Model updates land silently — the same saved prompt answers differently overnight, and users discover the swap through degraded outputs, not through the product.
← Back to patterns
Guidelines for Human-AI Interaction (HAX Toolkit, G1-G18)
Trust & Verification
EmergingModel Changelog
The model changed — hear it from the product, not from your broken outputs
Model updates land silently: the same saved prompt starts answering differently overnight, and users discover the swap through degraded outputs, not through the product. Model Changelog announces the change where the work happens — what improved, what behaves differently, scoped to what you actually do — with a before/after diff on your own pinned prompts instead of version arithmetic. HAX guidelines 14 and 18 (update cautiously, notify about changes) have named this obligation since 2019; almost nothing ships it.
Metaphor
Release Notes
Interaction
Update Card, Behavioral Diff, Notification Rhythm
Use case
Model Updates, Workflow Stability, Regression Triage
Failure mode
Silent Model Swap
Control mechanisms
In-Product Change Notice · Behavior-Scoped Notes · Before/After Diff on Your Prompts · Acknowledge Control
Reference examples
ChatGPT release notes · Anthropic model release notes + system cards · Cursor changelog
Interactive Example
The live side-by-side is the comparison.
This prototype already contains the contrast between chat and the pattern response. It stands as the primary comparison surface for this page.
Nothing changed on your end. Everything changed under it.
The saved handoff prompt returned four bullets, 26 words. Two days later, same prompt, same ticket: 74 words of prose with three hedges folded in. The model went v3.2 → v3.3 at Tuesday 02:00 and the interface said nothing about it, so the first thing the user debugs is their own prompt.
The announcement belongs where the work is. An update card names what actually moved — hedging, table formatting, +38% median output tokens — then re-runs the two pinned prompts against the new version: the handoff regresses, the KPI table gains an aligned target column. The change is measured on your work, not on a benchmark.
Which is also why the notify control sits next to it. Every version bump interrupting is how users learn to dismiss all of them; a weekly digest lets a regression run for days. Scope the card to behavior rather than version arithmetic, and let the verdict — regression, improved, unchanged — come back from the person it landed on.
A model that changes silently isn't updated, it's swapped — and the user is the last to know.
DOPE Lens
How this pattern improves AI interaction quality.
Observability
The update is announced in-product with behavior-scoped notes, not buried in a blog post the user never sees
Directability
Acknowledge, snooze, or re-run your checks — the user controls when the new behavior enters their workflow where the platform allows it
Predictability
A before/after diff on the user's own pinned prompts shows how the change lands on their work, before it surprises them
Explainability
Each notice names what category of behavior moved and why, instead of a bare version number
Fit
Where it earns its place.
Use it
when
when
Users build workflows on model behavior that changes underneath them, and silent updates cost them trust or output quality.
Avoid
when
when
The user pins model versions themselves (versioned API snapshots) — the notice belongs where change is imposed, not chosen.
In the wild
Where this pattern already ships.
The pattern is shipping but still maturing across products. Entries below show how different teams have approached it; expect variation as the space settles.
Screenshot pending
ChatGPT · Release notes
OpenAI
A running in-help changelog of model and feature changes — the notice half of the pattern. Behavior-scoped diffs on your own prompts remain unshipped anywhere.
Visit OpenAI ↗
Screenshot pending
Anthropic · Model release notes + system cards
Anthropic
Versioned release notes plus system cards documenting behavioral characteristics per model — the most behavior-scoped public change documentation in the industry, but it lives in docs, not in-product.
Visit Anthropic ↗
Screenshot pending
Cursor · Changelog
Cursor
Names model swaps and behavior changes per release, surfaced to developers who treat the tool as infrastructure. In-product update notes appear on version bumps.
Visit Cursor ↗
Research Frame
Why this pattern exists, and where it breaks.
Announce the change in-product, scoped to behavior: what improved, what moved, with a before/after diff on the user's own pinned prompts and a control over when it takes effect.
Chat has no surface for "the system under this box changed last night" — same input, different distribution, zero notice.
Changelog fatigue — if every minor bump interrupts, users dismiss all notices; scope announcements to behavioral change, not version arithmetic.
Common questions
What people ask about this pattern.
When should I use the Model Changelog pattern?
+
Users build workflows on model behavior that changes underneath them, and silent updates cost them trust or output quality.
When should I avoid the Model Changelog pattern?
+
The user pins model versions themselves (versioned API snapshots) — the notice belongs where change is imposed, not chosen.
What problem does Model Changelog solve?
+
Model updates land silently — the same saved prompt answers differently overnight, and users discover the swap through degraded outputs, not through the product.
Why is chat the wrong fit for this?
+
Chat has no surface for "the system under this box changed last night" — same input, different distribution, zero notice.
Grounded in Research
Not just opinion.
The papers, guidelines, and usability studies this pattern rests on.
Saleema Amershi et al.
Microsoft Research · 2019 · refreshed 2026-07
The canonical 18-guideline taxonomy for human-AI interaction. Strongest anchors: G9 efficient correction, G11 why-did-it-do-that, G17 global controls. Its three zero-coverage guidelines — G6 bias mitigation, G14 cautious updates, G18 change notification — became this library's Bias Flag and Model Changelog patterns in July 2026. Nearly the whole Agent Workflow category sits outside its 2019 single-assistant frame.
Related Patterns
Curated connections, not just neighbors.
Hand-picked relationships — what this pattern pairs with in practice, what it replaces, what it builds on, and what builds on it.
Often paired with
Capability Boundary
Capability Boundary declares what the system can do; Model Changelog announces when that declaration moves.
Often paired with
Model Selector
Model Selector lets you pick the model per task; Model Changelog tells you when the model behind that pick changes.
Often paired with
Seed Pinning
Seed Pinning reproduces an output only on the same model; Model Changelog tells you when that model changed.