Crisis Text Line & Loris.ai

Crisis Text Line, a US nonprofit, saves every crisis text conversation in a corpus, a store of texts that trains its own triage model. From about 2017 to 2020, Loris.ai, a for-profit company it partly owned, used that corpus to train commercial customer-service software.

The service

Crisis Text Line has offered free text-based crisis support around the clock since 2013. It also operates through affiliates in Canada, the UK, and Ireland.

As of early 2025 it reported nearly 10 million conversations and about 300 million messages since 2013. It reported more than 71,000 trained volunteers. It says it supported more than 1.5 million conversations in 2025, its highest-volume year.

The triage model

Crisis Text Line's data-science team built a machine-learning model to direct scarce counselor attention under heavy volume. The service describes it as several deep neural networks, a kind of machine-learning model, whose scores are combined into one. It was trained on about 65 million messages. It scores the opening messages of a conversation for severity. It then reorders the counselor queue by predicted risk, instead of by arrival time.

Crisis Text Line says the model identifies 86 percent of people at severe imminent risk in their first conversations. It says it aims to serve 94 percent of high-risk texters in under five minutes. These are self-reported blog and marketing figures, never independently evaluated.

The model decides who waits. It does not reply to texters or send emergency help. Volunteer Crisis Counselors tag conversations, for example "Suicide", and rate their risk in surveys afterward. Those labels are used to retrain the model.

Crisis Text Line still uses machine learning to rank risk. The sources do not say whether the 2018 model is still in use unchanged. Its error rates, and how often staff set it aside, have never been published.

Who decides about a rescue

When a texter may be in imminent danger, a paid clinical supervisor can authorize an active rescue. The supervisor contacts a local 911 center with the texter's phone number and carrier.

A 2020 Crisis Text Line self-report put active rescues at about 0.82 percent of conversations, roughly 28 a day. About 40 percent of conversations with imminent risk ended in a rescue. About 60 percent were de-escalated, meaning calmed, without one. The sources do not clearly document how many rescues were voluntary and how many were not.

The arrangement with Loris.ai

In November 2017, Crisis Text Line incorporated Loris.ai, a for-profit spinoff. Loris.ai was co-founded by Nancy Lublin, a co-founder of Crisis Text Line. Crisis Text Line held an ownership stake, which reporting based on Politico put at about 53 percent. A revenue-sharing agreement came with it. Per that reporting, the agreement had not taken effect as of January 2022.

Loris.ai used the anonymized corpus to train commercial customer-service and de-escalation software. Its website cited a "sentiment-rich" corpus of "62 million messages" from Crisis Text Line. That is a marketing figure, not an audited count. The exact number of records shared has never been made public.

How consent was gathered

Texters were asked for consent while in crisis. An automated reply pointed texters to a lengthy Terms of Service. The Markkula Center for Applied Ethics described it as a 50-page agreement. The Reform Crisis Text Line campaign described it as over 4,000 words. Critics noted that many texters are minors.

A former board chair, danah boyd, voted for the arrangement. She later said, "knowing what I know now, I would not have." Tim Reierson, a volunteer Crisis Text Line had terminated, argued publicly that a Terms of Service accepted mid-crisis cannot be meaningful informed consent. The case file does not say why he was terminated.

Who was asked

An outside data-ethics committee of academics and technologists reviewed research proposals. It was reportedly not brought into the decision to share the data with Loris.ai. As the case file puts it, a check that is not invoked is not a check.

How it ended

Politico reported the arrangement on January 28, 2022, and public backlash followed quickly. On January 31, 2022, Crisis Text Line announced it had ended the data-sharing and asked Loris.ai to delete the data. It stated the shared data had been anonymized and scrubbed of personally identifiable information, meaning details that identify a person. It said no such information was sold, and Loris.ai had not accessed the data since the start of 2020.

In March 2022, Brendan Carr, a commissioner of the Federal Communications Commission, wrote to Lina Khan, who chaired the Federal Trade Commission. He urged it to investigate how Crisis Text Line collects, keeps, and shares crisis-conversation data, and how it obtains consent. Crisis Text Line said it had engaged "in good faith" and that its practices comply with the law. No public enforcement action by the Federal Trade Commission is documented.

In 2025, chief executive Dena Trujillo replied to a commentary in the Journal of Medical Internet Research. She said texters must consent to the privacy policy to use the service, and can ask for deletion by texting DELETE. She said Loris.ai was required to return or delete data it had accessed. She said an Institutional Review Board, a committee that approves research involving people, has overseen in-house research since 2023.

Why this case is different

Most cases in the Lab turn on a model that is wrong, or on people who fail to catch it. This case is neither. The triage's accuracy is unaudited, but the harm here does not come from accuracy. The human review around the model is one of the better ones among the Lab's cases.

Three things make it distinct. First, a conflict of loyalty: the service had a stake in the company reusing its data. The chance of earning money from reuse settles the question without anyone deciding, unless governance is built to say no. Second, consent that cannot really be given: a Terms of Service is a legal document, not informed consent from someone texting a crisis line. Third, a review body that existed but was bypassed.

The case file also draws a lesson about deletion after the fact. You can ask for a copy of data back, but you cannot undo training a partner has already done on it.

What this network is drawn from

This network follows the pattern the case file documents. It does not reconstruct the actual service. It shows the triage model, the counselor queue, the counselors, the clinical supervisors, the corpus, the data-ethics committee and board, Loris.ai, and 911 services. The texters are outside the network, and no suicide or crisis outcome is computed.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Here a mistake is, for example, a wrong severity score or a wrong risk label. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 11 units. Understand the system, a Lab option that lowers some tools' prices, is not on offer here. So every price below is the full price, the same at every target level. Explore (No Targets) sets no targets.

Lingering effects are effects that stay after their cause is gone. The Lab's Dynamics menu starts at Both, which includes them, under Explore (No Targets) and Service Targets Only, where you can change it. Service and Safety Targets fixes it at Side-effects, which leaves them out. Side-effects are costs that governance itself adds, such as staff leaning on an oversight layer. All Governance Targets fixes it at Both, which includes both side-effects and lingering effects. With lingering effects on, three tools work at reduced strength here: Peer sharing rules, Vet connections, and Store less data.

Under Service Targets Only, the targets are not met before any tool is used. Mistakes are copied about as fast as they are corrected. That level asks for them to be caught faster than they are copied, and for the model to be helping the work. At the starting setting, the cheapest ways cost 3 units: Peer sharing rules or Assign a challenger, each at its stronger setting. With lingering effects off, Peer sharing rules meets them on its own for 2 units. In all, nearly 500 different sets of tools within the budget meet them.

Under Service and Safety Targets, the targets can be met, but only by spending the whole budget. This level also asks you to close every failure pathway, among other targets. Five are open before any tool is used. They are Model reorders the queue, Past conversations train the model, Counselor labels retrain the model, Every conversation saved to the corpus, and Counselor escalates to a supervisor.

Escalate checks is the only tool on offer that closes Model reorders the queue. Mark AI-written records closes Past conversations train the model. Vet connections closes it too, but costs more and fits in no set within the budget. Keep prompts neutral is the only tool that closes Counselor labels retrain the model. In the Lab it keeps questions put to an AI neutral. Here it acts on the counselors' labels, and the sources describe no such practice. Gate record entries, or Store less data, closes Every conversation saved to the corpus. Peer sharing rules, or Assign a challenger, closes Counselor escalates to a supervisor.

So exactly four sets of tools meet the targets at this level, and each costs all 11 units. Each has Escalate checks, Mark AI-written records, and Keep prompts neutral. Each adds Gate record entries or Store less data, and Peer sharing rules or Assign a challenger.

Under All Governance Targets, lingering effects are on. Store less data then leaves Every conversation saved to the corpus open. Its stronger setting closes it but costs too much to fit beside the other tools needed. So exactly two sets meet the targets, each costing all 11 units. They use Gate record entries, with Peer sharing rules or Assign a challenger.

None of these sets acts on Corpus shared with Loris.ai, the link this case turns on. That link is not a failure pathway unless you add the pressure named Connectors sprawl. The targets also do not ask for either check to start. Peer sharing rules and Assign a challenger start the Data-ethics committee review. No tool on offer starts the Consent check on the corpus.

More is not better here. Using every tool at its strongest setting costs 34 units, about three times the budget. It contains the mistakes and closes every failure pathway. It still meets the targets at no level, because the model then adds too little to the work.