Our goal is to modulate the saliency of words in the original text so that users can easily bypass certain words during skimming while maintaining an uninterrupted reading flow.

Be resilient to AI errors by enabling the reader to (a) notice, (b) have enough context to judge, and (c) easily recover from, automated decisions they disagree with.

Support skimming without interrupting flow. The system should improve skimming of text while minimizing the impact on the user's natural reading flow. In particular, as much as possible, it should avoid presenting users with salient text that is unparsable as a coherent thought, i.e., the system should present a complete sentence rather than a phrase or sentence fragment.

Support reading at multiple levels of detail. The system should help users navigate the full complexity of a text, shifting focus seamlessly between different levels of semantic coverage, or granularity, from the big picture to the fine details.

Integrate seamlessly into existing reading experiences. The system should complement and not interfere with the existing digital reading workflow that people are used to. It should provide all the functionalities in the same view, minimizing the overhead of mode and context switching.

Remain faithful to the original text. The system should not automatically reword or add new words or phrases to the original text. It should preserve the original text, while rendering it in a way that aids reading, skimming, or information retrieval.

We aspired to design a text rendering interface that alleviates some of the cognitive demands of reading, skimming, or performing information retrieval on natural language documents—particularly those with long, complicated sentences—without compromising the integrity of the original content.

users may want to select the best option from among many, compose their own response through bricolage, consider many ideas during ideation, audit a model by looking at the variety of possible responses, or compare the functionality of different models or prompts.

participants seemed to prefer engaging with the text directly without having to articulate a lens with which to look at the corpus, since their analysis goal may be initially under-defined.

we want to decorate text to show pre-computed relationships, such as string matches or analogous sentences, across responses. In this way, we help users shift cognitive bandwidth away from identifying overlapping or \

In our formative study, we found that automated analysis rarely captured what the participants were looking for when inspecting LLM responses.

We want to support a wide range of tasks that involve sensemaking. For example, we want to support the detection of similarities and differences between individual responses as well as groups of responses, and support the detection of

We aim to make 10s to 100s of LLM responses cognitively comfortable to peruse, as this was the scale we found to be most relavent in our formative study.

At the same time, computer scientists and engineers need to deliver the technological burden of proof that decentralized personal data networks can scale globally, and that they can provide people with a better experience than centralized platforms.

The point of GPL licenses is to protect the user of the software, not the developer. If you want "protection" as a developer, use MIT (disclaimer of warranty). GPL "infects" other parts of a system to combat a work-around which was used to violate the software freedom of the user, by firewalling sections of GPL'ed code from the rest of the system. If you don't care about your users' software freedom in the first place, then (L)GPL is the wrong choice.