Chapter 1 · Peoplarium

Research

How do you find out what is true?

We ran 28 research tasks in parallel with AI research agents, in four waves, and then checked the decision-relevant claims against their primary sources. The checks caught real errors, including a tenfold cost mistake and a product described as dead that had relaunched, so every published finding carries its source, its date and its check status.

As of September 2026 · Peoplarium is a Digi117 venture in development.

  • AI
  • Trust and safety
  • Growth

How the research ran

Each task had a written brief: one question, the decision it fed, and the format of the answer. AI research agents ran the tasks in parallel and returned a short answer, numbered findings and dated links. Every finding was labeled as a fact, an opinion or an interpretation. Sources older than twelve months were flagged.

Wave What it answered Tasks
A Decisions pending now: verification, AI-sent messages, pricing, voice, profiles, onboarding 6
B MVP shaping: fatigue, safety, location, communities, hosts, launch city, video 7
C Later phases and costs: personals, growth loops, AI unit costs, video rooms, pricing that backfired 7
D Next-round decisions: Vancouver launch, emergency help in Canada, CAD pricing, badge wording, onboarding 8

Research is fast and cheap this way. It is also wrong often enough that no answer goes into a plan unchecked.

How every finding was checked

Three reviewers read every result in full. They checked 15 of the most decision-relevant claims against the original source, and 9 of the 15 were wrong or only partly right. For this site, each published number was opened again at its source.

The checks follow one rule: a primary source wins over a summary, and a model's paraphrase is never evidence.

  • Numbers are traced to the page that states them. If the page cannot be opened, the number is dropped.
  • Quotes from social media that the research paraphrased are treated as color, not evidence. None are published.
  • Dates matter. A 2019 settlement figure had been superseded by a 2026 one; an app launch was dated two years late.

What the checks caught

These corrections are the best advert for checking. Each one would have changed a decision.

  • A tenfold cost error. The first per-member cost model put identity checks at 15% of 1,000 members × $1.20 = $18. The right answer is $180. Corrected, AI and trust costs came out at roughly double the first estimate. Finding 08
  • A product described as dead. Couchsurfing Hangouts was reported dead. It relaunched in 2026 with a coarse, privacy-safe map, which supports the design choice rather than undermining it. Finding 11
  • Unverified statistics. Hinge video-prompt figures (+62% likes, 3.4× conversation starters) exist only in a coaching blog citing internal data. They are not used. Finding 20
  • A rule applied too widely. Tinder's Face Check was mandatory first in California, not for every new US user. Its results were confirmed. Finding 03
  • A misread paper. "Three times more willing to send through your own AI" was an engagement lift from routing, not a preference. Finding 13
  • A public price called private. A voice-moderation vendor's prices were described as unpublished. They are on its website.
  • The wrong country. Tinder's USD 60.5M age-pricing settlement is a California class action, not a Canadian one. Finding 18

The findings

23 of the 28 results are published here. The other five are kept back: they cover roadmap decisions that are not public, vendor comparisons and an internal memo. Filter by theme, research wave, claim type or check status; the "corrected" filter shows where the first answer was wrong and what fixed it.

23 findings

  • Finding 01 · Demand and fatigue

    Swipe fatigue is a loop, not a one-time churn

    A Tawkify survey of 1,000 single Americans found users swipe about 58 times before a meaningful conversation and delete and reinstall dating apps about six times a year. Fatigue is a repeating cycle, not a single exit.

    Corrected: Those shares and quotes were synthesized by the model from qualitative sampling, not measured or quoted. They are not published; only the survey figures are kept.

    CorrectedFactWave BChecked 2026-09-28

  • Finding 02 · Identity and profiles

    Personals work on structure; karma gates don't stop bots

    Text-first personals on Reddit sort people by structured posts, not photos, but reputation gates such as karma no longer stop bots. In 2026 Reddit itself moved to human verification for accounts that look automated, while removing about 100,000 accounts a day.

    ConfirmedInterpretationWave CChecked 2026-09-28

  • Finding 03 · Identity and profiles

    A verification badge builds trust a public face cannot

    Verification lifts trust without forcing faces: Feeld reached about 38% of profiles verified in under a year, and Tinder's Face Check cut exposure to likely bad actors by over 60% and bad-actor reports by over 40%.

    Corrected: Face Check was mandatory for new users in California and seven countries, and was then being expanded across the US. The 60% and 40% figures are confirmed.

    CorrectedInterpretationWave AChecked 2026-09-28

  • Finding 04 · Identity and profiles

    A 26-minute AI interview gets dates but is too long a gate

    Known's voice AI interview averages 26 minutes and the company claims 80% of its introductions led to in-person dates; personality similarity, by contrast, explains very little of how satisfied couples are.

    ConfirmedInterpretationWave AChecked 2026-09-28

  • Finding 06 · Growth and launch

    Invite links grew Partiful to 500,000 monthly users

    The strongest consumer-social growth loop of 2023 to 2026 was the event invite link: every invitation brings in people who were not users. Partiful reached about 500,000 monthly active users in early 2025, up about 400% in a year, mostly on that loop.

    ConfirmedFactWave CChecked 2026-09-28

  • Finding 07 · Growth and launch

    Vancouver wins as launch city on access, not size

    Across eight candidate cities, a weighted scorecard put Vancouver first, mainly on founder access, newcomer share and room beside competitors rather than singles density. Timeleft is already live there; a claimed gap for 222 could not be confirmed.

    Corrected: Timeleft's locations page claims 160+ cities in 48 countries, Vancouver included. 222's absence from Vancouver could not be confirmed, so it is not counted as an advantage.

    CorrectedInterpretationWave BChecked 2026-09-28

  • Finding 08 · AI and costs

    AI and trust cost per member was understated by half

    The first research put AI and trust costs at about USD 0.27 per member a month at medium use; a 10x error in the verification line meant the real figure was about USD 0.54. In the current model, AI model usage is the largest cost at about CAD 0.21 per member a month.

    Corrected: 15% of 1,000 members at USD 1.20 is USD 180, not 18. Corrected totals are about USD 0.26, 0.54 and 0.86 per member a month. The USD 1.20 check price itself was not a published price.

    CorrectedFactWave CChecked 2026-09-28

  • Finding 09 · Safety

    Tinder's panic button does not work in Canada

    Noonlight, the emergency service built into Tinder, is available only in the US. A Vancouver launch cannot borrow the panic-button story; it needs its own path to emergency help and trusted contacts.

    Corrected: Match Group's 22 October 2025 release says Face Check was required for new users in California and seven countries, and was then expanding across the US. The 60% and 40% reductions are company figures. The Noonlight finding itself is confirmed.

    CorrectedFactWave BChecked 2026-09-28

  • Finding 10 · Safety

    Live voice moderation catches what users never report

    In Call of Duty, AI voice moderation cut the share of players exposed to severe toxic voice chat by about 50%, and only one in five of the severe cases it flagged had also been reported by players.

    ConfirmedFactWave AChecked 2026-09-28

  • Finding 11 · Safety

    People-nearby features only work with coarse location

    Products that show who is nearby right now hold up when location is approximate, interest is mutual and meetings happen in public places; live maps of people create stalking risk. Couchsurfing relaunched Hangouts in May 2026 on that model.

    Corrected: Couchsurfing relaunched Hangouts with a map in May 2026. Others see only a general area, never a member's exact location. The relaunch supports the coarse-location rule rather than showing the idea failed.

    CorrectedInterpretationWave BChecked 2026-09-28

  • Finding 12 · Pricing and money

    People pay for time saved, not to be seen

    Dating users resent paying for visibility and scarcity, and the law has punished the worst tactics: Match Group paid $14M to the FTC in 2025 over guarantees and cancellation, and Tinder's $60.5M age-pricing class action was finally approved in June 2026.

    Corrected: That earlier deal was superseded. A $60.5M settlement of the California class action received final approval on 9 June 2026. It is a California case, not a Canadian one.

    CorrectedInterpretationWave AChecked 2026-09-28

  • Finding 13 · AI and costs

    Only 4–13% of random pairs would accept AI-sent messages

    A survey of 2,617 users of a large dating platform found that, under random pairing, only 4–13% of sender and receiver pairs combine a sender willing to use an AI agent with a receiver willing to engage with it.

    Corrected: That misreads the paper. Its 3.1 times figure is the engagement lift when agent messages are routed to the most receptive users. The 4–13% pair figure is right.

    CorrectedFactWave AChecked 2026-09-28

  • Finding 14 · Growth and launch

    Discord bans bots that profile users' relationships

    Discord's Developer Policy forbids using API data to profile users, their identities or their relationships with other users, and forbids scraping. A bot that matches people from their server membership would break platform policy, not just feel intrusive.

    ConfirmedFactWave BChecked 2026-09-28

  • Finding 15 · Demand and fatigue

    Video-first social apps fade; video works as a consented step

    Video-first social products rise fast and rarely last: Houseparty drew about 50 million sign-ups in lockdown and shut in October 2021, and the video speed-dating app Filteroff was sold in October 2024. Video holds up as a consented step before meeting, not as the product.

    Corrected: The October 2024 sale is confirmed; the later closure could not be confirmed in a primary source. The cost figures belonged to the older stack. The current plan runs voice rooms on Cloudflare Realtime, so those vendor prices are not published.

    CorrectedInterpretationWave CChecked 2026-09-28

  • Finding 17 · Growth and launch

    Stranger meetups can grow without paid table hosts

    Creator-led apps such as Pie pay organizers per RSVP, while Timeleft grows with no host at the table: an algorithm forms each group and venue staff seat it. The research recommended paying Peoplarium hosts on show-up; the current plan pays no hosts.

    Corrected: Pie's current Creator Club help pages do say 1 point equals $1, earned per accepted RSVP, but TechCrunch reported $5 to $10 per RSVP in November 2024. The rate has changed over time, and neither figure is used as a benchmark.

    CorrectedFactWave BChecked 2026-09-28

  • Finding 18 · Pricing and money

    Pricing tricks cost Match brands USD 14M and USD 60.5M

    In August 2025 the FTC ordered Match Group to pay USD 14 million over misleading guarantees, hard cancellation and suspending users who disputed charges. Tinder's age-based pricing ended in a USD 60.5 million California class settlement, given final approval in June 2026.

    Corrected: The current settlement is USD 60.5 million, a California class action approved in June 2026. The source checked says users 29 and over, and later 28 and over, paid more; it gives no figure for how much more.

    CorrectedFactWave CChecked 2026-09-28

  • Finding 19 · Identity and profiles

    Profile music works when the viewer presses play

    Instagram brought back the MySpace profile song in August 2024 with a clip of up to 30 seconds that never autoplays; Spotify's widget terms bar its Play Button from commercial use.

    ConfirmedFactWave AChecked 2026-09-28

  • Finding 20 · Identity and profiles

    Profile video may help; video feeds stall

    Short video works as an optional part of a profile, while feed-first video dating apps such as Snack stalled. Widely repeated figures for Hinge video prompts, 62% more likes and 3.4 times more conversation starters, could not be traced to Hinge.

    Corrected: Hinge launched video prompts on 20 September 2022. Its announcement gives no such figures; they appear only in a coaching blog citing unnamed internal data, so they are treated as unverified.

    CorrectedInterpretationWave BChecked 2026-09-28

  • Finding 21 · Growth and launch

    Stranger dinners already run in Vancouver

    Vancouver already has stranger-dinner clubs: Timeleft lists it among six Canadian cities, and Well Fed seats quiz-matched tables of six. The gap is not dinners but verified members who pick their own intent and keep meeting.

    Corrected: Meetable's site lists no Canadian city. 222's Vancouver status is still unconfirmed; the public review names only US cities. No public source was found for the verification price.

    CorrectedFactWave DChecked 2026-09-28

  • Finding 22 · Safety

    Noonlight is US-only, and Canada has no dispatch API to plug in

    Noonlight's emergency response does not work outside the United States. Life360 offers 24/7 emergency dispatch in Canada, but only to its own paid members, with no API for other apps.

    ConfirmedFactWave DChecked 2026-09-28

  • Finding 23 · Pricing and money

    BC law makes cancel-anytime the floor for a monthly plan

    From 1 August 2026, British Columbia allows a subscription of 60 days or less to renew automatically only if the customer can cancel at any time, without penalty, and get a refund. Peoplarium's Plus is set at CAD 19.99 a month or CAD 191.90 a year.

    Corrected: Side claim only: the Tinder settlement is a California class action, not Canadian law. It still argues for one price for everyone. The BC subscription finding itself is confirmed.

    CorrectedFactWave DChecked 2026-09-28

  • Finding 24 · Law and policy

    A face-private profile can say "Verified Real" if the badge is exact

    Canada's Competition Act judges claims by their general impression, so a verification badge on a profile without a public face must say exactly what was checked: "Verified Real · ID checked", with an explainer.

    ConfirmedInterpretationWave DChecked 2026-09-28

  • Finding 25 · AI and costs

    Mutual hang-again is the first behavior signal worth logging

    A 26-minute AI interview is a ceiling, not a start: a new member should get a first suggestion within 3 minutes and at most one question a day. The one behavior signal to log first is mutual hang-again, both people saying they would meet again.

    ConfirmedInterpretationWave DChecked 2026-09-28

Ask about this, or request more analysis

Your message goes to Vadym Chernega, founder of Digi117, not a bot. Reply within one business day.

We use your email only to answer. No list, no newsletter. Want this kind of work on your own product? See what we make.