Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Crisis Text Line & Loris.ai

The corpus and the spinoff: governing crisis-conversation data

A crisis line's in-house model reorders which texter a counselor sees first, and the human loop around it genuinely works. The exposure is somewhere else: the same conversations become a corpus, and that corpus was routed to a for-profit the nonprofit owned a stake in, to train commercial software - a decision the ethics committee was reportedly never asked about. Modeled on Crisis Text Line and the Loris.ai controversy. The lever here is not the model; it is who may connect to the most sensitive data, and whether the people in it ever meaningfully agreed.

Stylized model of a documented deploymentBehavioral-health & crisis triage

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Crisis-Text-Line-class crisis-corpus governance network: 8 components and 12 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 5 assumed · 3 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the crisis-conversation data-provenance and consent-governance pattern documented in the Crisis Text Line / Loris.ai case file - not a reconstruction of the actual service, its ranker, or the exact terms of the arrangement. The scope here is deliberately the data layer, not the clinical accuracy of the triage.

  • baseline

    This shape's defining feature is an absence: two check pathways are drawn empty at baseline. The committee check is empty because the external data-ethics committee reviewed research proposals but was reportedly not looped in on the decision to route the corpus to the for-profit, so no standing review gated the commercial sharing - the relationship was ended in three days only after an outside expose. The consent/provenance check is empty because no gate verified the corpus records carried meaningful informed consent before retention or onward routing.

  • baseline

    The data-sharing pathway out of the org is the documented Loris.ai corpus-reuse pathway: the same anonymized conversation corpus was used to train commercial customer-service machine learning at a for-profit the nonprofit financially owned (an ownership stake reported at about 53 percent), open roughly 2017 to 2020 and terminated in January 2022. It is drawn low-but-real, not empty, because it demonstrably happened. The 62-million-message figure is Loris's own marketing rather than an audited count, the exact number of records shared is not public, and Crisis Text Line (CTL) disputes the framing (anonymized, no personally identifiable information sold, not accessed after early 2020). The Lab shows the pathway out, never the harm at the far end.

  • baseline

    Consent is drawn as an exposure on the conversation-capture edge: every conversation is captured into the corpus, gathered via an automated reply pointing texters to a long Terms of Service accepted at the moment of acute crisis (described as a 50-page agreement, or a 4,000-plus word document), by texters who include minors. Critics including a former board chair, who voted for the arrangement and later said she would not have knowing what she knows now, argued that a Terms of Service is not meaningful informed consent for people in crisis.

  • assumed

    Unlike most tools in this Atlas, the human review here is drawn active, not latent: the ranker only reorders the queue, counselors keep full within-conversation discretion, and supervisors authorize active rescues at their own judgement. The counselor-to-supervisor handoff is a genuine, staffed loop. The case's ungoverned surfaces sit around that working loop - in the data flow - not inside it.

  • assumed

    The active-rescue edge is the second consent exposure: a supervisor can authorize an emergency dispatch using the texter's phone number and carrier, which a 2020 CTL self-report put at about 0.82 percent of conversations, sometimes without the texter's real-time consent. The voluntary-versus-involuntary split of those dispatches is not clearly documented, so no rate beyond the 2020 self-report is asserted. A rescue here is an institutional signal, never a life.

  • assumed

    One national ranker orders every queue, so the model-to-model self-loop encodes correlated allocation: the same severity model's tilt applies to every conversation at once, with no second reading anywhere to disagree with it.

  • assumed

    No suicide or crisis outcome is modeled here. This Lab models institutional propagation only, and served texters are not in the dynamics. CTL's self-reported triage metrics (86 percent detection, 94 percent of high-risk served in under 5 minutes, a roughly 65-million-message training set) are unaudited marketing claims and are not asserted; the ranker's error and override rates have never been published; and the nearest peer-reviewed crisis-text error benchmarks come from SafeUT, a separate service, and are not attributed to CTL.

What this example does not show

  • This Lab models institutional propagation only. It never models suicide, crisis, or any clinical outcome, and the texters this service helps are not in the diagram - a conversation, a severity score, or an active rescue here is an institutional signal, never a person in crisis. The service's clinical value and any harm to the people it serves are documented in the case file and measured outside any diagram like this one.
  • The corpus-reuse figures are hedged as the sources hedge them. The '62 million messages' is Loris's own marketing language, not an audited count; the exact number of records shared with Loris is not public; the roughly 53 percent ownership stake and revenue-share trace to Politico-derived reporting of CTL's financials, carried here as reported rather than as an audited primary figure; and CTL disputes the framing, stating the data was anonymized, no personally identifiable information was sold, and Loris had not accessed data since early 2020. The FTC referral came from an FCC commissioner, and no public FTC enforcement action is documented - nothing here implies one occurred.
  • This scenario is scoped to data provenance and consent, not to the triage's accuracy. CTL's published detection metrics (86 percent detection of severe imminent risk, 94 percent of high-risk texters served in under 5 minutes, a roughly 65-million-message training set) are self-reported blog and marketing claims, never independently evaluated, and the ranker's false-positive, false-negative, and override rates have never been published. The nearest peer-reviewed crisis-text triage error benchmarks come from SafeUT, a separate Utah-Idaho-Nevada service, and must not be attributed to Crisis Text Line; the active-rescue rates are a 2020 CTL self-report.
  • The dialect finding carried on this cell is in exactly that register. It is peer-reviewed benchmark evidence about crisis-text classifiers as a class, entered with no magnitude and with its primary study absent from this repository's reference snapshot, and it is never attributed to this service's own unpublished false-positive, false-negative or override rates. It is not a fairness measurement, and no gauge on this diagram reads a texter.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • Crisis Text Line, a national nonprofit crisis service, built an in-house machine-learning severity-triage model that reorders which texters volunteer counselors reach first; from about 2017 to 2020 the same anonymized crisis-conversation corpus was routed to Loris.ai, a for-profit spinoff CTL held an ownership stake in — reported by Politico-derived reporting at roughly 53% — which used it to train commercial customer-service software. After a January 28, 2022 Politico exposé, CTL ended the arrangement within three days and requested that the data be deleted; an FCC commissioner referred the matter to the FTC in March 2022, and no public FTC enforcement action is documented. CTL states the shared data was anonymized and never sold as personally identifiable information, and the exact number of records shared has not been made public.

    empirical
    • Reference Crisis Text Line (Wikipedia, tertiary encyclopedia entry) (2026) https://en.wikipedia.org/wiki/Crisis_Text_Line
    • Vendor Crisis Text Line, An Update on Data Privacy, Our Community and Our Service (2022) https://www.crisistextline.org/blog/2022/01/31/an-update-on-data-privacy-our-community-and-our-service/
    • Advocacy Reierson, Reform Crisis Text Line (advocacy site) (2022) https://reformcrisistextline.com/
    • Trade press Benton Institute for Broadband and Society, FCC Commissioner Carr Calls for FTC Probe of Crisis Text Line (2022) https://www.benton.org/headlines/fcc-commissioner-carr-calls-ftc-probe-crisis-text-line

  • Crisis Text Line obtained consent for its data collection through an automated reply directing texters to a lengthy Terms of Service — described as a roughly 50-page or 4,000-plus-word document — accepted at the moment of acute crisis by users who include many minors; critics including a former board chair, who voted for the data-sharing arrangement and later said she would not have "knowing what I know now," and a terminated volunteer argued that a Terms of Service is not meaningful informed consent for people in crisis. CTL says texters must consent to its privacy policy to use the service and can request deletion by texting the word DELETE, and that since 2023 its in-house research has been overseen by an Institutional Review Board.

    empirical
    • Academic Markkula Center for Applied Ethics, Santa Clara University, Crisis Data: An Ethics Case Study (2022) https://www.scu.edu/ethics/focus-areas/internet-ethics/resources/crisis-data-an-ethics-case-study/
    • Academic Eysenbach, Crisis Text Line and Loris.ai Controversy Highlights the Complexity of Informed Consent on the Internet and Data-Sharing Ethics for Machine Learning and Research (Journal of Medical Internet Research, editorial, 2025) https://pmc.ncbi.nlm.nih.gov/articles/PMC11799832/
    • Advocacy Reierson, Reform Crisis Text Line (advocacy site) (2022) https://reformcrisistextline.com/
    • Academic Trujillo, Response From Crisis Text Line to Commentary on Protecting User Privacy and Rights in Academic Data-Sharing Partnerships (Journal of Medical Internet Research, 2025) https://pmc.ncbi.nlm.nih.gov/articles/PMC11799801/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Behavioral-health & crisis triage domain page.

Levers available here and the patterns behind them

Documented case histories