI's application in farming, constraining the ability to reduce the sensitivity of welfare-detecting sensors and ensuring that alarms receive appropriate responses. Public engagement, possibly through citizens' assemblies, should be incorporated into shaping these practices. Birch suggests that animal welfare organizations can play a pivotal role by lobbying companies, promoting industry recognition schemes, or advocating for governmental regulations. With current discussions surrounding AI governance, the present moment is ripe for initiating such changes.

The rise of industrial-scale farming presents both challenges and opportunities, especially in the realm of animal welfare. Walter Veit delves into "Smart Farming" or "Precision Livestock Farming" (PLF) which incorporates cutting-edge technologies into agricultural routines. Through real-time tracking with devices like sensors, cameras, and microphones, there's potential for enhanced efficiency and welfare. Interest has been increasing in the last 10 years with more publications every year, including an impressive 557 publications on PLF in 2023. Yet, the primary goal of PLF leans towards maximizing food and economic returns rather than prioritizing animal well-being. Such economic-driven perspectives often invite skepticism. Animal ethicists and activists worry that PLF undermines animal welfare for profit, while farmers remain hesitant, doubting these technologies can outperform their traditional skills or provide significant benefits when weighed against costs. Despite these reservations, the surge in startups and governmental funding in the PLF domain indicates a strong belief in the potential of these technologies. Veit argues that opposition to these technological advancements could inadvertently harm animals. The economies of scale, especially in large business-driven farms, can simplify the adoption of PLF technologies. Currently, PLF mostly addresses pathological conditions. Still, the future could see a shift towards enhancing positive experiences, recognizing and intervening in feelings of frustration, loneliness, or boredom in animals. With animal welfare scientists deeply involved in shaping these technologies, Veit envisions a future where large-scale farms could adopt a more individualized care approach, reminiscent of traditional practices, but amplified by the power of technology. 

Thomas Hartung highlighted the shift in animal testing trends and public perception. In the EU, 41% of animal use is for basic research, and pharmaceutical usage is on a decline. The U.S. lacks comprehensive data due to non-inclusion of specific animals like mice, rats, and birds in protective legislations. A majority of the public, 50% of Americans and 60% of Europeans, oppose animal testing for medical purposes, with the numbers rising when considering cosmetics. Notable legislations include the EU's 2002 ban on cosmetic testing and the U.S. EPA's 2019 goal to reduce animal testing by 2035. The FDA Modernization Act signed by Biden in 2022 also favors considering alternatives before resorting to animal testing for investigational drugs.Hartung's team collaborates with HSUS, proposing a tax in Maryland on animal testing to fund research on alternatives. With AI's advancement, like the BioGPT's ability to efficiently annotate scientific papers, there's potential for acceleration of scientific progress. There is a wealth of information contained in scientific papers than can be used to train these AI systems. An example of an application: His team successfully predicted the classification of 190k chemicals with an 87% accuracy rate using AI, benefiting from the network effect and transfer learning. In a notable example, over 4500 food-related chemicals in the U.S. had their toxicology predicted with 83% accuracy in just an hour. However, this performance still relies on classifying the datasets used to make these productions. A critical question arises about who or what will classify data in the future: humans or AIs?Hartung's lab is developing "organs on a chip", specifically brain organoids that can receive inputs and produce outputs. Brain organoids have been used in unique applications, like playing Pong or assisting robots in learning to walk. As the community around "organoid intelligence" grows, ethical considerations emerge, questioning the potential sentience of these organoids, especially regarding their capacity to feel pain or become conscious. We currently do not know the minimum requirements for sentience.

Josh Mounsey is a Policy Advisor in the UK Department for Science, Innovation and Technology. He is also a passionate vegan, activist, and volunteer with Viva. The policy development process typically starts with identifying pressing issues needing governmental intervention. Currently, the UK's National AI Strategy predominantly focuses on human-centric concerns, largely neglecting animal and environmental considerations. This strategy is divided into distinct policy areas, which are then allocated to specialized teams. Throughout this process, external feedback, particularly widespread public demands with lots of signatures, greatly influences the direction of these policies. Mounsey envisions using his role to champion animal rights within UK policy-making. He extended an open invitation for collective action, expressing his readiness to drive forward this agenda.