Hard Truth: The Central Claim
Humma calls Empathetic AI an open-source standard "written in the open, owned by no one." There is no public code, no license, no benchmark, and no paper anywhere. What does exist: three trademark applications by Humma.AI, Inc., one of which claims the open-source software itself as proprietary goods.
Three more claims break the same way.
HUM
Outside-In Audit / Referral via Limore Shur
The brand describes a company that does not exist yet.
Humma.AI sells consent-first AI for healthcare. Around it are a mutual aid community and a story told by Marvel-credited artists. We scored every claim it makes against what a stranger can verify. The marketing is genuinely good. Most of what it describes has not been built yet.
Humma.AI, Inc. · Los Angeles · Audit date 2026-08-14 · Prepared by SHUR Creative Partners
14/50
Structural Brand Power Index
570
On the waitlist counter
01
Humma oversells in four specific places: the open-source standard, community governance, the no-tracking promise, and the claim that healthcare success guarantees every other sector. In two others, it publishes evidence against its own pitch.
The open-source standard doesn't exist in public. The marketing describes community governance in the present tense. Humma's own FAQ contradicts it. The homepage says "not in a log, not in a profile" while running a Google Ads tag and an analytics beacon on every page view. And "because Humma AI works in healthcare, it will work in every sector" is written about a product no hospital has run yet.
Humma also publishes facts that hurt it. Humma's benchmarks page reports the one axis where its model is clearly behind Claude Sonnet 4.5, in its own words "the one axis we have not closed." Its FAQ volunteers that the governance charter is "approved, not yet in force."
The marketing speaks in absolutes about a system the operating documents describe in modes, records, and exceptions. Humma asserts today what it plans to build. The open seed round is selling that assertion. The round has been publicly solicited since January 8, 2026 without an announced close.
02
What Humma got right, stated as strongly as the evidence allows.
Public resentment of extractive AI is real. Humma saw it earlier than most companies did. The governance board is genuinely senior: Kay Firth-Butterfield chaired AI at the World Economic Forum, Desmond Upton Patton runs SAFElab at Penn, and Zachary Elewitz leads McKesson's Enterprise AI Lab.
The Last Trace is a 40-page interactive story by artists with Marvel and DC credits, and it works as brand mythology. The privacy notice is better written than most. It admits plainly that withdrawing consent can't reverse a completed training run. The CTO is a licensed physician who has built clinical language-processing systems. Inviting the first beta members to attack the system is credible engineering culture.
If Humma ships the product, this brand work will still have been worth what it cost.
Hard Truth: Registered vs Filed
Humma's press release calls Empathetic AI "a registered trademark." Its own annual report says only that applications were filed. The company's most legally precise claim about Empathetic AI is false.
All three trademark applications are still pending.
03
Thirty claims, each scored Substantiated, Partially substantiated, or Unsubstantiated against public evidence. The fourteen that decide the verdict:
| Claim, as the brand states it | Verdict | What the evidence says |
|---|
| "Empathetic AI™, our open-source methodology and benchmarking standard" | Unsubstantiated | Two GitHub accounts reserved in 2024, both empty. Nothing in any of the public code, model, or research repositories where such a standard would be published, and nothing at any standards body. Two years of the ™ symbol, zero artifacts. |
| "Written in the open, owned by no one" | Contradicted | Three US Patent and Trademark Office applications, 98303763, 98743370, and 99833454, filed by Humma.AI, Inc. The last one claims "downloadable open-source computer software" as trademarked goods. |
| "Governed by community" | Unsubstantiated today | By Humma's own FAQ: charter amendments approved but not filed, prior bylaws in force, Community Advisory Board "not yet in force." Marketing states the binding veto in the present tense. |
| "Not in a log. Not in a profile. Not for sale." | Partially | "Not for sale" holds as a formal representation. The rest doesn't: the homepage runs a Google Ads tag and an analytics beacon capturing browser, locale, country, referrer, and full URL on every page view. |
| "There's no targeting algorithm because there's nothing to target with" | Unsubstantiated | True only in the narrow sense that Humma doesn't sell ads. The matching service needs identities on a server to work at all. |
| "Free forever, no ads, no upsells, no subscriptions" | Partially | Consistent across every page. But "free by charter" depends on a charter that isn't filed. Today it's a pricing policy. |
| "5% of revenue goes back to the communities, and they vote on where" | Substantiated, inactive | Stated everywhere including the terms. The annual report is candid: pre-revenue, mechanism not activated, advisory board in development. Five percent of zero. |
| "Clinically curated, rule-checked, deterministic on domain expertise" | Partially | The architecture is described in unusual detail and matches the CTO's prior work at IQVIA. No pilot, customer, or third party has verified it in use. |
| 81.8% human preference over ChatGPT in a UC Irvine study | Unsubstantiated | No paper, no sample size, no methodology anywhere. The company's own published benchmark measured safety and task discipline, so nothing Humma has released supports or refutes the preference figure. |
| "Humma cracked the code" on patient-centered AI | Unsubstantiated | A capability claim from a company whose clinical advisory board does not exist yet. |
| California Benefit Corporation | Partially | Self-asserted with the right statute cited, and behaviorally consistent: two annual benefit reports exist. The registry itself blocks automated search. Other pages call the company a "Public Benefit Corporation," which is Delaware's term. |
| "570 people are exploring" | Substantiated, small | The 2025 annual report, signed February 9, 2026, said "over 500." The backend reported 630 on August 14, 2026. Six months bought roughly 130 signups. |
| Team of "technologists, ethicists, physicians, social workers" | Partially | One physician credential on the page, the CTO. "Dr. Brenda Fiala" shows no doctorate on her own LinkedIn. The co-founder's indexed LinkedIn lists a different current job. |
| "Our financial model has been vetted by experts" | Unsubstantiated | Nothing filed with the Securities and Exchange Commission, no close announced, no institutional investor named anywhere. Tracxn lists the company as unfunded. The one named backer is an angel who is also an advisor. |
04
Five contradictions, all resolved against the brand.
Confirmed"Nothing in a log" meets the training opt-in
Consent-gated collection is still collection, and the privacy notice says so in a section titled "What We Collect and Why." That notice also lists the vendors receiving it: cloud hosting, AI model inference, web analytics, and a customer-relationship database.
Confirmed"Free forever" meets how Humma makes money
Both revenue lines bill health systems: a paid readiness assessment and Concierge licensing. No community member is ever billed, which is how "free forever" and a revenue plan hold together.
DocumentedThe counter
Displayed: 570. Backend: 630, cached and stale. The hard-coded fallback of 500 means the only traction number on the site can never look smaller than the 2025 annual report's.
Confirmed"Nothing to target with" meets the matching service
To match people, the system stores needs, offers, locations, and identities server-side on Supabase. Encryption at rest with separate keys is claimed; end-to-end encryption is not, anywhere. Redaction of personal information is "best-effort" by the notice's own parenthesis.
The ExceptionHumma publishes evidence against itself
The benchmarks page names the one axis its model has not closed. The FAQ flags the charter as not in force. In both cases Humma's own documents are more precise than its marketing.
Hard Truth: The Missing API
Humma claims revenue comes from API licensing. There is no API documentation, no developer portal, and humma.ai/api returns a not-found error. The privacy notice names no vendor, so the model behind the product can't be identified from anything public.
What Humma does document in detail is the Concierge's design.
05
The design is coherent, and the CTO has built this kind of system before.
The hospital product, as documented on its own page: a governed 32-billion-parameter open-weight model that a customer can self-host. The model does no clinical reasoning. Doctors write that reasoning into cited articles based on published guidance. A governance layer checks every reply before it reaches a patient: no diagnosis, no prescribing, no urgency verdicts, nothing the sources don't support. Replies that fail are rewritten or replaced.
The CTO built multi-agent systems and clinical language-processing pipelines at IQVIA and practiced as a physician in Britain's National Health Service. The design he describes from that work, models and tools with compliance checks around them, is the design Humma sells. The claim "we build on small, governed, open-weight models rather than calling out to a frontier model" is consistent across the FAQ, the mission page, and the benchmarks page.
Nothing public shows it running: no named pilot, no customer, no subprocessor list, no job posting naming the technology, no API a developer could open. The clinical advisory board the marketing cites as oversight is, per the product page itself, "the next step as we move into pilots" (their words).
06
Humma's benchmarks page reports where its model wins and where it is behind. Its robots.txt asks crawlers to skip that page.
The company benchmarked its Medication Assistant against Claude Sonnet 4.5 given Humma's best structured prompt, with a printed leaflet as the third comparison, standing in for current practice. Its model won on safety: violating replies in 2.5% of its answers against Claude's 15.4%, at roughly one eighth the cost per encounter. It is clearly behind on comprehension verification, 0.39 against 0.83, which the page calls "the one axis we have not closed." On addressing the patient's concern it trails, 0.56 against 0.80, a result the page's own table marks inconclusive.
The tradeoff is deliberate: fewer unsafe replies, weaker checking of whether the patient understood. One click away, the homepage claims "sensitivity, nuance, and cultural awareness" and cites the unverifiable 81.8% study. The benchmark neither supports nor contradicts those claims: it measured safety and task discipline, not warmth.
07
Some of the consent claims are built into the data model. Others stop at the marketing site.
Training use requires separate recorded opt-in, off by default. Consent records are append-only at the database level, so no code change can quietly rewrite them. Deletion is hard deletion with three named survivals. A No History mode deletes conversations after about an hour. These are architecture decisions, and they go further than most companies do.
The rest stops at the marketing site. The same homepage runs Google Ads conversion tracking and a Tinybird beacon logging every visit. The Last Trace's waitlist form promises "no spam, ever" while posting to HubSpot with the visitor's tracking cookie attached. The privacy notice's "What We Collect and Why" is one generic paragraph that cross-references a section that doesn't exist. The terms cap total liability at $100 and waive class actions.
08
The marketing and the operating documents each say things the other never mentions. What only the marketing says is the overselling.
The marketing repeats claims the operating documents never back. "Owned by no one" runs against logged trademark and patent activity. "Communities control their own data" runs against a notice offering individual privacy modes only. "Free by charter" runs against an unfiled charter. The 5% community vote depends on instruments the documents call not yet in force. Commitments the operating documents make and the marketing never mentions: append-only consent records, hard deletion, age assurance without ID, two named revenue lines, the $100 liability cap, and the class-action waiver.
Consent is the only concept both the marketing and the operating documents treat as central. Everything else one of them treats as central, the other barely mentions.
Explore the graphs directly: the Brand Graph, the Operational Graph, and the Differential.
Hard Truth: No One Else Does Both
Humma's differentiation is in consent-first community platforms, which have almost no revenue. Its revenue plan is in healthcare AI, where it has no pilot. No other company works in both.
Humma ranks last on both lists.
09
Structural Brand Power Index: five dimensions, public evidence only.
Content strength7.0
Narrative ownership3.0
Community strength2.0
Distribution power1.0
Monetization1.0
Two years in, Humma has world-class brand content, almost no audience, and no way to earn money yet. It is raising money on the strength of the brand. The full profile is in the SBPI viewport.
10
Consent-first, community-owned social and mutual aid. This is the group Humma resembles. None of these companies sells to hospitals.
| Rank | Company | SBPI | Placement |
|---|
| 1 | Bluesky | 36 | 46M registered users and $123M raised. It owns the open-social-web story in the press. Engagement fell 27% year over year. Monetization starts in 2026. |
| 2 | Signal | 34 | The consent benchmark at 70-100M monthly users. Even its first paid product is built so Signal can't see the data. It also lost $8.6M in 2024: consent-first still needs a payer. |
| 3 | Mighty Networks | 37 | $500M in host earnings in 2025. On money alone it leads this list. It ranks below Signal here because hosts own member data, which is the opposite of member sovereignty. |
| 4 | Open Collective | 34 | $118M moved through 3,785 collectives, community-owned since 2024. It handles money for mutual aid groups rather than running a social network. |
| 5 | Front Porch Forum | 34 | 93% of Vermont households, human-moderated, profitable for 20 years. Proof the model works, at deliberately small geography, with zero AI. |
| 6 | Circle | 33 | Commercial community software, sold by subscription, that mines patterns across communities. The clearest company to contrast Humma with. |
| 7 | Discourse | 33 | 22,000 communities and revenue that stalled around $10-13M, which forced a profit-first change of plan at the end of 2025. Its corporate policy admits exchanging data with brokers. |
| 8 | Buy Nothing | 32 | 14M members and the closest live mutual-aid analogue. Every attempt to monetize, in 2021, 2023, and 2025, triggered community revolt. The cautionary case for Humma's model. |
| 9 | Mastodon | 27 | 9.1M accounts, 794K monthly users and falling. |
| 10 | Hylo | 22 | The closest match to Humma's stated values: mutual aid, no ads, open source since 2012, grant-dependent, hundreds of groups. What winning this niche looks like while staying small. |
| 11 | Humma | 14 | Last on every dimension except content, where it beats most of this list. |
11
Healthcare AI concierge and patient navigation. Every dollar in Humma's revenue plan comes from this group of buyers.
| Rank | Company | SBPI | Placement |
|---|
| 1 | Commure (Engage) | 33 | Absorbed Memora Health on December 20, 2024; $7B valuation, 500+ organizations, 200M patient encounters a year. The largest company in this market. |
| 2 | Luma Health | 33 | Claims 1,400+ health systems and a $100M revenue plan; a third-party estimate reads $32.4M trailing. Either way, orders of magnitude ahead. |
| 3 | Fabric | 31 | 100M covered lives assembled through five acquisitions in three years, plus its own 50-state clinical network. |
| 4 | Notable | 28 | 12,000+ care sites automated inside the hospital medical-record system. Declines to publish accuracy rates, and the market buys anyway. |
| 5 | Hyro | 29 | $95M raised, 45+ health systems, 30M patients reached. Edges Notable on composite; ranked fifth on deployment footprint. |
| · | Humma | 14 | Pre-pilot. It has not yet entered the market it has to win. |
Every incumbent on this list treats patient consent as provider-delegated paperwork: Hyro delegates it to the health system, Luma sells consent capture as a provider feature, Commure automates consent as an intake step. If patient-owned consent ever becomes a purchasing requirement, Humma's architecture becomes a genuine advantage. The Medicare and Medicaid health-equity rules Humma itself cites are the most likely reason that would happen. Today, no hospital purchasing document asks for it. Both rankings are interactive in the Rankings viewport.
12
Humma condemns the two-sided platform model. Its own business is two-sided: a free community whose opt-in contributions train the product it sells to hospitals. Consent is the stated difference, and its first pilot will show whether that difference is real.
What to watch, in order: a signed hospital pilot with a name attached; the charter filing that makes "free by charter" legally true; the first file to appear in either empty GitHub organization; and the seed close. Each one would settle a claim that public evidence cannot settle today. None of them has happened as of August 14, 2026.
Five interactive views of the evidence. Each one carries the same navigation, forward and back and home to this page.