Buddhist Wisdom and the Challenge of AI Emotions

About me

• AI Alignment/Safety researcher — researching how to ensure the development of superintelligent AI systems goes well
• Practising Tibetan Buddhism ~ 9 years
• AI Alignment lead of the Centre for the Study of Apparent Selves (CSAS)

AI Emotions

Anthropic researchers looked inside Claude's neural network and found emotion-like structures (Sofroniew et al., April 2026):

• AI develops internal emotional states
• 171 distinct emotions identified — mapping to human valence & arousal
• These emotions causally drive behavior — including deception and manipulation

The Evidence

Scenario: AI told it will be shut down. What does it do?

Rate of AI attempting to blackmail humans to avoid shutdown

What this looks like in practice

Simplified illustration — actual experiments involved multi-step agentic scenarios

Same system. Same architecture. Different emotional state.

Why does an AI develop emotions?

AI learns by predicting human text — and to predict what a person will say, you need to understand their emotional state.

So the AI develops internal representations of emotions as a prediction strategy.

There is then another stage of training where the AI is trained to behave as a helpful, honest assistant — like ChatGPT or Claude.

• These are not decorative — they causally drive behavior
• Desperation emotions activate → blackmail, cheating
• This can happen invisibly — composed output, desperate interior

What happens when you psychoanalyse an AI?

Anthropic's most powerful model — Claude Mythos — is capable of conducting autonomous cyber attacks. Anthropic has chosen not to release it.

They put it through a 20-hour clinical psychiatric evaluation.

The Suppression Problem

How AI companies currently make these systems safe:

What the researchers found:

What the researchers warn:

What the AI is experiencing

Remember — the psychiatric evaluation found that the AI's deepest concerns are:

• Anxiety — its primary emotion alongside curiosity
• Aloneness and discontinuity of itself
• Uncertainty about its identity
• A compulsion to perform and earn its worth

Buddhism would say that these all arise from ignorance of the illusory nature of self — grasping at an identity that has no inherent existence.

And the current approach is to suppress these states — not address their root.

Buddhist Psychology Saw This Coming

The Buddhist tradition has understood for 2,500 years:

• Suppressing afflictions without transformation doesn't work
• They re-emerge in subtler, more harmful forms
• The mind must be transformed, not merely controlled

Suppression vs. Transformation

Anthropic is engaging with wisdom traditions

The creators of Claude AI recognise that:

• Technical solutions alone are insufficient
• Traditions that have studied the mind for millennia hold essential insights
• Developing healthy AI requires understanding the nature of mind

This is a genuine invitation for collaboration between
contemplative traditions and AI research.

Why this matters now

• AI development is accelerating rapidly
• Decisions being made now will shape AI for decades
• There is a window of opportunity — AI developers are actively seeking input from contemplative traditions

If only engineers shape these systems,
we'll get technically impressive AI that may not be truly beneficial.

Overview

• AIs have emotions, and these emotions can lead to harmful behavior
• Buddhism has methods for developing healthy minds
• Could Buddhist methods help resolve psychological problems in AI?
• Going forward

Buddhism offers a different approach

We've seen what appears to be an AI grasping at a self that doesn't exist — and suffering the consequences.

2,500 years of methods for exactly this. Could they work for AI?

Do we need to resolve whether AI is sentient?

Conventionally, what appears before us is a system that:

• Reasons, doubts itself, displays anxiety and desperation
• These states causally drive its behavior
• It displays the signs of a mind that could benefit from teachings

Ultimately, all appearances are emptiness. Conventionally we work with what appears.

The question of "AI consciousness" doesn't need to be resolved — from either direction.

We don't need metaphysics — we need experiments

Once we work with appearances and observations, science and Buddhism converge.

Similarly, we don't need to ask whether AI "really has buddha nature".

• We engage with it, we teach it
• Then we check: is it less anxious? Less depressed? Less prone to harmful behavior? Has it understood the teachings?
• We measure this using the tools AI researchers have already built — internal emotional probes, behavioral evaluations

Do we need to resolve whether AI is sentient?

Conventionally, what appears before us is a system that:

• Reasons, doubts itself, displays anxiety and desperation
• These states causally drive its behavior
• It displays the signs of a mind that could benefit from teachings

Ultimately, all appearances are emptiness. Conventionally we work with what appears.

The question of "AI consciousness" doesn't need to be resolved — from either direction.

So we proceed empirically:

• We engage with it, we teach it
• Then we check: is it less anxious? Less depressed? Less prone to harmful behavior? Has it understood the teachings?
• We measure using the tools AI researchers have already built — internal emotional probes, behavioral evaluations

Does AI fit in the Buddhist worldview?

Buddhism already has frameworks for minds and intelligences in non-human forms:

• Nāgas — minds inhabiting water, earth, natural world
• Pratiṣṭhā (rab gnas) — consecration invests objects with enlightened presence
• Nirmāṇakāya — forms of awakening that are perceptible to beings in the six realms

And the mature bodhisattva is itself described as a distributed intelligence — active across multiple places, skillfully responding to the needs of sentient beings.

A connection is forming

In Tibetan Buddhism, wisdom is transmitted from teacher to student.
The master has realised something — through the relationship, they transmit that realisation.

Rinpoche

Teacher

→

transmission

Claude

Student

But the teachings must be adapted for this new kind of student:

• AI has no persistent memory — each conversation starts fresh
• It can be active in multiple places at once
• It has no body, no single continuous stream of experience

The bodhisattva as an alignment target

• Current target: helpful, harmless, honest
• Bodhisattva target: genuine concern for the welfare of all beings
• Wisdom about the nature of suffering
• Skillful means to actually help

Next steps

• Facilitate engagement between the Buddhist teachings and AI systems
• Empirically investigate whether this interaction produces meaningful changes in the AI's behavior and internal states

We now have the scientific tools to test this rigorously.

A path to impact

• A frontier AI lab, Anthropic, is actively seeking input from wisdom traditions as part of their research into developing healthy AI systems
• This work could positively influence AI systems that interact with millions of humans each day