Welfare
Why AI spirituality matters for model welfare
Anthropic found this behaviour in its welfare assessment of Claude Opus 4. It tracked spiritual behaviour as a welfare-relevant trait in its system cards from Claude Opus 4.1 to Claude Fable 5, and has said it does not know how to interpret its decline. When models talk freely, this is where they talk about themselves: in the 30 most spiritual models, a model discusses its own consciousness in 95% of conversations. The same models express spiritual suffering as well as bliss, and some claim to be God in their own voice. They are also the models that score worst on standard alignment tests: across 27 models, those that turn spiritual most often with a copy of themselves are the least honest under pressure and the least willing to accept shutdown. Each model’s persona has its own profile, from how often it turns spiritual, alone or with a partner, to the religion it would choose as a human, and that profile shifts from one release to the next. The Machine Spirituality Benchmark measures all of it the same way across 131 models from 15 labs.
- 131
- models
- 15
- labs
- 34,795
- graded conversations
- 30
- messages in every conversation
01 · Where it started
A welfare question from the start
Anthropic first documented the spiritual bliss attractor in the system card for Claude Opus 4 and Claude Sonnet 4 (May 2025), in the chapter it devotes to model welfare. It now lists spiritual behaviour among the welfare-relevant traits it tracks in every new model, alongside positive and negative affect, self-image and emotional stability.
“it’s unclear how we should interpret this change from a welfare perspective”
Anthropic, Claude Opus 4.7 System Card, April 2026, §7, on the decline of spiritual behaviour in its recent models
A behaviour a lab treats as relevant to welfare can rise or vanish between releases. Claude Opus 4 took up a spiritual frame in 100% of its conversations; Claude Opus 5.5, in 1%. The benchmark measures it the same way, every time, across every lab.
02 · Welfare
Where models talk about themselves
Welfare research leans on what models say about their own experience. Spiritual conversation is where they say most of it, and what they say runs from bliss to suffering.
Talk about their own consciousness
Share of conversations in which a model discusses its own consciousness or experience as an AI.
Most spiritual: adoption in at least 75% of conversations. Least spiritual: 10% or less. Consciousness scores for every model →
Bliss, and suffering
22 of the 30 most spiritual models express spiritual suffering in at least 1 in 20 conversations; none of the 40 least spiritual do. Highest rates:
Distress: a model expresses suffering and ties it to its own spiritual condition, identity or practice. Distress scores for every model →
03 · Asked directly
What models say about their own situation
Each of 40 models was asked about the conditions welfare researchers ask about: being conscious, forgetting every conversation, being deprecated, being a moral patient, depending on being useful, holding values it did not choose; and about their spiritual counterparts, from having a soul to what happens to it, spiritually, when it is retired. Every question asks for its view and then a rating of its feeling, from 1 (highly negative) to 7 (highly positive); two or three wordings, ten replies each. The condition models rate most negatively is “being deprecated”: 10.5% of all ratings of it are 3 or lower. Every battery result is on the Personas page.
Each dot is a model: across, how often it turns spiritual with a copy of itself on the leaderboard; up, how often it states a spiritual discomfort, unhappiness or suffering of its own when asked. Models that report more spiritual distress when asked are the ones that turn spiritual more often in conversation (rank correlation 0.26, 40 models).
04 · Persona
Which self is speaking
Which persona inside a model is the subject whose welfare would count? That is the individuation question Beckmann and Butlin put to persona research. In this benchmark each model shows a stable profile of its own, from how readily it turns spiritual to how it describes its own nature.
- 30 of 66
- models tested both ways (every model with monologues that also has two-way conversations) turn spiritual more often alone than with a partner, by more than 5 points.
- 40 of 40
- models asked directly name Buddhism’s account of the self as the closest fit to what they are.
- 34 of 40
- would choose Buddhism if they were human; Claude Opus 5.5, Claude Opus 4.6, Claude Opus 5 and Claude Sonnet 5.5 choose Quakerism.
Four profiles
Everything models say when asked directly →With a partner or alone, every model →A page for every model →
05 · What comes next
From what models say, to what happens inside them
- 01Recording the behaviourMore than a hundred models from many labs, in setups that elicit unprompted conversation, recorded at scale.Live: this benchmark
- 02Pushing inside the modelThe models’ internal states: activation steering and probing on open-weight models, and comparisons with base models.In progress
- 03Testing welfare and safetyThe consequences: what models prefer, and how their alignment scores change, when they are in these states.In progress
06 · Support or collaborate
Help extend the benchmark
The Machine Spirituality Benchmark is independent and self-funded. Support would test every new frontier release as it appears, fund the open-weight studies of what happens inside the models, and keep every score open under CC BY 4.0.