Interactive Paper Companion

Expectation, Backlash, Recovery, and Excitement

How model releases shape Reddit perceptions of conversational AI systems

Vahid Rahimzadeh, Yury Zhauniarovich, Savvas Zannettou

Delft University of Technology · CySafeLab

Reddit discussions of conversational AI systems, October 2022 to December 2025

A guided tour of six releases

Every model release is a user-facing event

Between October 2022 and December 2025, Reddit communities wrote hundreds of thousands of posts about conversational AI systems. We treat each major model release as an observation point and measure how sentiment and discussion themes shift in the weeks before and after. Six releases tell six very different stories.

Scroll to begin

What providers can take away

Five recurring patterns across the six releases, phrased as practical lessons for planning future model updates.

Close the expectation gap

Announce only what users can access at launch, or say clearly when and for whom advertised capabilities arrive. The distance between demo and reality dominated GPT-4o's reception.

Expectation Gap +19.4 pp after GPT-4o

Preserve model continuity

Removing models that users built workflows — and relationships — around triggers the sharpest backlash. Transition periods matter.

Feature Removal Frustration +6.8 pp after GPT-5

Ship capacity with capability

A technically admired model that people cannot reliably access still produces a negative reception. Release planning is also infrastructure planning.

Negative share +16.0 pp after DeepSeek R1

Release with a concrete use case

The most favorable shift in our data came from a model that arrived bundled with a clear workflow: Claude 4 with coding and agentic automation.

Positive +5.3 pp, negative −10.5 pp after Claude 4

Monitor perceptions continuously

Discussion themes move fast and backlash can be partially reversed: GPT-5.1 de-escalated much of the GPT-5 revolt within weeks.

Negative −4.3 pp after GPT-5.1

Technical appendix

The full interactive dashboard: model-mention structure, sentiment timelines, subreddit composition, the canonical concept bank, and release-delta explorers. All data is aggregate; no raw post text or IDs are shipped.

The appendix loads when you scroll here.

Summary Workbench

Release Events As User-Facing Interventions

Loading aggregate paper data...

Selected provider sentiment trend

Hover points for counts and rates

Release sentiment atlas

Selected release is outlined

Top concept deltas for selected release

Top 5 rising and top 5 declining

Model Extraction

Discussion Prevalence And Co-Discussion Structure

All controls in this section use the same selected taxonomy level and month range.

Selected-level mentions

Monthly prevalence by selected level

Overall top entities

Conditional co-discussion

P(target | source)

Jaccard co-discussion

Top 20 selected-level items

Top co-discussion pairs

Subreddit mention entropy

Validation

Single vs. Multi Mention

How Often Posts Compare Or Combine Models

Mention multiplicity totals

Monthly multi-mention rates

"Multi-model" and "multi-generation" are the same thing in this dataset

Multi-mention rate by entity

Share of each entity's posts that also mention another at the same level

Multi-mention by subreddit group

Sentiment

Provider Trends And Release-Window Atlas

SVG marks include hover titles with post counts, percentages, and release-window deltas.

Weekly sentiment shares

Muted area = weekly volume

Trend estimates

Release sentiment atlas

Subreddit Structure

Where Model Discussion Concentrates

Aggregate subreddit and subreddit-group views show coverage and specialization without exposing individual posts.

Coverage vs. multi-model discussion

Selected subreddit top entities

Specialization cells

Concept Landscape

Concept Prevalence And The Public Concept Bank

Prevalence data is exported as monthly/quarterly and aggregate scope tables, not raw post-level records.

Top-10 concept prevalence over time

Dashed line = top-10 coverage

Concept family cells

Canonical concept bank

Sortable aggregate prevalence

Concept Sentiment

Human Concept-Valence Annotation And Agreement

This section separates concept-level valence from prevalence and release-delta visualizations.

Annotation summary

Majority valence

Annotator agreement

Concept annotation table

Release Delta Ranking

Interactive Concept Moves Around Releases

Paper-focus releases are the main narrative cases; additional selected releases are shown for completeness.

Ranked release deltas

Release Concept Explorer

Rising And Falling Concepts Around Releases

Paper-focus releases are marked in the profile; other selected releases are included for completeness.

Concept delta bars

Names are neutral; bar direction carries change

Release profile

Methods

Public Data And Provenance

The website ships aggregate tables only. Raw Reddit text, Reddit post IDs, and local absolute paths are excluded.

Filters

Generated bundle