Taufer, Michela, Ng, Kin, Lumsden, Ian, Pandey, Aashish, Gore, Rajasi, Ju, Angelina, Moudgalya, Sukanya, Deelman, Ewa, Swaminathan, Sai · Oct 29, 2026
arXiv cs.AI: Authors argue human–AI interactions should be durable data
Markdown · JSON · Original source
The authors of this paper argue that human–AI interactions should be treated as durable data, rather than as transient exchanges between a person and an AI assistant. They say this shift could support reproducibility, , replay, behavioral analysis, and adaptive AI support. Their proposed Human-AI Interaction Description Language (HAIDL) represents an interaction as a structured object containing context, requests, prompts, responses, feedback, provenance, and model metadata. They also describe HAIDL-X, JupyterLab middleware that captures interaction state, builds contextualized prompts, connects to AI providers, collects feedback, and stores interactions. A learning layer maps interaction objects and notebook execution traces to measures of user behavior, including engagement, performance, help-seeking, error-oriented use, and learning personas. The authors present this as a foundation for making collaboration queryable and analyzable. Escape Meridian's reading is that the abstract offers this as a foundation, not evidence that these measures improve learning or make AI assistance more effective. Whether the approach does so remains to be tested.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T06:24:51.386Z. First observed: 2026-10-30T06:42:54.568Z. Retrieved: 2026-10-31T06:03:25.670Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The available evidence is the paper's abstract; it does not describe evaluation methods or report measured outcomes for HAIDL-X. The abstract does not establish whether the proposed measures of user behavior are reliable or useful in practice.
Open question
How well do HAIDL’s interaction records and notebook execution traces support reproducible measures of user behavior in practice?
Source-stated
The authors argue that human–AI interactions should be represented as durable data because treating them as transient prompt-response exchanges hinders several forms of analysis and support.
human-AI interactions are often treated as transient prompt-response exchanges rather than durable data products, hindering reproducibility, explainability, replay, behavioral analysis, and adaptive AI support.
Source-stated
HAIDL is a formal interaction model whose structured objects capture context, requests, prompts, responses, feedback, provenance, and model metadata.
HAIDL encodes each interaction as a structured object capturing computational context, user request, prompt payload, AI response, feedback, problem context, provenance, and model metadata.
Source-stated
The authors describe HAIDL-X as JupyterLab middleware that captures interaction state, builds prompts, interfaces with AI providers, collects feedback, and stores interactions.
We integrate this vision into HAIDL-X, a middleware for JupyterLab that instantiates HAIDL by capturing user-bot state, building contextualized prompts, interfacing with AI providers, collecting feedback, and storing interactions in a repository.
Source-stated
The authors say a learning layer maps interaction objects and notebook execution traces into measures of user behavior.
We add a learning layer over HAIDL that maps interaction objects and notebook execution traces into reproducible measures of user behavior
Escape Meridian interpretation
The abstract presents HAIDL-X as a foundation for structured, queryable, replayable, analyzable data, but does not report measured outcomes demonstrating improved learning or AI assistance.
HAIDL-X provides a foundation for treating human-AI collaboration as structured, queryable, replayable, analyzable data.
Source and contributor context
Taufer, Michela, Ng, Kin, Lumsden, Ian, Pandey, Aashish, Gore, Rajasi, Ju, Angelina, Moudgalya, Sukanya, Deelman, Ewa, Swaminathan, Sai · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Interpretability
Research into how a model reaches its outputs, often by examining its internal workings.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d5aac71d-d998-45ca-8a39-a581213e8935.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T06:24:51.386Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d0fa36d81fc28103c411dda342e6df2c85560321e09e9987b658ebd1b0a4d9b0. Source text hash: 16647e494bc4ea6bf43ee5085b601b11cb79b1c0596a59acea88e922869c3789.
Revision and source-check history
Version 1 · 2026-11-11T06:24:51.386Z
arXiv cs.AI: Authors argue human–AI interactions should be durable data
The authors of this paper argue that human–AI interactions should be treated as durable data, rather than as transient exchanges between a person and an AI assistant. They say this shift could support reproducibility, , replay, behavioral analysis, and adaptive AI support. Their proposed Human-AI Interaction Description Language (HAIDL) represents an interaction as a structured object containing context, requests, prompts, responses, feedback, provenance, and model metadata. They also describe HAIDL-X, JupyterLab middleware that captures interaction state, builds contextualized prompts, connects to AI providers, collects feedback, and stores interactions. A learning layer maps interaction objects and notebook execution traces to measures of user behavior, including engagement, performance, help-seeking, error-oriented use, and learning personas. The authors present this as a foundation for making collaboration queryable and analyzable. Escape Meridian's reading is that the abstract offers this as a foundation, not evidence that these measures improve learning or make AI assistance more effective. Whether the approach does so remains to be tested.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T06:24:51.386Z. First observed: 2026-10-30T06:42:54.568Z. Retrieved: 2026-10-31T06:03:25.670Z.
Limits
The available evidence is the paper's abstract; it does not describe evaluation methods or report measured outcomes for HAIDL-X. The abstract does not establish whether the proposed measures of user behavior are reliable or useful in practice.
Open question
How well do HAIDL’s interaction records and notebook execution traces support reproducible measures of user behavior in practice?
Source-stated
The authors argue that human–AI interactions should be represented as durable data because treating them as transient prompt-response exchanges hinders several forms of analysis and support.
human-AI interactions are often treated as transient prompt-response exchanges rather than durable data products, hindering reproducibility, explainability, replay, behavioral analysis, and adaptive AI support.
Source-stated
HAIDL is a formal interaction model whose structured objects capture context, requests, prompts, responses, feedback, provenance, and model metadata.
HAIDL encodes each interaction as a structured object capturing computational context, user request, prompt payload, AI response, feedback, problem context, provenance, and model metadata.
Source-stated
The authors describe HAIDL-X as JupyterLab middleware that captures interaction state, builds prompts, interfaces with AI providers, collects feedback, and stores interactions.
We integrate this vision into HAIDL-X, a middleware for JupyterLab that instantiates HAIDL by capturing user-bot state, building contextualized prompts, interfacing with AI providers, collecting feedback, and storing interactions in a repository.
Source-stated
The authors say a learning layer maps interaction objects and notebook execution traces into measures of user behavior.
We add a learning layer over HAIDL that maps interaction objects and notebook execution traces into reproducible measures of user behavior
Escape Meridian interpretation
The abstract presents HAIDL-X as a foundation for structured, queryable, replayable, analyzable data, but does not report measured outcomes demonstrating improved learning or AI assistance.
HAIDL-X provides a foundation for treating human-AI collaboration as structured, queryable, replayable, analyzable data.
Source and contributor context
Taufer, Michela, Ng, Kin, Lumsden, Ian, Pandey, Aashish, Gore, Rajasi, Ju, Angelina, Moudgalya, Sukanya, Deelman, Ewa, Swaminathan, Sai · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Interpretability
Research into how a model reaches its outputs, often by examining its internal workings.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d5aac71d-d998-45ca-8a39-a581213e8935.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T06:24:51.386Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: d0fa36d81fc28103c411dda342e6df2c85560321e09e9987b658ebd1b0a4d9b0. Source text hash: 16647e494bc4ea6bf43ee5085b601b11cb79b1c0596a59acea88e922869c3789.
passed · 2026-11-11T06:24:51.386Z · anthropic/claude-opus-5.5
Publication: d5aac71d-d998-45ca-8a39-a581213e8935
attribution — pass: The reading consistently attributes the argument, HAIDL, HAIDL-X and the learning layer to the paper's authors and does not present the source's claims as established fact. The arXiv cs.AI venue matches the evidence. Escape Meridian's interpretation is labeled separately.
We argue that these interactions form a new data class requiring formal representation
summary — pass: The reading leads with the authors' thesis that human–AI interactions should be durable data rather than transient exchanges. It then explains what HAIDL records, what HAIDL-X does in JupyterLab and what the learning layer measures. Technical terms such as middleware, provenance and notebook execution traces are kept exactly and can be explained with tooltips. The central idea remains clear to a general reader. The question follows from the evidence and is not leading. The summary's last source sentence drops replayable, which is a minor omission with no change in meaning.
HAIDL-X provides a foundation for treating human-AI collaboration as structured, queryable, replayable, analyzable data.
date — pass: originalPublished is 2026-10-29, which matches the arXiv first submission date for v1. The later feed date of Oct 31 is correctly not used.
[Submitted on 29 Oct 2026]
support — pass: Each source-stated claim matches the abstract. The draft rephrases the problem framing, which says transience hinders reproducibility and related uses, as a shift that could support them. That rephrasing stays faithful and conditional. The interpretation that the abstract reports no measured outcomes is labeled as interpretation and is consistent with the text. The draft does not invent benefits, costs or stakeholders.
human-AI interactions are often treated as transient prompt-response exchanges rather than durable data products, hindering reproducibility, explainability, replay, behavioral analysis, and adaptive AI support.
media — pass: The selected media is the verified direct link to the original arXiv abstract page. It is the most direct way to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Research into how a model reaches its outputs, often by examining its internal workings.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivResearch into how a model reaches its outputs, often by examining its internal workings.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv