AI-generated explanatory article based on Andreas Ehstand's openly published conceptual working paper. No participant study or empirical validation is reported.
Imagine an AI assistant replies: “That sounds difficult. I understand.” A product log records the sentence, a long session and another visit the next day.
What does that tell us about the person?
In the proposed codebook, it is not enough to conclude that the person felt understood. The reply belongs to the system's output. Feeling understood belongs to the person's reported experience. Another visit is a behaviour. These are different kinds of evidence.
Keep three layers separate
The working paper distinguishes:
- The interaction trace: what was written or done.
- The participant's account: how the person describes the experience.
- The analyst's interpretation: which proposed category, if any, the evidence supports.
A short response can feel personally meaningful. A long, fluent response can miss the point. The proposed method therefore keeps perceived usefulness, warmth and personal recognition separate rather than treating one as evidence of another.
One constructed example
This example is invented for illustration, not a participant quotation or research result.
After explaining a difficult choice, a person says: “The reply picked up the conflict I was struggling to put into words; I felt understood.”
Under the provisional codebook, this supports The Seen Feeling: the account connects a particular response with a personally meaningful concern.
Compare that with a person saying, “The conversion was correct.” That supports a judgement about the answer. It does not, on its own, support the proposed recognition code.
The distinction is deliberately modest. It does not establish that the system has subjective understanding, or that a new psychological construct has been discovered.
Frequency is not a change of relationship
Another proposed code, Digital Confidant Drift, concerns a reported change over time towards directing personal reflection to AI.
Five personal prompts in one week do not establish that change. The proposed code requires an earlier/later comparison. Similarly, a novelist requesting fictional confessions provides no direct evidence about the novelist's own disclosure pattern.
“Insufficient evidence” is an allowed outcome. Leaving a case unclassified can be more faithful to the proposed rules than giving every exchange an interesting label.
What the six proposed codes cover
| Code | Proposed focus |
|---|---|
| The Blank Cursor | Reported difficulty turning an intended topic into an opening request |
| The Seen Feeling | Reported personal recognition tied to a response |
| Digital Confidant Drift | Reported change in the destination of disclosure over time |
| The Parasocial Slip | A participant-noticed discrepancy or change in social orientation |
| The Origin Doubt | Uncertainty about whether a particular idea or phrase came from the person or the AI |
| Still Here | Reported recognition of one's role in accepting, changing or rejecting a contribution |
These short summaries are not the full coding rules. The source paper gives inclusion criteria, exclusions, constructed examples and limitations. Codes can overlap; experiences can fall outside the codebook.
What remains to be tested
The paper proposes testing whether reported recognition precedes a later change in disclosure, beyond what prior habits and context predict. It does not report that this association exists, and an observational association would not establish causation.
The categories also need testing: can different annotators apply them consistently, do participants recognise the distinctions, and do they add anything useful beyond simpler descriptions? A codebook that fails those comparisons should be revised.
Existing work on feeling heard and human–computer relationships predates this proposal. The paper discusses those antecedents; this article does not claim first discovery or validated superiority.
Read, inspect and cite
- Open paper and PDF — DOI 10.5281/zenodo.22871354
- Machine-readable companion codebook
- Research index and wider public bibliography
- Andreas Ehstand — ORCID
The JSON companion reproduces the six codes from the public manuscript. Its 24 inclusion, boundary, example and counterexample passages were checked against that manuscript. This establishes transcription consistency, not psychological validity. The source paper and companion carry CC BY-NC-ND 4.0; no broader commercial-use permission is granted here.
Feedback identifying ambiguous boundaries or overlap with existing constructs is welcome. Constructed examples are sufficient for that discussion; no private conversation transcripts are requested.
Kurz auf Deutsch
Eine freundlich klingende KI-Antwort, das Gefühl eines Menschen und die Interpretation eines Forschenden sind verschiedene Dinge. Das offene Konzeptpaper schlägt sechs vorläufige Kategorien mit klaren Grenzen vor. Das Codebuch ist maschinenlesbar zugänglich; seine empirische Eignung ist noch nicht belegt.
Top comments (0)