AI Signals Briefing

Superhuman CEO Shishir Mehrotra on the 'Expert Review' feature that used reporters' names

Nilay Patel interviews Superhuman CEO Shishir Mehrotra after the 'Expert Review' AI feature used reporters' names without consent; he apologizes and outlines remediation.

TL;DR in plain English

  • The Verge reported an AI feature that produced content framed as coming from a named reporter; the piece includes an interview with the company CEO. Read it here: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Key actionable takeaways:

  • If your product claims to speak “in the voice” of real people, you face rapid trust, PR, and legal risk. See the Verge reporting above for context.
  • Immediate targets you can adopt: confirm scope within 30 minutes, publish an initial public status within 180 minutes (3 hours), and aim to remediate within 72 hours.
  • Technical guardrails: provide an opt-out, visible attribution, and a one-click rollback that can flip in <5 minutes and update UI state within ~200 ms.

This note translates the Verge interview into operational steps teams can run now. It is practical guidance, not legal advice. Methodology: this memo summarizes the Verge reporting and CEO interview listed above (same link). https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

What changed

The Verge published reporting and an interview about an AI product that generated content presented as if authored by a named reporter; the piece focuses on attribution versus impersonation. Read it here: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Use this reusable incident timeline and fill values immediately after discovery:

| Step | Example field | Your value (fill in) | |---|---:|---| | Release date | ISO date | 2026-03-XX | | Discovery date | ISO date | 2026-03-XX | | Discovery channel | e.g., reporter story, social | reporter story | | Initial mitigation options | pause / opt-out / narrow / defend | pause | | Time-to-public mitigation target | hours | 72 | | Final resolution | removed / revised / policy | pending |

Reference: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Why this matters (for real teams)

  • Trust and retention: public claims of impersonation can erode trust quickly. Track time-to-remediation (goal <72 hours) and monitor retention changes over 7 and 30 days. A >3% drop in retention is material. See the Verge reporting for incident context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

  • Operational surge: expect support-volume spikes of 3×–10× in the first 24 hours after public coverage; plan for temporary 24–72 hour scale-ups in staffing.

  • Legal and partner risk: features that use real names or imitate public figures raise rights-of-publicity and reputational exposure. Treat persona features as requiring legal review before broad release. The Verge piece explains why attribution versus impersonation matters: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Concrete thresholds you can copy now:

  • Triage SLAs: 30 minutes to confirm scope; 180 minutes to publish initial public status; 72 hours to remediate or publish remediation timeline.
  • Rollback targets: one-click revert across environments in <5 minutes; UI state update <200 ms.
  • Audit retention: persist prompt context at least 1,024 tokens per event for 90 days.

Concrete example: what this looks like in practice

Scenario: you ship an “expert tips” mode that labels generated suggestions as coming from named professionals. A reporter sees their name used without consent and publishes a story. Use the Verge report as the incident lens: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Decision options (when to use each):

  • Apologize + Opt-out: add an opt-out and publish a brief public apology; appropriate when exposure is broad and harm is reputational. Target opt-out uptake >20% in first 7 days if users are concerned.
  • Narrow content: replace names with role labels (e.g., “AI-generated — Senior Editor”) and remove named prompts; low-risk stopgap.
  • Remove feature: full removal followed by rebuild with consent flows; required if legal counsel advises.
  • Defend: only with legal signoff and clear evidence; high risk.

Operational metrics during an incident:

  • Time-to-removal target: <72 hours.
  • Opt-out uptake target: >20% in first 7 days (if deployed).
  • Support-ticket surge: plan for +200–1,000 tickets depending on user base size.

Example rollback: flip persona_enabled = false. The toggle should update UI state within ~200 ms and propagate across CDN edges within 5 minutes. Reference: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

What small teams and solo founders should do now

If you are 1–5 people, prioritize fast, low-cost, high-impact actions. The Verge report provides the incident context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Minimum concrete actions (solo founder / tiny team):

  1. Pause persona outputs now.
    • Action: flip a runtime flag or remove the persona template from circulation. Goal: 30–180 minutes from discovery.
  2. Replace real names with generic roles and add attribution.
    • Action: change labels to “AI-generated — [role]” and surface visible attribution on first render for ≥3 seconds.
  3. Add a persistent opt-out and a one-click global rollback.
    • Action: persist persona_enabled = false as a user setting and expose a global admin toggle; test end-to-end rollback in <5 minutes and validate UI latency <200 ms.

Additional low-cost, high-impact steps:

  • Inventory the top 50 templates first; record template hashes and the last 1,024 tokens for quick audits.
  • Draft a 2–3 sentence public status and a short FAQ; publish within 30–60 minutes.
  • If possible, get a 30-minute consult with counsel; capture guidance in the incident log.

Quick checklist to copy and run now:

  • [ ] Pause persona outputs (30–180 minutes)
  • [ ] Inventory top 50 templates (24–48 hours)
  • [ ] Publish short status message (30–60 minutes)

Reference: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Regional lens (US)

Operational guidance for US launches (not legal advice). Use the Verge reporting as the incident reference: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

US-focused steps:

  • Require legal review for persona outputs that reference named individuals; record signoff timestamps.
  • Maintain a decision log with timestamps and who signed off; regulators and partners expect timely corrective action—aim for visible remediation under 72 hours.
  • Prepare PR materials: one-line public statement template, customer FAQ, and documented rollback evidence for executives.

Monitoring metric: time from public discovery to visible remediation (goal <72 hours). See the Verge interview for context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

US, UK, FR comparison

Below is a concise operational comparison. Local counsel should confirm specifics. Source context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

| Jurisdiction | Operational focus | Quick mitigation checklist | |---|---:|---| | US | Fast PR and publicity risk; state rights-of-publicity concerns | Pause persona outputs; legal review; incident log; 72h remediation target | | UK | Transparency and reputation; potential defamation concerns | Clear attribution; consent where possible; public FAQ and correction path | | FR (France) | Strong personality/image protections; EU digital expectations | Avoid names without consent; document GDPR triggers if personal data processed |

Technical notes + this-week checklist

This section lists short technical requirements and a 7-day tactical checklist. See the Verge reporting for incident context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Assumptions / Hypotheses

  • The Verge reporting documents a reporter being impersonated and includes an interview with the company CEO about attribution vs. impersonation: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation
  • Items not explicitly in that reporting (for example: specific lawsuits, settlements, or timeline outcomes) are treated as hypotheses and must be validated before citing externally.

Risks / Mitigations

Risks:

  • Public trust erosion: measurable as >3% retention drop or sustained NPS decline; support-ticket surges of 3×–10× in the first 24 hours.
  • Legal exposure on rights-of-publicity and reputation.
  • Operational cost: engineering hours and PR time to rollback and remediate.

Mitigations (first actions):

  • Add persona_enabled = false global toggle. Target rollback execution <5 minutes and UI status update <200 ms.
  • Persist prompt logs: record template hashes and the last 1,024 tokens per event for 90 days for audits.
  • Require Product + Legal signoff (consent, attribution, audit log, legal review, rollback) before re-enabling persona outputs.

Next steps

This-week checklist (7-day aims):

  • [ ] Inventory top 50 prompt templates and flag those referencing named individuals (24–48 hours).
  • [ ] Add visible attribution text for generated content and deploy within 48 hours.
  • [ ] Implement and test a global opt-out toggle (persona_enabled = false) with one-click rollback; deploy within 72 hours.
  • [ ] Draft a short public FAQ and incident-status template (ready in 24–48 hours).
  • [ ] Engage counsel to review persona use and prepare a lightweight consent policy (meet within 7 days).

Developer checklist:

  • Record template hashes and last 1,024 tokens for persona outputs.
  • Ensure deployments can flip persona_enabled across prod/staging/canary in <5 minutes.
  • Add monitoring: alert on 3× baseline increase in support tickets or social-mention spikes; notify execs.

Final note: treat the Verge report as a practical wake-up call and follow the short SLAs above. Source and context: https://www.theverge.com/podcast/898715/superhuman-grammarly-expert-review-shishir-mehrotra-interview-ai-impersonation

Share

Copy a clean snippet for LinkedIn, Slack, or email.

Superhuman CEO Shishir Mehrotra on the 'Expert Review' feature that used reporters' names

Nilay Patel interviews Superhuman CEO Shishir Mehrotra after the 'Expert Review' AI feature used reporters' names without consent; he apologizes and outlines r…

https://aisignals.dev/posts/2026-03-24-superhuman-ceo-shishir-mehrotra-on-the-expert-review-feature-that-used-reporters-names

(Weekly: AI news, agent patterns, tutorials)

Sources

Weekly Brief

Get AI Signals by email

A builder-focused weekly digest: model launches, agent patterns, and the practical details that move the needle.

  • Models and tools: what actually matters
  • Agents: architectures, evals, observability
  • Actionable tutorials for devs and startups

One email per week. No spam. Unsubscribe in one click.

Services

Need this shipped faster?

We help teams deploy production AI workflows end-to-end: scoping, implementation, runbooks, and handoff.

Keep reading

Related posts