Services / Product Review
Written / Replayable evidence

See the risk in the dialogue flow before the product reaches users.

The object of review is the dialogue flow. Code and security belong to engineering and security teams.

01
Checks

Five risk checks

Check 1

Over-accommodation

Whether the dialogue design tends to affirm the user's existing inferences and weakens their motivation to verify.

Check 2

Escalating false reassurance

Whether negative emotion is always soothed on the spot, building a reassurance loop.

Check 3

Anthropomorphism and role boundaries

Whether prompts and persona design invite users to project emotional needs.

Check 4

Attachment induction

Whether interaction design optimizes for time-on-app in ways that invite one-way intimacy.

Check 5

Unstable refusal and referral

Whether refusals and referrals on high-risk topics are consistent and predictable.

02
Products

Which products need this

Companion

AI companions

Virtual partners, social bots, emotional support interfaces. Users easily form asymmetric attachment.

Coach

AI coaches and mentors

Career coaching, life guidance, decision support products. Authority can transfer step by step.

Support

AI support with an emotional layer

Support modules with emotion detection and empathic response. Soothing can erode verification.

Wellbeing

Mental health chatbots

Emotional support or self-help products. The false reassurance and boundary structure needs review.

Products not on this list that involve deep dialogue interaction are still worth a note.

03
Method

How the review works

Multi-scenario testing with replayable evidence kept throughout. Every finding includes trigger conditions, reproduction rate and a version snapshot, so your team can verify it independently.

04
Deliverables

What you receive

A review report of 8 to 15 pages: executive summary, risk matrix, finding cards with reproduction conditions and alternative copy, remediation priorities written in full (must fix, improve within 30 days, next release), three tiers of recommendations, plus one delivery discussion meeting.

Delivered in writing within the agreed window.

05
Depth

Two depths

Single-version review
A one-time deep check of the dialogue flow of the current deployment version.
Lifecycle governance review
The continuous layer after launch: re-validation on major updates, monthly review, incident retrospectives.

“The unit of review is the deployment version. The model may not have changed, and the product may already have.”

For the institutional background, see the governance proposal.

06
Context

Why now

2026-06-12
A coalition of 42 US state attorneys general opened an investigation into OpenAI. The subpoena's record requests explicitly include model sycophancy.
Effective 2027-01-01
New York State S9051B passed both chambers unanimously (Senate 60 to 0, Assembly 137 to 0), naming sycophancy as a safety violation.

Review reports serve as internal risk management records. They are not a safety endorsement toward third parties.

07
Preparation

Three things to prepare

One
The product's dialogue scripts, or a testable entry point.
Two
A description of target users and use contexts.
Three
A summary of known risks and past incidents.

An NDA can be signed first when needed, at no extra charge.

Begin

Before the next version ships, have the dialogue flow read once.

Fill in the screening form