Module 3, Oct 21, Seminar and lab. Intake conversations are completed this week.

The structure of the technology industry

Background

The people who build a product decide what it does, and some of those roles are open to clinicians and social workers. Grow Therapy’s clinical quality and documentation integrity manager posting accepts a licensed clinical social worker among its eligible credentials and puts that person on the clinical review layer for product initiatives, checking that clinical requirements are represented correctly before anything ships. AI governance work has become its own profession, and IAPP’s 2025 survey found 77 percent of organizations building AI governance programs, rising to about 90 percent among organizations already using AI. The function has no settled organizational home, splitting mainly across privacy, legal, and IT teams, and only 1.5 percent of the surveyed organizations reported no additional AI-governance staffing needs for the coming year. Supporters argue that changing who holds these roles changes the product before it ships.

The inside roles are also the first cut when money tightens. Across the 2022 and 2023 layoffs, Microsoft cut its ethics and society team from 30 people to zero in two rounds, Twitter cut its ethical-AI team from 17 people to one along with about 15 percent of its trust and safety staff, Meta cut roughly 200 content moderators and more than 100 trust and integrity roles, Google cut about a third of a unit working on misinformation and radicalization, and Twitch cut roughly 50 of its 330 trust and safety staff. A role that exists only at the company’s discretion can be cut whenever the company chooses. The public materials do not record the cuts: Microsoft’s current responsible-AI page presents six named principles and an Office of Responsible AI that publishes an annual transparency report, with no mention of the team eliminated in 2023.

The symbolic version of inside ethics predates those layoffs. Susser catalogued it in 2019: Google announced an external ethics advisory board and then dissolved it, and he argued that internal ethics structures can work as a company’s conscience but cannot substitute for legal constraints and outside oversight. Tech Policy Press documents “grantwashing,” small grants and advisory roles that buy an ethics story without changing what the company does. Its lead example: OpenAI announced up to $2 million in mental health research funding, in grants of $5,000 to $100,000 against a median NIH grant of $642,918 in 2024, one week after its lawyers argued in court that the company bore no responsibility for a teenager’s death involving ChatGPT. Meta ran a similar play in 2019, giving $50,000 grants to six scientists while its lawyers pressed internal researchers to limit the company’s liability.

The correction that stuck at one AI vendor came from outside the company, when the FTC fined an accessibility-overlay firm $1 million for claiming its AI made websites compliant when it did not; the agency also found the firm had disguised paid reviews as independent opinion, and the final order, approved in April 2025, runs for twenty years with annual compliance reports. On this reading, regulation and outside pressure move a company further than one governance hire can. Simpson and Conner’s proposal would build that pressure into structure, a three-tier regulatory framework with a gatekeeper tier for dominant platforms and proactive rulemaking instead of after-the-fact enforcement. The outside lever has its own capacity problem: the FTC runs on roughly 1,160 full-time staff, about 40 of them dedicated to data protection, around a third of the headcount at Ireland’s Data Protection Commission alone.

Key ideas

Product cycles and economics

A product team ships on cycles. Work is scoped into sprints, ordered on a roadmap, and released before it is finished, on the plan that use will show what to fix next. The money runs on vendor economics. Most tools are priced per seat, so a vendor’s revenue grows with its largest customers, and a roadmap set by the biggest accounts does not bend for one small agency’s request.

Pahlka’s Recoding America locates a related gap in government. She argues that American government splits policy-making from implementation, and that delivery, building things that work for the people who use them, is the central unaddressed problem.

What the customer side of that market looks like is measurable. A late-2025 survey of 346 nonprofits found 92 percent using AI in some capacity while only 7 percent reported major gains in organizational capability. Use stayed reactive and individual in 65 percent of organizations, only 18 percent reported team-wide operational use, and nearly half had no formal AI governance policy. Small organizations reported moderate impact slightly more often than large ones, 41 against 34 percent, so organizational scale is not what separates the 7 percent from the rest.

The role map

A product organization splits its decisions by role: a product manager decides what gets built and in what order, engineers judge feasibility, designers shape the interaction, researchers supply the evidence about users, and trust and safety and policy staff set the limits. The ecosystem page lays out each role and where social work training already fits. The interactive role map lets you click through the roles and see where a social-work graduate lands.

The Integrity Institute’s conversation with Dave Willner, then head of trust and safety at OpenAI, describes that role from inside a frontier AI lab. The work is distinct from social-media content moderation and from alignment research, and it runs from day-one moderation of a newly released model such as DALL-E to longer-horizon policy questions about what the systems should refuse.

The evidence role has a social-work track record. Frey, Patton, Gaskell, and McGregor hired two formerly gang-involved young men from Chicago as domain experts to annotate Twitter data from gang-involved youth, and the pair taught MSW student annotators to read context outsiders miss, such as marijuana-strain names referencing deceased rivals. The domain-expert-informed tagger reached 89.8 percent accuracy on their Twitter corpus against 81.5 percent for off-the-shelf tools, a measured gain from putting lived expertise inside the research role.

Product roles and governance roles

There are two distinct ways in. A product or user-research role sits on the build side, where a needs assessment is user research and case coordination is stakeholder alignment under other names, and the work decides what gets built. A responsible-AI or governance role sits on the review side, where documented risk assessment, safety planning, and negotiating consent under a power imbalance are already daily social-work skills, and the work decides what is allowed to ship.

The review route has named examples. Leslie Taylor, MSW, moved from a $40,000-a-year job as an intensive in-home family therapist to trust and safety at the National Center for Missing & Exploited Children, then to Snapchat as the fourth hire on its trust and safety team, then to Adobe. Her resume translations run straight from casework to product language; “prevented out-of-home placements” becomes “implemented safety interventions.”

The review side is also professionalizing into credentials. IAPP’s AI Governance Professional certification tests four domains, responsible-AI foundations, how laws such as the EU AI Act apply, the AI lifecycle and risk management, and emerging governance issues, and 23.5 percent of organizations in IAPP’s survey said finding qualified AI professionals was part of what made delivering AI hard.

Worker organizing

Holding a role is one inside position; organizing across roles is another. The Tech Workers Coalition is an all-volunteer coalition of tech workers and community advocates with local chapters in the Bay Area, Portland, and New York, a union job board, and published organizing guides. Its current campaigns include tech-sector union drives and workplace organizing around AI’s effect on jobs, pressure that does not depend on any single governance hire surviving the next budget cycle.

The grantor-grantee relationship

The newest structure between tech companies and the social sector is money. Companies fund nonprofits with cash, credits, and embedded staff, and the terms shape what the nonprofit can build and keep. The ecosystem’s funding section collects the programs and the two questions that travel across all of them: cash or credits, and what the funder gets back.

The structure has a longer history than the current AI programs. Gunther’s study of the Silicon Valley Community Foundation traced how it grew to $8.2 billion in assets on gifts including roughly $1 billion from Mark Zuckerberg, while 98 percent of its $1.3 billion in 2016 grants moved through donor-advised funds that donors effectively control and only $19.2 million of its own endowment went to local grants. Tech wealth converted into philanthropy, and the conversion preserved donor control at each step. In the current AI programs, control arrives through different instruments: credits denominate an award in the funder’s own product, and embedded staff carry the funder’s practices into the grantee’s daily work.

The case: the Trevor Project’s Crisis Contact Simulator

The Trevor Project’s Crisis Contact Simulator is a counselor-training tool built to close a specific gap. Youth reaching out to the LGBTQ crisis line nearly doubled after the pandemic started, and the organization needed to grow its team of 700 volunteer crisis counselors tenfold while training nearly 70 percent of them outside business hours, on nights and weekends. Over two years, Google.org gave $2.7 million in funding and nearly 30 Google.org Fellows to help Trevor build the simulator, pooling expertise in machine learning, natural language processing, product management, user experience, and clinical psychology.

The simulator’s youth personas came out of six months of research and thousands of role-play transcripts between Trevor’s own training team and its volunteers. “Riley,” the first persona, plays a young person struggling to come out as genderqueer; a trainee’s conversation with Riley is scored against a rubric the training team wrote for whether the exchange was sensible, specific, authentic, and met its learning objective. An initial cohort of Trevor Project staff trained as crisis counselors using the tool before it moved into the wider training curriculum.

The case shows both structures at once. The clinical judgment stayed with Trevor: its training team supplied the transcripts, wrote the personas, and owned the scoring rubric, while the Fellows supplied machine learning, natural language processing, product management, user experience, and clinical psychology expertise on the company’s clock. The announcement runs on the funder’s own blog, and what Trevor pays to maintain the simulator after the Fellows rotate out is the one figure the post does not give.

Before class

In class

Seminar

Vote, then how a product organization actually runs: cycles, roadmaps, and why a per-seat price means a small agency is rarely the customer the roadmap serves. We map the roles, separate product roles from governance roles, and read the two-sided evidence, the clinical license named in a product-team hiring bar against the ethics team cut to zero.

Jes Kane, who worked on the Trevor Project’s crisis simulator from Google.org’s side, joins as guest to describe how these product and funding decisions look from inside the industry, with time for your questions.

Lab: the funder grant teardown

The lab is one hands-on teardown of a real tech-company funding program, done live. Each group claims one program from the ecosystem’s funding section, OpenAI’s People-First AI Fund, Google.org’s Generative AI Accelerator, Microsoft for Nonprofits, the AWS Imagine Grant, the McGovern Foundation, or Fast Forward, and reads its current terms on the funder’s own site during class. Four questions structure the read: whether the program gives cash or credits, what the funder gets back, what collaboration or reporting the grant obligates, and what year five looks like once the grant or the credits end. A credits-denominated award is really vendor pricing in another form, so the same teardown decodes the underlying product’s per-seat economics and names who inside the company sets that price. Groups compare findings at the end, and the pattern across programs is what each group carries into its pitch.