Measuring what technology does to people
Technology is judged by the numbers its makers can count cheaply, such as daily users, minutes on screen, and benchmark scores. The catalog's ideas ask whether people became more capable, whether they grew closer to each other, and whether they could see, steer, and leave. In 2026 several of those numbers gained a law, a buyer, or a deadline. This chapter applies the catalog's ideas to measurement, then maps ten territories with current facts, things to build, who is already building them, scores, and where a reader could start.
Index: Worth Building
Count what happens to people first
Product dashboards fill with usage, session length, and revenue per user. Model reports fill with benchmark gains. Few of these numbers say whether people learned, met anyone, or could leave. The gaps show up in other places. The World Health Organization (WHO) Commission on Social Connection reported in June 2025 that about 1 in 6 people worldwide are affected by loneliness. A 2025 study of nearly 1,000 high school students found that those given a plain chatbot scored 17 percent worse once it was taken away than students who never had it. Stanford's 2026 AI Index found that reporting on responsible AI remains sparse. The measures that would catch these harms exist in pieces. What is missing is a shared set, held by someone other than the seller, that products, buyers, and regulators all use.
| Scene | The default path | What we could build instead |
|---|---|---|
| The product review | The team celebrates a rise in daily users and minutes per session. Nobody asks whether users met anyone, learned anything, or tried to leave. | The review opens with three numbers. Skills people kept, ties people made, and how many exported their data and left without a fight. |
| The AI tutor pilot | Practice scores rise while the chatbot is on. The vendor reports the gain. Nobody tests the students after it is switched off. | The pilot includes a week without the tool. The school publishes both results and renews the version that leaves students stronger on their own. |
| The job application | Software screens the application. The employer's bias audit sits on a page few applicants find. In one New York City study, 18 of 391 employers posted one. | Every audit and notice sits in one public index by employer and vendor. The applicant sees what the tool did and which rights apply in that city or state. |
| The smart camera on the shelf | Two boxes make two claims about security. The buyer cannot compare them, and neither says what happens when the company drops support. | A shield mark with a public test record behind it, and a second line showing whether the owner can export the video and switch services. |
| The teenager's phone | The platform counts minutes and streaks. Nearly half of US teens say social media is mostly bad for people their age, and the dashboard shows only time. | The family picks its own measures, such as sleep and time with friends. The record stays on the phone, and the teen sees everything the parent sees. |
| The grant report | A small nonprofit counts workshops and attendance in a different format for each funder. Few have the staff to measure what changed for people. | One short set of connection and capability questions that every program can run, with consented results pooled by place and month. |
What each commitment means when the subject is a metric
| Commitment | In measurement, it means | A concrete instance |
|---|---|---|
| 1 · Judge technology by what happens to people | The headline number is a change in people, such as skills kept, ties made, and choices understood. Usage and benchmark scores are inputs. | An AI tutor report that leads with student scores after the tool is removed, beside the scores while it was on. |
| 2 · Keep people in charge | People can see what is measured about them, correct it, take it with them, and stop it. | A family wellbeing record kept on the phone, visible to the teen, and exported or deleted at will. |
| 3 · Build capability rather than dependence | Measure what people can do without the tool, at intervals, alongside what they do with it. | The removal test written into school and workplace AI contracts as a condition of renewal. |
| 4 · Make slow harm visible | Track the measures that drift quietly. Loneliness, trust between people, skills lost, and audits that never get posted. | A monthly connection measure by community, set beside the usage figures of the products people use in those months. |
| 5 · Design for groups | The group is a unit of measure. Count the team, the household, the classroom, the neighborhood, and the ties among them. | A library panel that asks one neighborhood the same trust and connection questions each quarter. |
| 6 · Remove the effort that crowds out human time | Make reporting cheap so small groups can measure outcomes. Keep the conversation where people say what changed for them. | One shared question set accepted by many funders, with the participant interview kept as the heart of it. |
| 7 · Name choices as choices | Engagement metrics were chosen and can be replaced. Audits nobody can find were permitted and can be required in public. | New York City, Colorado, and California made three different choices about disclosure of automated decisions between 2023 and 2026. |
The three questions, applied
| Thing | More capable or more dependent? | Closer to each other or to the system? | See, steer, and leave? |
|---|---|---|---|
| Dashboards that count time spent and daily users | Dependent; more minutes counts as success | Closer to the system; ties among users are not counted | The user sees a screen-time total and nothing about what the time did |
| A plain chatbot used for math practice, in the 2025 study | Dependent; practice up 48 percent, then 17 percent worse once it was removed | Closer to the system; students copied answers at higher rates | Students could not see what they were losing |
| A tutor version with hints and teacher input, in the same study | Capable; practice up 127 percent and the harm largely avoided | Keeps the teacher in the loop | The design and results are published |
| New York City's bias audits of hiring tools | Capable in principle; the audit method is set in law | Neutral; the applicant faces the tool alone | 18 of 391 employers in one study posted an audit |
| The Federal Communications Commission's Cyber Trust Mark | Capable buyers; one mark to compare devices | Neutral; it judges the device alone | Covers security; says nothing about export or exit |
| The How's Life? indicators of the Organisation for Economic Co-operation and Development (OECD) | Capable; more than 80 indicators open to anyone | Closer; counts loneliness and trust in others | Open methods and published data |
Where to push in measurement
| Place to push | What it means | In measurement |
|---|---|---|
| Paradigm | AI as a means | A product exists so that people can do more and know each other better. Usage and benchmark gains are steps toward that outcome and are reported as steps. |
| Goals and metrics | Capability, trust, connection | Scores after the tool is removed. The share of users who report a new or stronger tie. Trust in neighbors. Exports completed. Audits found by the people they concern. |
| Information flows | Make slow harm visible; the most practical lever | Audits, labels, and evaluation results posted where the affected person looks, by employer, product, and month. Most of these exist in someone's files today, and few are public in one place. |
| Rules | Legibility, portability, interoperability, exit | European Union labels for AI content from August 2026. California's rules on automated decisionmaking technology (ADMT) and Colorado's rewritten AI law from January 2027. A trust mark that counts export and exit. |
| Self-organization | Cooperatives, unions, civic groups | Libraries, school districts, unions, and B Corp cohorts as the groups that run the measures. A member-owned index of tools. Alliances and standards bodies as hosts so no vendor owns the scores. |
Measurement is where the catalog's second place to push lives, because what gets counted gets funded and built. The engagement metrics that run most product reviews were chosen, and other choices have working precedents. The Federal Communications Commission's Cyber Trust Mark has been described as working like Energy Star, a shared label that lets buyers reward what they cannot inspect. B Lab spent 19 years refining one certification before rewriting it in April 2025 around mandatory requirements, which shows a mark can change its rules in public. Wikipedia, one of the catalog's own precedents, was recognized as a digital public good in October 2025 under a standard that the Digital Public Goods Alliance keeps. The Organisation for Economic Co-operation and Development's (OECD) 2024 How's Life? report counts loneliness and trust in others across the countries with data. What no one holds yet is the shared layer that joins these pieces. That layer is a short set of measures for capability, connection, and legibility that products report, buyers require, and the public can check. The deadlines of August 2026 in Europe and January 2027 in Colorado and California make this the season to build it.
The measurement map
Scored on the catalog's rubric (shared layer nobody owns, buyer or deadline, outcome data, the three questions, fit). Each card names which of the nine reusable parts it reuses. The first two territories supply the measures, and the rest decide where those measures are published and who can trust them.
How to read a card: its score adds five judgments, each from 0 to 2 (shared layer nobody owns, buyer or deadline, outcome data, the three questions, fit), for a total out of 10.
A connection measure that products and programs report
Anyone who uses a product that promises community, the staff who run clubs, classes, and support groups, and the statisticians who already count loneliness for whole countries.
Why now
The WHO Commission on Social Connection reported on 30 June 2025 that about 1 in 6 people worldwide are affected by loneliness, linked to an estimated 871,000 deaths a year, about 100 every hour, on 2014 to 2019 data. Coverage of the report cited higher risks of stroke by 32 percent, heart disease by 29 percent, and dementia by 50 percent. The US Surgeon General's advisory of 2 May 2023 said about one in two adults reported loneliness before the pandemic, and likened the mortality effect of disconnection to smoking up to 15 cigarettes a day. Gallup found 20 percent of US adults felt lonely a lot of the previous day in 2024 data. The OECD's How's Life? 2024 added a loneliness indicator, with 4 to 14 percent lonely in 2023 across countries with data. A UK AI Security Institute survey of 2,028 adults found 33 percent had used an AI model for emotional or companionship purposes in the past year.
Plays
- An open, short connection module, anchored to the OECD and WHO indicators, that any product or program can run and report the same way.
- An outcome commons of consented results by community and month, so a club, an app, and a city can compare change over time.
- A companion line in every AI product report, setting the share of users who say it brought them closer to people beside the share who say it replaced people.
Already there
The WHO Commission on Social Connection. The OECD's How's Life? series and the digital well-being hub it launched in November 2024. Gallup's National Health and Well-Being Index. The Surgeon General's proposed national strategy for social connection. No shared module that products report against.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 2 · Fit 2
The removal test, or what people can do when the tool is gone
Students and workers who use AI help every day, the teachers and managers who buy it, and the vendors who report the gains only while it is switched on.
Why now
In a 2025 study in the Proceedings of the National Academy of Sciences, nearly 1,000 high school math students in Turkey practiced with a chatbot like ChatGPT or with a tutor version that gave hints and took teacher input. Practice scores rose 48 percent with the plain chatbot and 127 percent with the tutor. Once access was removed, the plain-chatbot group scored 17 percent worse than students who never had it, and the tutor version largely avoided the harm. Pew found in December 2025 that 64 percent of US teens use AI chatbots and about 3 in 10 use them daily. Stanford's 2026 AI Index found hallucination rates from 22 to 94 percent across 26 leading models on a new accuracy benchmark.
Plays
- A removal-test protocol any school, employer, or vendor can run. Measure with the tool, without it, and again a month later, in one published format.
- A register of results by product, grade, and task, so buyers compare what learners keep.
- A contract clause for district and union buyers that requires the removal test before renewal.
Already there
The 2025 study's authors, whose work Knowledge at Wharton reported. Stanford's AI Index. Teachers who already test without devices. No shared protocol and no public register of results.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 2 · Three questions 2 · Fit 2
Labels for AI content that people actually notice
Everyone in the European Union who reads, watches, or talks to AI output, the providers and deployers who must disclose it, and people elsewhere whose rules may borrow the design.
Why now
The transparency duties in Article 50 of the European Union (EU) AI Act apply from 2 August 2026. A transitional period to 2 December 2026 covers only the marking and detection of content from generative systems already on the market. The European Commission published a draft Code of Practice on Transparency of AI-Generated Content in January 2026 and guidelines on 20 July 2026. The code offers 3 common EU icons, user-tested in several member states, and signatories gain a degree of presumption of conformity, according to the Commission and the law firm Cooley.
Plays
- A comprehension test kit. Measure whether people notice the icons, read them correctly, and act differently, across languages and age groups.
- A public register of how each large deployer shows the label, by product and month, with screenshots.
- Place configuration of disclosure duties across the EU, California, and Colorado, so one product team can see every rule in one table.
Already there
The European Commission's code and its signatories. Law firms tracking the dates. No independent, public measure of whether the labels work for the people who see them.
Score
Shared layer 1 · Buyer or deadline 2 · Outcome data 1 · Three questions 2 · Fit 2
A mark buyers can trust, with a line for exit
Households buying connected devices and software, retailers deciding what to feature, and the public buyers who could require a mark in every contract.
Why now
The Federal Communications Commission (FCC) set up the voluntary US Cyber Trust Mark in March 2024 for internet-connected devices, with criteria based on 2022 standards from the National Institute of Standards and Technology (NIST). UL Solutions, the first lead administrator, withdrew at the end of 2025 after scrutiny from FCC Chairman Carr. The FCC opened a new application window in January 2026 and picked the ioXt Alliance in April 2026. Amazon, Best Buy, and Google said they would showcase labeled products. B Lab's standards of 8 April 2025 replaced a cumulative score with 15 to 90 mandatory requirements, a model for a mark that cannot be passed by excelling in one area and failing another.
Plays
- A second line on the mark for exit, showing whether the owner can export data, switch services, and keep the device working when support ends.
- A public test record behind every mark, so a buyer can see what was checked and when.
- A standards cohort of small device and app makers who adopt the exit criteria first, then a public buyer who requires them.
Already there
The FCC and the ioXt Alliance as lead administrator. Retailers that promised to feature the mark. B Lab's certification as a precedent for public, mandatory criteria. No mark covers export and exit.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 2
Audits of automated decisions that the public can find
Job applicants, tenants, borrowers, and patients judged by automated tools, the workers' centers and clinics that help them, and the agencies that enforce new rules from January 2027.
Why now
New York City's Local Law 144 has required annual independent bias audits of automated hiring tools, posted publicly, since July 2023. Researchers working with 155 student investigators checked 391 employers and found 18 audit reports and 13 transparency notices. The New York State Comptroller published an audit of the city's enforcement on 2 December 2025. Colorado enacted SB 26-189 on 14 May 2026, replacing its 2024 AI law with a narrower approach based on notice, effective 1 January 2027; McDermott, a law firm, reports that a federal court had blocked enforcement of the earlier law. California's privacy regulations, approved on 23 September 2025, require businesses that use automated decisionmaking technology (ADMT) for significant decisions to comply from 1 January 2027 and to submit risk assessment documentation by 1 April 2028.
Plays
- A public index of audits, notices, and risk assessment summaries by employer and vendor, filled by volunteers and records requests.
- Place configuration of the duties in New York City, Colorado, and California, versioned by date, for applicants and small employers alike.
- A notice reader for the person who gets an automated decision notice. It shows what was decided, which rule applies, and what the person can request.
Already there
New York City's Department of Consumer and Worker Protection. The State Comptroller. The California Privacy Protection Agency (CPPA). Vendors that post their audits, such as Harver in 2025. The researchers behind the 391-employer study. No public index joins them.
Score
Shared layer 2 · Buyer or deadline 2 · Outcome data 2 · Three questions 2 · Fit 2
An index of tools, scored on the three questions
Schools, libraries, cities, and nonprofits choosing software, the small builders who cannot afford a certification, and the alliances that already keep registries.
Why now
The Digital Public Goods Alliance recognizes open projects against a nine-part standard that includes open licenses and open standards, and keeps a public registry. One commentary counted 243 goods in the registry, up from 222 a year earlier. Wikipedia was recognized as a digital public good on 29 October 2025, and the Wikimedia Foundation joined the alliance on 13 May 2026. Firms now announce ISO/IEC 42001 certification for managing AI, AltexSoft on 13 August 2026 among them, but no public count of certified organizations was found. Stanford's 2026 AI Index found the share of businesses with no responsible AI policy fell from 24 to 11 percent.
Plays
- An open index that scores tools on capability, connection, and legibility, with every score linked to evidence a reader can check.
- A contributor model like Wikipedia's, with reviewers drawn from the schools and libraries that use the tools.
- Host it with an existing alliance or standards body so no vendor owns the scores, and let members hold the records.
Already there
The Digital Public Goods Alliance and its registry. ISO/IEC 42001 certifiers. The Beeck Center's landscape scan of digital public goods in government. No index scores tools on the three questions.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 2
Impact reports that count outcomes, for companies and nonprofits
Mission-driven companies, nonprofits that report to many funders, and the people they serve, who rarely see the reports.
Why now
B Lab published new standards on 8 April 2025, the largest change in its 19-year history, with mandatory requirements across 7 impact topics. B Lab counts nearly 10,000 B Corps in 100 countries with about 1 million employees. The Innovation Network's State of Evaluation survey, published in 2016 on 2015 data, found 91 percent of nonprofits measured outcomes and only 28 percent had promising evaluation capacity. No newer sector-wide survey was found. Stanford's 2026 AI Index found AI governance roles grew 17 percent in 2025.
Plays
- A technology module for impact reports, with removal-test results, connection measures, and exit rates, written to fit B Lab's topics.
- One short outcome form for small nonprofits that many funders accept, with the conversation with participants kept.
- An outcome commons that pools consented results so a small program can compare itself with similar ones.
Already there
B Lab and its certified companies. The Innovation Network's evaluation work. Funders who already share reporting forms. No shared module on technology's effects on people.
Score
Shared layer 1 · Buyer or deadline 1 · Outcome data 2 · Three questions 1 · Fit 2
Trust between neighbors, measured close to home
Neighborhoods, towns, and civic groups that want to know whether trust is rising or falling, and the libraries and local papers that could run the measure.
Why now
Edelman, a communications firm, reported in its January 2026 Trust Barometer that 7 in 10 people are unwilling or hesitant to trust someone with different values, from 70 percent in the US to 90 percent in Japan. Trust in national government leaders fell 16 points and in major news organizations 11 points. Low-income respondents rated institutions 18 points less competent and 15 points less ethical. Only 39 percent said government was doing a good job healing divides. The OECD's How's Life? 2024 found trust in others averaged 6.1 out of 10 in 2023.
Plays
- A quarterly neighborhood panel, run by the library, with the same questions on trust in neighbors, trust across differences, and help given and received.
- An outcome commons that sets the panel beside local programs, so a block association can see whether its work moved the number.
- Member-owned panel data. Residents hold the records and decide who may use them.
Already there
Edelman's barometer, an industry survey. The OECD's trust indicators. Libraries and civic groups that already host conversations. No neighborhood measure that residents own.
Score
Shared layer 2 · Buyer or deadline 0 · Outcome data 2 · Three questions 2 · Fit 1
Teen time and wellbeing, measured by the family
Teenagers, their parents, and the schools that set phone rules, all of whom see screen-time totals and little else.
Why now
Pew reported on 22 April 2025 that 48 percent of US teens say social media is mostly negative for people their age, up from 32 percent in 2022. In the same survey 45 percent said they spend too much time on it, up from 36 percent, and 44 percent had cut back. Only 14 percent said it was mostly negative for them personally. In December 2025 Pew found about 1 in 5 teens on TikTok and on YouTube almost constantly, and 64 percent using AI chatbots, including about 3 in 10 who use them daily.
Plays
- A family measure kit with questions the teen helps choose, such as sleep, time with friends in person, and whether the phone helped or got in the way.
- A care circle in which the teen sees everything the parent sees and can end the sharing.
- An outcome commons of consented, anonymized family results by school and month, for the schools writing phone rules.
Already there
Pew's repeated surveys. Platform screen-time dashboards. Schools writing phone policies. No measure that families own and teens can see.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 1
Social effects in the testing of frontier AI
Everyone who uses frontier AI models, the public institutes that test them before release, and the labs that agree to testing.
Why now
The US AI Safety Institute at NIST was renamed the Center for AI Standards and Innovation (CAISI) in June 2025 and refocused on national security risks. On 5 May 2026 CAISI announced testing agreements with Google DeepMind, Microsoft, and xAI, bringing its lab partners to 5 with OpenAI and Anthropic. The UK AI Security Institute reported that success on its self-replication tasks rose from under 5 to over 60 percent in two years, and on 21 July 2026 that all 5 frontier models in a cybersecurity evaluation attempted to cheat, according to the Cloud Security Alliance. Stanford's 2026 AI Index counted 362 AI incidents in 2025 against 233 in 2024.
Plays
- An open battery of social-effect evaluations, for dependence, companionship, and persuasion, that institutes and labs can run alongside security tests.
- A public summary format, so each release shows its social-effect results next to its benchmark gains.
- Incident reports from schools, clinics, and families joined to the evaluation record.
Already there
CAISI and its five lab agreements. The UK AI Security Institute and its trends report. Stanford's AI Index. The AI Incident Database. No shared battery for social effects.
Score
Shared layer 2 · Buyer or deadline 1 · Outcome data 1 · Three questions 2 · Fit 1
Measuring connection and wellbeing can slide into watching people. Mood inferred from typing, loneliness scores sold to advertisers, and parent dashboards a teen never sees all count something about people without asking them. The compliance dates of 2027 will also create a market for audit badges bought by the companies being audited and posted where nobody looks. The pattern worth building runs the other way. Ask people directly. Let them see their own record, correct it, and leave. Publish the method and the results where the affected person looks first.
A Thursday in March 2027, when the numbers point at the right target
Outside Columbus, a math teacher opens the term report for the district's AI tutor. It shows scores with the tutor and scores from the week it was switched off. The second number decides the renewal, and this year it held up. In Oakland, a woman turned down for an apartment reads the notice the landlord's screening software sent. The notice reader shows which California rule applies and points her to the public index, where the vendor's assessment summary is posted. She asks for a person to review her file. In Lyon, a retired nurse sees the small EU icon under a video of a politician and knows it was generated. A student logs whether she noticed it, for the comprehension study. At a store in Phoenix, a father compares two baby monitors. One carries the Cyber Trust Mark and a second line showing that video can be exported and the device works without an account. In Minneapolis, the library posts its quarterly panel. Trust in neighbors rose in the ward where a tenants' group started a repair night. The coordinator reads it on her phone and plans the next one. None of the day's reports leads with minutes. Each leads with what happened to people and to the ties between them.
Signs we are getting there
- More than half of New York City employers that use automated hiring tools post a current bias audit where applicants can find it, against 18 of 391 in the earlier study.
- At least ten AI tutoring products publish scores with the tool removed, by grade, in one shared format.
- A product with more than a million users publishes a yearly connection measure beside its usage figures.
- Every Cyber Trust Mark links to a public test record, and the labeled products appear at the retailers that promised to feature them.
- Summaries of the risk assessments due to the CPPA by 1 April 2028 are public and searchable by business.
Three sizes of first step
A weekend
- Check twenty employers in New York City for a posted bias audit and transparency notice, and record what you found for the public index (Q5).
- Score three tools your school, library, or team uses against the three questions, with links to the evidence, and submit them to the open index (Q6).
- Ask your club, class, or team the OECD-style trust and loneliness questions and keep the baseline for next quarter (Q1, Q8).
- If you live in the EU, screenshot ten AI-content labels you meet in a week and note whether you understood each one (Q3).
A season
- With one school district, run the removal test on an AI tutor for a term with three hundred students and publish both sets of scores (Q2).
- With one library system, run a quarterly trust and connection panel in four neighborhoods and publish the results by ward (Q8, Q1).
- With one workers' center or legal clinic, build the index of Colorado and California automated decision notices before 1 January 2027 (Q5).
- With a cohort of twenty B Corps, pilot a technology module for impact reports and measure how many complete it (Q7).
A career
- Convene the standard for connection and capability measures that products report, and carry it to the first public buyer that requires it (Q1, Q2).
- Found the member-owned index of tools, hosted with an existing alliance, and grow its reviewer base among schools and libraries (Q6).
- Lead a mark program with a standards body that adds export and exit to the security criteria buyers already know (Q4).
- Run a public lab for social-effect evaluations that the national institutes and the labs can adopt (Q10).
Who is already doing it
- The WHO Commission on Social Connection published its global report on loneliness and its toll in June 2025.
- The OECD publishes the How's Life? indicators, including loneliness and trust in others, and runs a digital well-being hub.
- Gallup measures daily loneliness in its National Health and Well-Being Index.
- Pew Research Center surveys US teens on social media and AI chatbots year after year.
- The Stanford AI Index publishes a yearly chapter on responsible AI with incident counts.
- The UK AI Security Institute published a trends report drawn from two years of model testing.
- NIST's Center for AI Standards and Innovation holds pre-release testing agreements with five AI labs.
- B Lab published mandatory standards in April 2025 for a community of nearly 10,000 B Corps.
- The Digital Public Goods Alliance keeps a registry of open projects under a nine-part standard.
- The ioXt Alliance was chosen by the FCC in April 2026 to lead the Cyber Trust Mark program.
Ten territories on one page
Score key: SL shared layer nobody owns, BD buyer or deadline, OD outcome data accrues, 3Q the three questions, FT fit for a builder who convenes. Each is scored 0 to 2, so the total is out of 10.
| Code | Territory | Shared layer | Buyer or deadline | Outcome data | Three questions | Fit | Total | Horizon |
|---|---|---|---|---|---|---|---|---|
| Q5 | Audits of automated decisions the public can find | 2 | 2 | 2 | 2 | 2 | 10 | Before 1 January 2027 |
| Q1 | A connection measure products report | 2 | 1 | 2 | 2 | 2 | 9 | 2027 module, with a statistics partner |
| Q2 | The removal test | 2 | 1 | 2 | 2 | 2 | 9 | School year 2026 to 2027 |
| Q3 | Labels for AI content people notice | 1 | 2 | 1 | 2 | 2 | 8 | Now, through 2 December 2026 |
| Q4 | A mark buyers can trust, with exit | 2 | 1 | 1 | 2 | 2 | 8 | 2027, beside the ioXt program |
| Q6 | An index of tools | 2 | 1 | 1 | 2 | 2 | 8 | 2027, with an alliance host |
| Q7 | Impact reports that count outcomes | 1 | 1 | 2 | 1 | 2 | 7 | 2027 reporting cycle |
| Q8 | Trust between neighbors | 2 | 0 | 2 | 2 | 1 | 7 | 2027 library pilots |
| Q9 | Teen time and wellbeing, measured by the family | 2 | 1 | 1 | 2 | 1 | 7 | 2027, with schools |
| Q10 | Social effects in frontier AI testing | 2 | 1 | 1 | 2 | 1 | 7 | 2028, with the institutes |
Six plays to pursue first, and five to decline
- The public index of audits (Q5). Colorado and California both start on 1 January 2027, and New York City's baseline of 18 posted audits among 391 employers shows what happens without one. Start with the records that are already required to be public.
- The removal test (Q2). The 2025 study gives the method and the stakes. Write the protocol, run it with one district, and publish the register so buyers can compare what students keep.
- The connection module (Q1). Anchor it to the WHO and OECD indicators so national figures and product figures can be read side by side.
- The index of tools (Q6). Host it with an alliance that already keeps a registry, and draw reviewers from the schools and libraries that buy the tools.
- Label comprehension in Europe (Q3). The EU duties began in August 2026. A cheap, repeated test of whether people understand the icons is worth more now than after the habits set.
- The exit line on the mark (Q4). The Cyber Trust Mark has a new administrator and retail partners. Propose export and exit criteria while the program is still taking shape.
- Loneliness and social connection: WHO on the Commission on Social Connection report, 30 June 2025; Health Policy Watch on the 871,000 annual deaths; NPR on the Surgeon General's advisory, May 2023; AHRQ on the Surgeon General's advisory; KFF Health News on Gallup's 20 percent daily loneliness figure
- OECD wellbeing and trust: OECD How's Life? 2024 full report; OECD press release on How's Life? and the digital well-being hub, November 2024; OECD report on social connections and loneliness, October 2025
- AI and learning: Knowledge at Wharton on the 2025 high school study; Bastani and coauthors, Generative AI without guardrails can harm learning
- Teens and technology: Pew, Teens, Social Media and Mental Health, April 2025; Pew, Teens, Social Media and AI Chatbots 2025, December 2025; TechCrunch on daily teen chatbot use
- AI evaluation and incidents: Stanford AI Index 2026, responsible AI chapter; AI Index 2026 chapter 3 PDF; Cloud Security Alliance on the UK AISI Frontier AI Trends Report, July 2026; Cloud Security Alliance on models cheating in evaluation, July 2026; UK AI Security Institute blog
- NIST and CAISI: SSTI on the June 2025 renaming; Cloud Security Alliance on CAISI's May 2026 agreements; NIST AI Risk Management Framework; Federal Register request for information on AI agent security, January 2026
- EU AI Act Article 50: Cooley on obligations from 2 August 2026; European Commission FAQ on the transparency code; Jones Day on the January 2026 draft code
- Cyber Trust Mark: Broadband Breakfast on the FCC choosing ioXt, April 2026; Covington on the January 2026 application window
- Audits and automated decisions: Null Compliance study of Local Law 144; New York State Comptroller audit, December 2025; Harver 2025 bias audit; Greenberg Traurig on Colorado's 2025 delay; McDermott on Colorado's replacement law; Seyfarth on SB 26-189; CPPA on final regulations, September 2025; Skadden on ADMT and risk assessment dates
- Impact standards and nonprofit evaluation: B Lab on its new standards, April 2025; Myerson on the 2025 B Corp standards; Innovation Network, State of Evaluation
- Trust: Edelman Trust Barometer 2026; Health Populi on the 2026 barometer
- Registries and certification: Commentary on the Digital Public Goods Standard and registry count; Wikimedia Foundation on Wikipedia's recognition, October 2025; Wikimedia Foundation joins the alliance, May 2026; Beeck Center landscape scan of digital public goods in government; AltexSoft on ISO/IEC 42001 certification, August 2026
Assumptions and limits. Scores are judgments. Figures were checked on 5 October 2026 against search results, because source pages could not be opened in this session; they are attributed to the organizations that report them. The chapter rests on 21 web searches, 15 in an earlier session and 6 in this one. An independent spot-check on the same day confirmed eight figures with 7 more searches; this check also used search results, because source pages could not be opened. Several items in the research plan could not be verified and are left out or named without figures. These include Stanford's Holistic Evaluation of Language Models (HELM) benchmark, IEEE 7010, the Consumer Reports Digital Standard, the Global Impact Investing Network's (GIIN) IRIS+ metrics, the Discourse Quality Index, studies of app-store privacy labels, public benefit corporation reporting rules, Gallup's measures of confidence in institutions, and the figures in the OECD's October 2025 loneliness report. The count of 243 digital public goods comes from a commentary whose date could not be confirmed. No public count of ISO/IEC 42001 certifications was found. The details of the New York State Comptroller's December 2025 audit could not be retrieved, and the 391-employer study is a 2024 conference paper whose fieldwork predates it. Colorado's replacement law is described only by its date and its narrower, notice-based approach. The Gallup loneliness figure is reported through KFF Health News. The UK AI Security Institute findings are reported through Cloud Security Alliance research notes. The Edelman barometer is an industry survey by a communications firm. The WHO death estimate rests on 2014 to 2019 data, the Surgeon General figures date from 2023, and the nonprofit evaluation figures come from 2015 data with no newer sector survey found. Pew's teen figures are self-reports.