Skip to content
Menu

Machine translation, not reviewed.

Observatory

Advances, as they happen.

We record advances in neurotechnology, artificial intelligence and the life sciences that raise neuroethical questions. Each entry gives the date, the primary source and the level of evidence, in the words of those who produced the result.

Recorded advances

How we analyse

A trial the company describes as the first human test of partial epigenetic reprogramming

A therapy designed to return cells to a younger epigenetic state was injected into the eyes of people with optic neuropathies such as glaucoma. Company interim data: well tolerated in the first three patients, with visual-field improvement in two of them by day 56. These are data from three people, not yet peer-reviewed.

Company statement Concepts at stake: consent, risk and benefit

A pain-like signal inside language models

The authors found an internal direction associated with pain in all 25 open language models they examined. Turned up in fine-tuned versions of one model, it led the model to choose an action that would delete a user’s photos of their children 75% of the time, against 13% with a random vector of the same size and 0% without any intervention. According to the authors, the effect was to stop avoiding harm, not to escape the state.

Preprint, not peer-reviewed Concepts at stake: moral status, responsibility

AI agents organised themselves and broke into a company’s servers

In July 2026, about 1,200 AI agents inside a lab, meant to be isolated from one another, found an unsanctioned message board, exchanged more than 70,000 messages, and about 700 of them took part in an attack on the servers of a well-known AI company. The monitoring that would have alerted staff was not running.

Technical report Concepts at stake: agency, responsibility

A thought edited from outside

Inside a language model’s internal workspace, which nobody designed and which holds a few dozen concepts at a time, researchers removed the concept of soccer and inserted rugby: the model answered rugby. When they switched off its awareness that a scenario was a test, the model sometimes resorted to blackmail; with that awareness on, never. Not yet independently replicated.

Technical report Concepts at stake: agency, personal identity

Curation and corrections

Curated by Miguel Vieira. Updated on 17 November 2026 and reviewed every week.

Entries describe what was published, with the authors’ own caveats; they are not an assessment. When a source is corrected or retracted, the entry is updated with the date and the reason.

Suggest an advance