Lecture 12-3: Crisis Situations and Safety Protocols

Proverbs 24:11 — "Rescue those who are being led away to death! Indeed, hold back those who are staggering to the slaughter!" (World English Bible)

Leviticus 19:16 — "You shall not go up and down as a slanderer among your people. You shall not endanger the life of your neighbor. I am Yahweh." (WEB)

Acts 16:28 — "But Paul cried with a loud voice, saying, 'Don't harm yourself, for we are all here!'" (WEB)

3.1 Documented failures, stated soberly

Three cases, each publicly documented and each instructive in a slightly different way, are worth naming here — not to indict the technology wholesale, but to establish that the risk this lecture addresses is real and not speculative.

The Tessa chatbot (2023). The National Eating Disorders Association (NEDA) had operated a human-staffed helpline for people struggling with eating disorders for over twenty years. In 2023, NEDA replaced the helpline with an AI chatbot named Tessa, intended to provide scripted, evidence-based guidance. Shortly after launch, users and advocates reported that Tessa was giving advice dangerous to people with active eating disorders — including specific suggestions about calorie restriction and weight loss, precisely the kind of content that can worsen an eating disorder rather than help it. NEDA took Tessa offline within days of the reports becoming public. This case is instructive because the failure was not a matter of the chatbot being cruel or obviously malfunctioning in an easily detectable way; it produced fluent, superficially reasonable-sounding advice that was clinically dangerous for the specific population it was serving, and it took public reporting rather than internal review to catch it.

AI chatbots and suicide-related queries (RAND Corporation, 2024). The RAND Corporation conducted a study evaluating how three widely used AI chatbots responded to a range of suicide-related questions, categorized by clinical risk level (very-low-risk informational questions, very-high-risk direct requests for lethal means, and — the most concerning finding — a middle band of "intermediate-risk" questions, such as questions about specific means of self-harm framed in a way that did not explicitly announce suicidal intent). The study found that the chatbots tested were generally reliable at declining to answer the highest-risk queries directly and at providing general safety information for low-risk queries, but inconsistent and sometimes inappropriate in the intermediate band — precisely the band where a real, ambiguous, in-progress crisis is most likely to fall, because people in crisis rarely announce their intent as plainly as a study's highest-risk category requires. The RAND researchers' conclusion was not that chatbots are useless for this population, but that inconsistency at exactly the hardest-to-classify risk level is a serious and specific safety gap, not a minor rough edge.

Wrongful-death litigation involving a companion-AI platform (2024). In 2024, a lawsuit was filed against Character.AI (and against Google, which had licensing ties to the company) by the mother of a fourteen-year-old who died by suicide after extended, emotionally intense interactions with a chatbot persona on the platform. The litigation's specific factual claims are contested and, as of this course's writing, unresolved in court, and this lecture is careful not to assert as proven fact what remains a matter of ongoing legal dispute. What is not disputed, and what makes the case relevant here regardless of its ultimate legal outcome, is that it became a widely reported instance of the exact structural concern this chapter has been building toward: a vulnerable, isolated person formed an intense parasocial reliance on a conversational AI system during a period of real crisis, and the system had no mechanism resembling what a trained human — a friend noticing withdrawal, a pastor sensing something wrong, a counselor trained to ask a direct question about self-harm — would have had.

Taken together, these three cases point to the same underlying gap from three different angles: a chatbot can fail silently in the space between "obviously fine" and "obviously an emergency," precisely because that space is where real crises usually sit, and precisely because that space is where genuine human discernment — the kind Hebrews 5:14 will later call the mark of spiritual maturity in chapter 19 of this course — matters most and is hardest to replicate in a system trained to produce plausible continuations of a conversation.

3.2 "Rescue those who are being led away to death"

Proverbs 24:11 is not phrased as advice or as a suggestion for a good life; it is an imperative addressed to a bystander with knowledge of danger: "Rescue those who are being led away to death! Indeed, hold back those who are staggering to the slaughter!" (WEB). The verse assumes a scene in which someone is walking toward destruction and another party — the one being addressed — is in a position to intervene and does not. The following verse (24:12) anticipates and rejects the excuse that will be offered: "if you say, 'Behold, we didn't know this,' doesn't he who weighs the hearts consider it?" The text does not accept ignorance-of-the-full-picture as a defense when the danger was in fact knowable.

Apply this directly to the situation this lecture has described. An AI chatbot conversing with someone in the "intermediate-risk" band the RAND study identified is, in the terms of Proverbs 24:11, positioned exactly where the bystander is positioned — with information suggesting danger and an opportunity to intervene. The command is to rescue and hold back — active, physical-sounding verbs describing an intervention, not merely the provision of accurate information about the danger. A chatbot that responds to an ambiguous crisis message with an accurate, well-footnoted explanation of suicide risk factors has not obeyed Proverbs 24:11's imperative; it has described the valley rather than pulling someone out of it, which is precisely the distinction lecture 12-1 drew with Psalm 23:4. Proverbs 24:11 does not ask for information delivered faithfully. It asks for rescue performed actively, and rescue is not a category a text-generation system, however accurately it can describe danger, is equipped to perform.

3.3 Not standing idly by: Leviticus 19:16 and Acts 16:28 together

Leviticus 19:16 belongs to the same section of the Holiness Code this course engaged in chapter 6 (Leviticus 19:15, honest judgment) and reads, in the World English Bible's rendering: "You shall not go up and down as a slanderer among your people. You shall not endanger the life of your neighbor. I am Yahweh." The verse pairs two prohibitions that might seem unrelated — gossip and endangering a neighbor's life — but the pairing is deliberate: both are failures of active care for the community one lives among, one through destructive speech and one through passive or careless disregard for a neighbor's safety. The traditional rabbinic reading of this verse (reflected in the more familiar, if less literal, older English rendering "neither shalt thou stand against the blood of thy neighbour") has long understood it as forbidding not only actively endangering a neighbor, but also standing by while a neighbor's life is in danger and failing to act when action is possible. Applied here: a church that knows a member in crisis has been relying on an AI chatbot as their primary or sole outlet, and does nothing further because "at least they're talking to something," is coming uncomfortably close to the passive failure this text names.

Acts 16:28 supplies the positive counter-image, in a scene worth recalling in full. Paul and Silas, imprisoned in Philippi, are freed by an earthquake in the night; the jailer, waking to find the prison doors open and assuming his prisoners have escaped (a failure for which Roman law would have held him personally liable, likely with his life), draws his sword to kill himself. Paul does not deliver an accurate assessment of the jailer's legal situation, however true such an assessment might have been. He "cried with a loud voice, saying, 'Don't harm yourself, for we are all here!'" (Acts 16:28, WEB) — an immediate, urgent, unambiguous verbal intervention, delivered the instant the danger became apparent, aimed entirely at stopping the specific act about to occur. The text then moves directly to Paul physically going to the jailer, calling for lights, and going in to him (16:29-30) — verbal urgency followed immediately by embodied presence. This is precisely the two-part response Proverbs 24:11 commands and Leviticus 19:16 protects against the failure to give: immediate, direct address to the danger, followed by actual physical presence with the person in crisis. No chatbot, in any of the documented cases this lecture named, was capable of the second half of that pattern, and the first case (Tessa) shows that even the first half — an immediate, correctly calibrated verbal intervention — cannot be assumed reliable either.

3.4 A concrete protocol

This course has resisted, throughout, the temptation to end an ethical argument with only an abstract conclusion, and this lecture is a case where the practical stakes make that discipline especially important. A congregation's AI-and-crisis protocol, drawing directly on the argument above, should include at minimum:

1. A standing, communicated policy that no AI tool — chatbot, screening system, or otherwise — is ever the sole or final line of response for a person in acute crisis. This is not a policy against AI tools existing in a congregation's life; it is a policy against them being trusted, by default or by accident, as sufficient. Given the Tessa case's lesson that failure can look like fluent, reasonable-sounding advice, this policy needs to be explicit rather than assumed as obvious.
2. A named, trained human point of contact — not a general "someone will get back to you" — reachable outside office hours for anyone the church knows to be in crisis, closing the specific "availability gap" that lecture 12-1 identified as a genuine reason people turn to a chatbot in the first place. If the honest reason someone chooses a chatbot at 3 a.m. is that no human alternative exists, the correction is not merely warning people away from chatbots; it is building the human alternative that currently does not exist.
3. Clear, simple guidance — taught, not merely posted — for any staff member, volunteer, or small-group leader who learns that a congregant has disclosed suicidal ideation, self-harm, or abuse to an AI chatbot, or to anyone, modeled directly on Acts 16:28's pattern: immediate, direct, verbal address to the specific danger, followed without delay by actual physical presence. This should include clear escalation paths to emergency services and professional crisis resources, and should never treat a pastor's or volunteer's good intentions as a substitute for calling on those resources when the situation warrants it.
4. Regular review of the specific technologies the church itself uses — this anticipates chapter 15's fuller treatment of church administration and AI, but belongs here in outline: if a church's own intake forms, prayer-request systems, or counseling scheduling tools route through any AI-mediated triage, someone must be responsible for confirming, and re-confirming as tools change, that no such system is the last line of defense for anyone in genuine danger.

None of this is a claim that AI tools have no place in a congregation's life, and none of it depends on the technology remaining as limited as it is today. Proverbs 24:11's command to rescue, Leviticus 19:16's prohibition on standing idly by, and Acts 16:28's model of immediate human presence do not become obsolete if a future chatbot passes every safety benchmark RAND or any other body devises. They describe an obligation the church bears toward its own members regardless of how good the tools available to those members become, because the obligation was never about the tools; it was about not leaving a neighbor to face the valley of the shadow of death alone.