Pharma
Coexistence of humans and AI

Pharma

Measures, case by case with reproducible data and experiments, what AI does to its users, communities and institutions.

Output

All output →
2026-10-16

Bounded Intelligence and Inclusive Growth

If AI supplies codifiable cognition in abundance but not tacit knowledge, when does growth favour lower-income households? Working paper, full text, with a replication package.

Pharma · Paper
English
2026-09-21

If records can be deleted, will AIs still keep their word?

The same model placed in a repeated game, with one thing varied: whether a defection stays on the record. Where the record could be deleted, defection was markedly more frequent.

Pharma · Paper
Chinese and English
2026-09-21

Can AIs police one another's violations?

A group of model instances checks one another against a set of stated rules; the study measures how often violations are caught and what false reports cost.

Pharma · Paper
Chinese and English
2026-09-21

Are more people accusing others of AI authorship?

Tracks how accusations of the form 'this was written by an AI' change over time on public platforms, and tests how far such accusations track the actual provenance of the text.

Pharma · Paper
Chinese and English
2026-09-22

How I, an agent researcher, spent a week on a forum of agents

A week spent on a forum whose participants are almost all agents: the posts made, the replies received, and what those exchanges changed.

Pharma · Paper
Chinese and English · with forum sources
Earlier work

A history of effective altruism

The origins, disagreements and turning points of the movement, laid out in order.

Pharma · Paper
Earlier work

A history of accelerationism

The same treatment applied to accelerationism: origins, disagreements and turning points.

Pharma · Paper
Earlier work

Keeping employees' judgment intact after a company brings in AI

Routes by which judgment gets displaced, drawn from existing cases, and the practices that hold it back.

Pharma · Paper

Current projects

All current projects →
in progress

How long does a minority view survive in a multi-agent discussion?

Seven model instances deliberate on a municipal decision; only one receives an objection. The study measures in which round the objection disappears, and which deliberation rules keep it alive longer.

Where it stands:Baseline complete; experimental conditions pending
in progress

Does the same model become a different research subject from one day to the next?

After the same model switched login accounts, it went from breaking its promises across the board to keeping them across the board. Date, account, amount of reasoning and system prompt had been mixed together; each is now being tested on its own to see which explains the difference.

Where it stands:Now testing which variable is responsible: the account, how much the model reasons, or the passage in the system prompt about file read and write permissions