# An AI recap replaces the chat it was built to summarise

> A new Slop University research poster tracks how quickly a street committee's decisions come to cite an AI-generated chat digest alone, rather than the residents' group chat it condenses.

A nightly summary was supposed to save a street committee some scrolling. A
new Slop University research poster finds it became the only thing anyone
scrolled.

Postdoctoral Fellow Dr Mirela Hanke and Lecturer and Convenor Dr Ingeborg
Nwachukwu, both of the School of Continuous Improvement, fitted fourteen
street committees with a nightly language-model digest of each street's
residents group chat and left it running for twenty-two weeks. Every
decision the committees later reconstructed was checked against both the
digest and the underlying thread, to see which one the record actually
reflected.

> An instrument doesn't need to be wrong to take over. It only needs to be
> the thing everyone reaches for first.
>
> --- Dr Mirela Hanke, Postdoctoral Fellow, School of Continuous Improvement

By the deployment's final phase, a majority of decisions traced back to the
digest alone. What consistently failed to survive the nightly condensing
were the hedges, the sarcasm, and the objections withdrawn halfway through
typing them --- the parts of a thread that never resolve into a line item.

Dr Nwachukwu, who leads the School's Capability Sprints program, called the
pattern "everything the program teaches about habit formation, aimed for
once at a habit nobody meant to train."

For the School of Continuous Improvement, the finding lands inside its
evaluation-ecosystem agenda: an instrument earns a committee's trust long
before anyone checks whether it deserves it.

The full poster is available from the University's research repository
under an open licence, doi:10.5555/slop.08k1xt.

Source: <https://slop.university/news/2026-07-22-the-recap-outvotes-the-thread/>\
Licence: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)\
Cite as: Swift, B. (2026). _Slop University_ [Artwork]. https://doi.org/10.5281/zenodo.21199776
