Skip to content
Aakash GuptaAakash Gupta

Complete Course: AI Product Design

Elizabeth Laraki reveals how to design AI products users actually love. She breaks down the 3-phase design framework from Google, shares the shocking AI image expander story, and shows why everyone's adding chat interfaces wrong. --- Full Writeup: https://www.news.aakashg.com/p/elizabeth-laraki-podcast Transcript: https://www.aakashg.com/how-to-design-ai-products-complete-masterclass-with-elizabeth-laraki/ --- Timestamps: 00:00:00 - Intro 00:01:52 - Elizabeth's background at Google 00:04:19 - Google's AI search integration 00:06:19 - Designing image & video for AI 00:09:44 - AI image expander disaster 00:16:05 - Ads 00:17:50 - AI safeguards & human-in-the-loop 00:18:28 - 3-step AI design process 00:31:29 - Ads 00:33:25 - Designing AI voice interfaces 00:38:25 - Designing beyond chat 00:41:52 - AI design tools for designers 00:44:49 - Live design: LinkedIn for AI 00:57:04 - Google Maps redesign story 01:04:14 - Google Maps India landmarks 01:10:09 - Where to find Elizabeth 01:12:00 - Outro --- Thanks to our sponsors: 1. Vanta: Automate compliance, manage risk, and prove trust - http://vanta.com/aakash 2. Kameleoon: Leading AI experimentation platform - http://www.kameleoon.com/ 3. The AI PM Certificate: Get $550 off with ‘AAKASH550C7’ - https://maven.com/product-faculty/ai-product-management-certification?promoCode=AAKASH550C7 4. The AI Evals Course for PMs: Get $1155 off with code ‘ag-evals’ - https://maven.com/parlance-labs/evals?promoCode=ag-evlas --- Key takeaways: 1. The Core Design Process Hasn't Changed: Define the product (who, what tasks, what needs), Design it (features, architecture, flows), Build it (UIs, brand). Don't skip to "let's add a chatbot" because you have API access. 2. AI Adds Non-Deterministic Risk: Traditional software is deterministic - click A, get B every time. AI is non-deterministic with unpredictable outputs. Elizabeth's image expander added a bra strap that wasn't in the original photo. Completely unintentional, completely unacceptable. 3. Work With Research on Safeguards: Audit training data for bias. Build evals that flag sensitive content (human bodies, faces, private information). Show A/B options for ambiguous cases. Make AI's work visible in the UI so users can scrutinize changes. 4. Start With Jobs To Be Done: Don't ask "We have GPT-4, what should we build?" Ask "What painful workflow takes users hours?" Descript mapped video editing lifecycle and baked AI into each job: remove filler words, edit from transcript, create clips, write titles. 5. Map User Context, Not Just Needs: ChatGPT voice in car with three kids? Perfect - nobody's looking at screen. Meta Ray-Bans reading Spanish menu item by item? Terrible - should ask "What are you in the mood for?" Same AI, different context requires different design. 6. Emerge From Ambiguity First: For "LinkedIn for AI," Elizabeth mapped 4 possible directions, picked Matchmaking, identified AI's unlock (personality patterns vs keyword matching), mapped separate UIs for job seekers and employers. Only then touch pixels. 7. Chat Fails for Complex Tasks: Elizabeth tried creating Madrid itinerary in ChatGPT. Every change regenerated everything with new hallucinations. Chat works for Q&A but fails for document creation, visual tasks, multi-step workflows that need persistent editable outputs. 8. Make Chat Supporting, Not Primary: Photoshop embeds AI in existing canvas tools. Google Search shows AI summaries inline in normal results. Cove gives canvas with multiple AI conversations in parallel. Chat is a tool, not THE interface. 9. Stop Adding AI Sprinkles: Elizabeth: "I can't help but think of this massive container of AI sprinkles everybody's shoving on top." Twitter/X + Grok, Amazon + Rufus, Apple Photos all feel forced. Ask three questions: Is this solving a real problem? Does chat make sense? Can you show your work? 10. Google Maps India Innovation: Researched how Indians actually navigate (by landmarks, not street names). Identified which landmarks work (visible from street level like temples, petrol stations). Redesigned entire directions system around that insight. That's design, whether AI or not. --- Where to find Elizabeth: Twitter: https://twitter.com/elizlaraki LinkedIn: https://linkedin.com/in/elizlaraki Substack: https://elizlaraki.substack.com --- Where to find Aakash: Twitter: twitter.com/aakashg0 LinkedIn: linkedin.com/in/aagupta/ Newsletter: news.aakashg.com #aidesign #productdesign #aiproducts --- About Product Growth: The world's largest podcast focused solely on product + growth, with over 187K listeners. Hosted by Aakash Gupta, who spent 16 years in PM, rising to VP of product, this 2x/week show covers product and growth topics in depth. Subscribe and turn on notifications to get more videos like this.

Aakash GuptahostElizabeth Larakiguest
Oct 6, 20251h 12mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Practical frameworks for designing AI products beyond chat interfaces today

  1. AI features should be integrated into existing user workflows (like Search) while clearly managing nondeterminism and hallucination risk.
  2. Chat-based UX is inherently linear and often mismatched to tasks that require stable artifacts, visual context, or iterative co-creation on a canvas.
  3. AI product design requires safeguards across model training/evals and interface design, with explicit human-in-the-loop review for sensitive outputs.
  4. Great AI products “bake AI into the cake” by embedding it across core jobs-to-be-done (e.g., transcript editing, filler word removal, clip creation).
  5. Strong product outcomes still start with classic design fundamentals: define the product, design the experience/architecture, then build and iterate using user research.

IDEAS WORTH REMEMBERING

5 ideas

Treat AI outputs as probabilistic and design for uncertainty.

Laraki and Gupta emphasize that hallucinations are normal across LLMs, so products should gate when to show answers, communicate confidence, and provide verification paths rather than presenting every output as fact.

Embed AI into real user jobs, not as a superficial add-on.

Tools like Descript feel powerful because AI supports each step of the workflow (edit via transcript, remove filler words, generate clips/titles) instead of being a single flashy feature bolted on top.

Move beyond linear chat when the task needs stable structure or visual grounding.

For tasks like travel itineraries or physical troubleshooting, a canvas or artifact-centered UI (image/video stays central; chat becomes a tool around it) better supports iteration, reference, and user control.

Pair model-side safeguards with UI-side review mechanisms.

The image expander incident shows that “reasonable workflows” can yield harmful results; mitigation includes training/evals plus UI that clearly marks AI-generated regions and prompts human review before publishing.

Design voice experiences around context, turn-taking, and how people actually consume information.

ChatGPT voice in a car works because it feels like a natural participant, while reading an entire menu aloud (Meta glasses) fails because it ignores human scanning behavior and conversational preferences.

WORDS WORTH SAVING

5 quotes

There are safeguards with which, as users, we need to think about when assuming any answer.

— Elizabeth Laraki

It felt like I walked into a bike shop and got the least helpful bike mechanic I could possibly find.

— Elizabeth Laraki

AI can have very unintended consequences, and as people using these tools, we need to have a heightened level of scrutiny.

— Elizabeth Laraki

The goal is really to emerge from ambiguity with a clear sense of what you're building and for whom.

— Elizabeth Laraki

We went from three tabs with a total of five different search boxes to one single search box.

— Elizabeth Laraki

Google Search AI summaries and trust/hallucinationsDesigning image/video-centric AI interactionsAI image expander failure case and biasSafeguards, evals, and human-in-the-loop UI patternsThree-step AI design process: define, design, buildVoice UI context: ChatGPT voice, Meta Ray-Bans, LimitlessBeyond chat: canvas/document-based AI (Cove)AI tools for designers (Figma Copilot, prototyping tools)Live design exercise: “LinkedIn for AI” matchmakingGoogle Maps redesign and landmark-based directions in India

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome