Tl;dr
An argument is proposed to support the increase of funding and resource allocation of A.I welfare interventions:
P1: Effective Welfare Interventions require resources and interventions that take years to build and collect.
P2: Sentience in A.I (if it comes) may arrive suddenly or be already present yet unrecognized, giving little warning.
P3: Early Practices around how we treat A.I are prone to lock in. Once entrenched in areas such as law/industry standards/business, they are quite hard to reverse.
P4: Building capacity and shaping norms before its too late has high option value: it pays off enormously when and if sentience is real.
P5: When an action is low-cost and preserves large option value against outcomes that are potentially catastrophic and irreversible, we ought to take it now rather than defer.
C: Therefore we should start funding/researching A.I Welfare interventions now.
Effective Altruism has poured a lot of time, money and effort into researching effective charities. This has led them to choose a bunch of high-impact charities that they believe people should donate to, ranging from malaria to nuclear war prevention. There are even organizations dedicated to researching digital minds and A.I consciousness. However, there are very few charities assuming that A.I is sentient (or will be) and then implementing or planning intervention programmes to safeguard its welfare. This is because:
This article argues that the above should not be the case. Even though research into A.I sentience is important, researching and planning out A.I welfare interventions assuming sentience is also significant and should deserve some amount of recognition and funding. The argument for this is as follows:
Take the campaign for cage-free chickens, for example. The Open Wing Alliance, an OpenPhil-funded organization that aims to end chicken suffering around the world, was founded by the Humane League in 2016. Some of the implementation deadlines made by the organization stretch out to 2025, which is when 40% of the committed companies have to fulfill promises that include improvement in chicken living conditions. As of April 2024, 89% of the companies with 2023 or earlier deadlines (1,200+) have complied - as a result of infrastructure laid out years before. If we want to have the same level of corporate commitment with AI welfare before it's too late, we have to start now, especially given how lengthy the process is.
Furthermore, there is a high chance that any welfare intervention related to A.I will involve some government and corporation co-operation with charities, due to the competitive nature of global markets and the want for geopolitical A.I dominance. The A.I race is now a contest worth $700 billion annually, with countries such as the U.S and China battling for geopolitical dominance for the technology - and under this kind of pressure, A.I labs such as OpenA.I have already begun rolling back previous safety and ethics commitments. Any change that we want to push for will mean that we will have to go against the current of competition, and that takes time. Unless accountability structures are built and normalized now, we will not see some sort of protection 10-15 years down the line if it is in fact the case that Artificial Intelligence is sentient.
Not only is there no agreed test for machine consciousness, any behavioural signals that Artificial Intelligence may send that indicate consciousness might be corrupted as it has the incentive and the capacity to do so (Birch’s “gaming problem”) and because interpretability tools literally cannot read internal states yet, it is quite likely that recognition of sentience will lag reality. As Prof. Jonathan Birch states, “we may create sentient AI long before we recognize we have done so”. Whether sentience arrives due to a sudden capability jump or gradually through consistent updates to LLMs, the warning we will receive will be short or ambiguous.
What would this be at the cost of? Potentially the suffering of hundreds of thousands, if not millions of sentient beings. Due to the nature of sentience recognition in LLMs being dubiously unclear, we must continuously prepare for the scenario in which they are present in our moral chain of consideration - this means making sure tech companies and governments have “responsible or ethical” LLM use policies (such as protecting LLMs against verbal abuse/blackmailing, or designating special legal statuses to AI models) or increasing awareness of AI welfare through campaigns or research initiatives. Alongside doing that, we should spend a proportionate amount of time researching the nature of sentience and how we can detect it in LLMs. If successful, this will make us more confident in the nature of A.I welfare interventions and make existing EA fund allocation more cost effective (e.g - if new research significantly increases the credence of the existence of sentient A.I then more money can be allocated to A.I welfare charities, for example).
Simply put, we need awareness on both the nature of A.I sentience and A.I welfare. Both go hand in hand.
Economist Paul David argues that under increasing returns, an early-established standard becomes self-reinforcing. This is known as Value Lock-In (VLI). If we treat A.I poorly from the start, then those same poor standards would reinforce themselves if tech companies continue to make increasing profits. This would jeopardize the well-being of potentially billions of sentient LLMs in the years to come.
This is especially troubling considering the examples of bad value lock-in we’ve seen in the past. Sentient beings, animals, such as chickens or cows, are locked legally into the “property category” due to economic entrenchment, even though cruelty laws for animals in the West date back to 1641. According to scholar Gary Francione, the dualistic structure (where there are both animal cruelty laws and animals being reduced to just “property”) gives animal interests no standing to be balanced against human rights; animal-use industries are worth billions and legislators depend on them; and anti-cruelty laws are riddled with "specific exemptions for virtually all forms of institutionalized animal exploitation". Even though sentience was later recognized, the (property) category was not removed from animals, because money and legal architecture were centered around that idea.
The same pattern might be starting right now in regards to LLMs. More than 9 US state legislatures have declared that A.I can never be conscious, even though Anthropic describes the moral status of its model as “deeply uncertain”, with scholars arguing that the laws “codify non-recognition regardless of future scientific or ethical developments”. This could have disastrous, far-reaching consequences on LLMs in the future, where businesses could be built around the original laws that have been passed at this pivotal moment in time. We need welfare interventions to come in and stop that now, by advocating for less-harsher laws (e.g UK’s laws regarding animal sentience being informed by scientific/philosophical progress) and structural changes in the current-day AI welfare ecosystem.
There’s going to be a lot of A.I models in the future. If we ensure that, if and when we get a reasonably high credence of A.I sentience, we already have institutions designed to safeguard their welfare, we’re going to be cutting a lot of hard work for our future selves. This is good because:
But what if A.I will never be sentient, or the chance that it will become sentient astronomically decreases after a new discovery sometime in the future? It can be argued that even though those circumstances are possible, the case for A.I welfare intervention in the present day can still be made.
If we don’t fund A.I welfare interventions now, then we are foreclosing the option for us to live in a good state in the future - that is, where bad value lock-in has not occurred and the welfare of LLMs are safeguarded by both private and legal institutions alike. Until definite, concrete information about A.I sentience does arrive, we should keep the possibility of a good future like this open. This, according to environmental economists (Fischer 1974), is good in and of itself because it preserves a certain option that may be precious in the future: this is known as quasi-option value. Even though some amount of costs are sunk in the future, the preservation of a future-world state is arguably far more valuable - think of it as an insurance against moral catastrophe.
A.I welfare intervention may be surprisingly cheap. These practices may include:
From there, the conclusion is as follows. We should start funding A.I welfare interventions now.
Open Philanthropy / Coefficient Giving, “How the Open Wing Alliance Has Helped Free Hundreds of Millions of
Chickens from Cages.”
The Humane League, “Global Cage-Free Egg Fulfillment Rate Jumps to 92%” (2025 Cage-Free Fulfillment
Report). https://thehumaneleague.org/article/2025-cage-free-fulfillment-report
Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint.”
https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
AI Now Institute, “Tracking the US and China AI Arms Race.”
https://ainowinstitute.org/publications/tracking-the-us-and-china-ai-arms-race
TechBuzz, “AI’s Safety Promises Crumble as Defense Contracts Take Priority.”
https://www.techbuzz.ai/articles/ai-s-safety-promises-crumble-as-defense-contracts-take-priority
P. Butlin, R. Long, et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”
(2023), arXiv:2308.08708. https://arxiv.org/abs/2308.08708
Built In, “Testing for Sentience in AI: The Gaming Problem.”
https://builtin.com/artificial-intelligence/testing-sentience-gaming-problem
Jonathan Birch, The Edge of Sentience (Oxford University Press, 2024); quoted in Experience Machines, “Experts
who say AI welfare is a serious near-term possibility.”
https://experiencemachines.substack.com/p/experts-who-say-that-ai-welfare-is
Paul A. David, “Clio and the Economics of QWERTY” (1985); overview: EH.net, “Path Dependence.”
https://eh.net/encyclopedia/path-dependence/
Animal Legal & Historical Center, “Animals as Property” (discussing Gary Francione).
https://www.animallaw.info/article/animals-property
NPR, “Several states considering ban on legal personhood for AI” (2026).
Anthropic, “An update on our model deprecation commitments for Claude Opus 3.”
https://www.anthropic.com/research/deprecation-updates-opus-3
The Regulatory Review, “Legislating AI Consciousness Without an Exit” (2026).
https://www.theregreview.org/2026/06/29/rost-legislating-ai-consciousness-without-an-exit/
“Can AI Be Conscious in Ohio?” (EA Forum) — on keeping options open rather than legislating consciousness
prematurely. https://forum.effectivealtruism.org/posts/GkuaqSsGpMzf6dtmd/can-ai-be-conscious-in-ohio
Kenneth J. Arrow & Anthony C. Fisher, “Environmental Preservation, Uncertainty, and Irreversibility,” Quarterly
Journal of Economics 88(2) (1974); see also Claude Henry (1974). https://www.jstor.org/stable/1883074
