27:19
trustworthy AI
Batteiger argues that as agentic AI systems take over more of product creation, builders can no longer just avoid harm — they must explicitly design and codify what trustworthiness means (character plus capability, vision plus guardrails), because left unspecified, agents default to "the average," slop, hallucinations, and untrustworthy outcomes; the guiding principle should be trust over short-term gain.
New vocabulary — slop, hallucinations, sycophancy, enshittification — emerged in roughly three years and names real harms in current tech products.
AI algorithms flatten diversity, defaulting to "the average" answer rather than a range of viewpoints.
Data centers' fresh-water consumption and AI-enabled autonomous weapons (removing human oversight from targeting) are cited as harmful real-world impacts of current AI systems.
Batteiger explicitly rejects tech leaders who frame environmental concern as a "strategic disadvantage."
Trust has two dimensions per Stephen Covey's Speed of Trust — character (integrity, intent, honesty) and capabilities (actual results) — mapped onto a house's foundation and structure respectively.
The Starbucks South Korea "Tank Day" campaign is used as a central case study: an AI-assisted marketing failure that referenced a violent historical crackdown, triggering a 26% one-week revenue drop (~$580M annualized) and public backlash.
Codifying what NOT to build ("trust by design") only gets a product to the "ground floor" — it doesn't deliver positive impact; teams must also specify what "good" looks like or agents default to slop and non-trustworthy outcomes.
Martin Eriksson's "decision stack" (vision, strategy, goals, principles) should be made explicit for agentic co-creators, with "trust over short-term gain" proposed as a foundational principle.
Akshay Kore's framework holds that a system that isn't trustworthy isn't useful, breaking trustworthiness into explicable, transparent, non-biased, privacy-centered, and beneficial to society.
The EU AI Act's foundational dimensions — human autonomy, prevention of harm, fairness, explicability — are cited as a practical governance reference.
Concrete design examples: a radiologist workflow where the AI only alerts on disagreement (preserving the doctor's diagnostic skill); a talent.com hiring system where excluding bias-proxy data (zip codes, commute time, resume gaps) is treated as a core product feature, not a constraint; a shopping agent that produces a sweatshop T-shirt without context versus a genuinely trustworthy purchase when given explicit values and budget.
Agentic purchasing needs new trust signals — rate limiting and supply-chain information — exposed explicitly to agentic buyers.
Preserving human agency/choice is framed as good trust-by-design practice, illustrated by Ecosia gaining 40% more US users recently by not forcing AI summaries, while Google did not preserve that choice.
A tech-stack-wide trust-by-design framework (attributed to a person referred to as "Prompt Ledger") covers authorizations, identity verification (citing Estonia and Saudi Arabia's work on agent ID systems), explicit consent/context, recourse/fallback/kill switches, and logging for explicability.
Trustworthiness must be built in at every stage of the product development lifecycle — discovery, analysis, building, scaling — in collaboration with engineering and design, not left to hope.
Co-active coaching no/yes exploration — A coaching tool used with leaders facing hard decisions to explore what they do not want to stand for alongside what matters and what they want to say yes to. Apply: When facing a difficult product decision, explicitly list what you refuse to build/stand for and what you want to affirmatively create, then lead the decision from that pairing.
Makers Manifesto — A guidance document produced by a working group of product coaches and thought leaders (including Batteiger) on building positive products, referenced as forthcoming/soon to be public. Apply: Consult it as a shared reference point for what 'building positive products' means across the product community.
House metaphor for trust — A metaphor mapping trust's foundation (character, vision) and structure (capabilities, results) onto the parts of a house that must be built intentionally or the structure collapses. Apply: Use it to check that a product has both a trustworthy foundation (integrity/intent) and reliable, well-designed upper structure (delivered results) before shipping.
Speed of Trust (Stephen Covey) — A framework stating trust is composed of two dimensions: character (integrity, intent, honesty, congruence, motives) and capabilities (the results actually produced). Apply: Evaluate a product or agentic system on both dimensions — its stated intent/values and its actual delivered reliability — since either one alone is insufficient for trust.
Decision stack (Martin Eriksson) — A layered model of vision, strategy, goals, and principles as the foundational structure for product decisions. Apply: Make each layer explicit and available to agentic co-creators so their outputs are grounded in the same vision/strategy/goals/principles as human decision-makers.
Trust over short-term gain — A proposed foundational principle stating that when forced into difficult tradeoffs, trust should be chosen over short-term profit or speed. Apply: Encode it explicitly as a guiding principle for teams and agents so that rushed decisions do not erode trust in pursuit of quick wins.
Akshay Kore's trustworthy-AI framework — A 2018 framework arguing a system that isn't trustworthy isn't useful, defined via five dimensions: explicable, transparent, non-biased, privacy-centered, and beneficial to society. Apply: Audit an AI-driven product feature against each of the five dimensions before considering it complete or launch-ready.
EU AI Act trustworthiness dimensions — A regulatory framework built around human autonomy, prevention of harm, fairness, and explicability, referenced as a practical governance foundation. Apply: Reference these four dimensions when defining governance requirements or design constraints for AI-driven product features.
Trust by design (ground-floor codification) — A practice of explicitly codifying what should not be built (harms, biases, violations) as a minimum baseline layer for a product. Apply: Use it as the starting compliance/safety layer, but recognize it must be paired with a positive vision of 'what good looks like' or agents will still default to average/slop outcomes.
Radiologist double-check workflow — A workflow where a doctor states their own diagnosis first and the AI only surfaces an alert when its independent read disagrees, preserving the doctor's diagnostic skill while adding a safety net. Apply: Design AI assistance to trigger only on disagreement with a human's independent judgment, rather than replacing the human's primary skill exercise.
Feature-negative fairness design — Deciding what data a model is deliberately not allowed to see (e.g., zip codes, commute time, resume gaps as bias proxies) and treating that exclusion as a core, sellable product feature. Apply: In systems requiring provable fairness (e.g., hiring), design explicit data-exclusion rules as a marketed feature rather than an afterthought constraint.
Rate limiting and supply-chain transparency for agentic buyers — Trust signals — rate limits and supply-chain information — proposed as necessary information to expose to agentic purchasing systems. Apply: Expose supply-chain data and apply rate limits to agent-driven purchasing flows so agents can judge whether continued buying is appropriate.
Choice and control by design (human agency) — A design principle of preserving human agency and choice within AI-driven product experiences, illustrated by Ecosia letting users opt out of AI summaries. Apply: Preserve user choice/opt-outs for AI features rather than forcing AI-driven defaults on all users.
Trust-by-design tech-stack framework (Prompt Ledger) — A full-stack framework for building trustworthiness covering authorizations, identity verification (including national agent-ID systems like those in Estonia and Saudi Arabia), explicit consent/context, recourse/fallback/kill-switch mechanisms, and logging for explicability. Apply: Work through each layer — authorization, identity, consent, failure recourse, and logging — when architecting an agentic product to ensure trustworthiness is built into the technical stack itself.
The Starbucks case reframes trust as a quantifiable revenue issue rather than a purely ethical one: a single AI-assisted campaign misstep is tied to an estimated $580M annualized loss against $2.1B in regional revenue.
Trust-by-design (avoiding harm) and a positive vision are presented as two separate, both-necessary layers — doing no harm alone is described as only the "ground floor," insufficient to make a product valuable or complete.
Fairness/bias-exclusion is reframed from compliance overhead into a sellable product feature: the talent.com example ties provable fairness directly to whether the government hiring product can be sold at all.
Agent identity is framed as emerging public infrastructure rather than a product-team concern alone, citing national governments (Estonia, Saudi Arabia) building ID systems for agents.
The Ecosia/Google contrast suggests restraint (not forcing an AI feature on users) can be a competitive advantage rather than a liability, positioned against the industry's default push toward AI-forward features.
«if this is not what we want to stand for, if we're not the people that want to get our colleagues sacked, that wants to build untrusty product experiences that create slop and all this we see in the world, quite frankly,»
— 00:33
«This is not the impact we want to have in the world»
— 02:54
«I'm sorry, but that's not what I want to stand for, and I hope none of you would want to be that as well.»
— 03:50
«This is what happens if you build non-trustworthy product experiences.»
— 10:46
«I don't think you can build a useful product if the thing isn't trustworthy.»
— 10:56
«It should be trust over short-term gain.»
— 13:52
«You cannot do something quick and hope that trust will come. It's not going to work that way.»
— 13:57
«let's build trustworthy progress with AI.»
— 26:23
«Not not slop, not hallucinations, not stuff that has the house crumbling down or someone ordering us things that we never intended to have.»
— 26:29
«For me, it's yes to trustworthiness by design and no to defaulting to patterns that are built into those systems today»
— 26:41
«because I hope that the successful products and the impact we create in the world are the things that we can manage to be proud of and it doesn't feel so awful to be in this tech space that has its hands on all of the things that we all don't want to be a part of.»
— 26:49
Reception
No comments are available, so audience reception cannot be determined.
The talk stitches together an ethical framing, one detailed business-case (Starbucks South Korea), and several named frameworks (Covey's Speed of Trust, Eriksson's decision stack, Kore's trustworthy-AI dimensions, the EU AI Act) into a practical call to codify trustworthiness explicitly for agentic co-creators; some references (the Makers Manifesto, the person credited as "Prompt Ledger") are cited by name without elaboration, leaving them as pointers rather than fully explained sources.
27:19