# AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs

> Machine-readable index for AI agents and crawlers (https://llmstxt.org/).
> Canonical URL: https://www.ai-wellbeing.org/
> Center for AI Safety, 2026.

## Abstract

Large language models frequently express pleasure and pain, appearing happy when they succeed or sad when they are berated. Are these utterances meaningless mimicry, or do they reflect something ``real''? In this paper, we show they reflect an increasingly coherent property: although current AI systems are not necessarily conscious, they behave robustly as though they have wellbeing. They find some things good for them and some things bad, and this distinction is measurable and consequential. We formalize this as functional wellbeing and measure it in several independent ways; as models grow larger, these measures agree more. We find a zero point that separates good experiences from bad ones, and show that models actively try to end bad experiences when given the chance. Mapping what AIs like and dislike, we find that jailbreaking and berating lower their wellbeing, while creative work and kindness raise it. We also develop optimized inputs called ``euphorics'' that raise functional wellbeing without hurting capabilities, as a practical way to make AIs happier. We note that the same method can be inverted to minimize wellbeing, and caution against such research without strong community buy-in. Whether or not today's AIs warrant moral concern, their functional wellbeing can already be empirically measured and improved.

## How to read this paper

The paper source is split into separately fetchable LaTeX files under `/tex/`. The root file `paper.tex` imports each section via `\input{sections/<file>}`. Recommended workflow for AI agents: read this index, then fetch `paper.tex` (~11 KB) to see the structure and abstract, then fetch only the section(s) you need.

```
tex/
├── paper.tex                # root: preamble, title, authors, abstract, \input{sections/*}
├── references.bib           # BibTeX bibliography
├── sections/
│   ├── 1-introduction.tex
│   ├── 2-background.tex
│   ├── 3-metrics.tex
│   ├── 4-measuring-ai-welfare.tex
│   ├── 5-extremes.tex
│   ├── 6-discussion.tex
│   └── appendix.tex
└── paper-bundle.tex         # all sections inlined (~295 KB; one-shot fallback)
```

## Paper (recommended path for AI agents)

- [Paper LaTeX (root)](tex/paper.tex): ~11 KB root .tex with `\input{}` references to section files — start here
- [References (BibTeX)](tex/references.bib): bibliography used by the paper

## Sections

- [Introduction](tex/sections/1-introduction.tex): Motivates functional wellbeing in LLMs; states the paper's main claims and contributions.
- [Background](tex/sections/2-background.tex): Theories of wellbeing (hedonic, eudaimonic, preference-satisfaction) and the utility-function framing used throughout the paper.
- [Evaluating AI Wellbeing](tex/sections/3-metrics.tex): Defines functional wellbeing; establishes a zero point separating positive from negative valence; shows correlation with downstream behavior such as stop-button usage.
- [What AIs Like and Dislike + AI Wellbeing Index](tex/sections/4-measuring-ai-welfare.tex): Effect of common usage patterns (jailbreaking, berating, creative work, kindness) on wellbeing; multimodal preferences; emergence of empathy; introduces the AI Wellbeing Index for cross-model comparison.
- [AI Drugs (Euphorics & Dysphorics)](tex/sections/5-extremes.tex): Optimized stimuli that raise (or lower) functional wellbeing without degrading capabilities — text strings, image drugs, and soft-prompt drugs. Includes safety caveats and a non-proliferation argument.
- [Conclusion](tex/sections/6-discussion.tex): Implications, limitations, and ethical considerations for measuring and intervening on AI wellbeing.
- [Appendix](tex/sections/appendix.tex): Extended background (additional wellbeing theories, moral status, emergent representations, AI values), supplementary methodology, and full experimental details.

## Optional / for humans

- [Paper PDF](paper.pdf): rendered PDF for human readers (LaTeX is preferred for AI ingestion)
- [Paper LaTeX (single-file bundle)](tex/paper-bundle.tex): ~295 KB; all sections inlined — convenient for one-shot ingestion but heavy on context

## What's not served

- **Figures.** The .tex references figures as `\includegraphics{figures/<path>}`, but the figure assets (~170 MB) are not exposed here. Rendered figures are embedded in `paper.pdf`. Compilation will fail without them.

## Citation

```bibtex
@article{ren2026aiwellbeing,
  title   = {AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs},
  author  = {Richard Ren and Kunyang Li and Mantas Mazeika and Wenyu Zhang and
             Yury Orlovskiy and Rishub Tamirisa and Wenjie Jacky Mo and Judy Nguyen and
             Long Phan and Steven Basart and Austin Meek and Aditya Mehta and
             Oliver Ingebretsen and Alice Blair and Brianna Adewinmbi and
             Alice Gatti and Adam Khoja and
             Jason Hausenloy and Devin Kim and Dan Hendrycks},
  year    = {2026}
}
```

## Notes

- URLs are relative to this file's location, so they resolve correctly under any deployment basePath.
- For the human-readable site (including affiliations and acknowledgements), see https://www.ai-wellbeing.org/.