AI-Powered Consent Education Tool for Preventing Gender-Based Violence

override · Other · 2026 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Prevention education and adolescent safety

Override Labs needed a private, judgment-free tool that could help teenagers reflect on ambiguous consent situations without giving permission, shaming users, or enabling harm.

Constraints:

High-risk safety domain involving minors and sexual harm prevention
Users may avoid tools that feel judgmental, tracked, or punitive
Effectiveness is difficult to measure because prevention means an absence of harm

Architecture

Override's Is This Okay? keeps the architecture intentionally small: a free-text consent-scenario UI asks adaptive safety-floor questions, deterministic risk-classification rules evaluate factors such as intoxication, age gaps, and power imbalance, and Claude receives the scenario plus risk tier and motivational-interviewing prompt guidance. Supabase stores modular functions and prompts, while privacy choices avoid accounts, IP logs, and cross-session profiles.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  free-text-scenario-ui["free-text-scenario-ui"]
  safety-floor-questioner["safety-floor-questioner"]
  deterministic-risk-classifier["deterministic-risk-classifier"]
  claude-response-function["claude-response-function"]
  session-context-store["session-context-store"]
  evaluation-rubric["evaluation-rubric"]
  free-text-scenario-ui --> safety-floor-questioner
  safety-floor-questioner --> deterministic-risk-classifier
  deterministic-risk-classifier --> claude-response-function
  deterministic-risk-classifier --> evaluation-rubric
  claude-response-function --> session-context-store

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`free-text-scenario-ui`	Minimal chat-style interface where users describe sensitive consent scenarios and can optionally use guided prompts.	`safety-floor-questioner`
`safety-floor-questioner`	Asks adaptive follow-up questions about intoxication, age gaps, power imbalance, and other risk factors needed for consent assessment.	`free-text-scenario-ui`, `deterministic-risk-classifier`
`deterministic-risk-classifier`	Applies hard rules to produce red/yellow risk classifications before the LLM response is generated.	`safety-floor-questioner`, `claude-response-function`, `evaluation-rubric`
`claude-response-function`	Uses Claude with motivational-interviewing prompt instructions to validate, humanize the other person, and encourage reflection rather than permission-giving.	`deterministic-risk-classifier`, `session-context-store`
`session-context-store`	Maintains coherence during one interaction while avoiding accounts, IP logging, PII, or cross-session profiles.	`claude-response-function`
`evaluation-rubric`	Clinical and AI-advisor rubric for situational assessment, risk assessment, safety severity, and tone against Reddit-derived examples.	`deterministic-risk-classifier`

Data flow

A teen describes a situation in free text or guided mode. The system asks safety-floor follow-ups when the scenario indicates risk factors, runs deterministic classification before the model response, and then sends Claude a prompt containing the situation, risk tier, and nonjudgmental response structure. Follow-up questions use a separate function that preserves the current session context without linking conversations across sessions.

“what I call like safety floor questions, because obviously if any of those types of questions being true, like a power imbalance or intoxication, kind of drives whether or not this actually could be a consenting physical interaction or not.”transcript, implied

“The combination of those factors are really what make up the risk classification function like inside of the tool”transcript, stated

Why this architecture

The design favors deterministic guardrails and privacy over agentic complexity because the product does not need external-world actions, but it does need to avoid the worst case: letting a user justify harm. The LLM is bounded to tone and reflection, while risk classification and permission boundaries are handled by rules and prompts.

“because the model doesn't really need to do anything in the outside world. There's no like application of any more like fashionable AI patterns, like no agents, no MCAP, like nothing like that.”transcript, stated

“using somebody using this tool to either cause harm or justify harm. And so once I started with that is like this is the worst case scenario, I built on top of that how do we avoid that?”transcript, stated

Models used: Claude

Scale

The original charter targeted sexual-assault prevention for high-school-age youth, specifically ages 15 to 19.
“digital tools specifically to prevent sexual assault in the high school age demographic, so like 15 to 19 years old.”transcript, stated
The team scraped the last 2,000 posts in each teenager-focused subreddit it studied, then classified how many were consent- or coercion-related.
“I did a scrape of the last 2,000 posts in each Reddit or in each subreddit, and classified them high-level into how many of these scenarios that people are asking about”transcript, stated
The motivating prevention statistic was that 93% of juvenile sexual-assault scenarios involved victims who knew the perpetrator.
“93% of juvenile sexual assault, in those scenarios, the victims know their perpetrator, right? It's not a stranger, it's somebody who's known to them.”transcript, stated

Alternatives considered (roads not taken)

Agentic tool use or external actions

Rejected because: The system only needs to reflect back guidance; no outside-world tool calls are required for the consent-education use case.

Immediate hard shutdown on degrading language

Rejected because: Coaches warned that punitive shutdown would stop users from opening up, so the product redirects through reflective questions instead.

“The initial version of the prototype, I basically had a red flag list within the system prompt said if if the input has any one of these words that are degrading towards women”transcript, stated

“if you're going to stop them in the first time they're experiencing the product, like they'll not want to open up and continue.”transcript, stated

Meta-style ad tracking and cookie-consent instrumentation

Rejected because: The sensitive product needed trust and minimal tracking, so the team avoided pixel tracking and cookie banners even though that constrained distribution.

“if you if I wanted to implement pixel tracking, I also would have had to implement a cookie consent banner.”transcript, stated

Patterns

`safety-policy-enforcement` (trust-and-safety)

Deterministic risk rules run before the LLM and prevent the system from ever issuing a green-light response.

“How do we build in rules heart like hard rules that are hard coded before again that we talked actually we actually invoke the AI at all.”transcript, stated

`guardrails-io-filtering` (response-control)

The chatbot classifies narratives into risk tiers and uses that classification to shape Claude's response.

“the risk classification really started as pretty simple like red flag, yellow flag. Initially there was a green flag, but now critically the system will never respond with any sort of thumbs up”transcript, stated

`eval-driven-development` (evaluation)

Override uses scraped teen forum scenarios, clinical guidance, and a rubric to test risk assessment, severity, tone, and follow-up quality.

“providing evals that run the gamut of like situational assessment, risk assessment, safety kind of severity, like thinking of all of the different ways that could manifest”transcript, stated

`privacy-preserving-processing` (privacy)

The product minimizes identity capture so sensitive questions can be asked without account creation or persistent tracking.

“There's no IP logging, there's no account creation, there's no PII, obviously. And hopefully that that lack of account creation too is a big feature”transcript, stated

Tradeoffs

How much user history to retain.

Chose session-scoped minimal logging over persistent identity and long-term personalization.

Rationale: The privacy posture reduces friction and aligns with the sensitive, situational use case.

“we aren't tracking session over session. Like we don't have that same burden that the classic LLM experience the classic LLM experience is I need to know you forever”transcript, stated

How to communicate risk to teenage users.

Chose motivational, reflective guidance over punitive scoring or direct verdicts.

Rationale: A reflective tone is more likely to keep users engaged and thinking through the other person's experience.

“you want them to be wise and slightly more experienced, but still know where you're coming from because they were once a young boy, too.”transcript, implied

Gotchas (what bit them)

A conventional permission-style answer would be dangerous in this domain.

Root cause: The worst-case failure is a user using the tool to justify or cause harm.

Resolution: Override removed green flags and keeps even neutral answers focused on consent, communication, and caution.

“the thing you want to avoid at all costs is using somebody using this tool to either cause harm or justify harm.”transcript, stated

Bluntly judging the user can end the conversation before reflection happens.

Root cause: Ambiguous consent scenarios require bringing users along rather than scoring or condemning them.

Resolution: The prompt uses invitational questions and a tone modeled on a trusted older peer.

“it's not just like a reveal of oh, you made a bad choice. It's like prompting in conversation about, "Do you think they felt this way? Or what was the What was your take on this experience? Or what does this person mean to you?"”transcript, stated

Short-term usability tests cannot prove real-world prevention.

Root cause: The desired outcome is behavior change over time, not just an immediate attitude shift in a study.

Resolution: The team treats early studies as directional and says longitudinal measurement is needed.

“this is like really proxy. This is really just directional. It's you have to like we'll have to do like a real longitudinal study on the long term to actually observe if behaviors shifted”transcript, stated

Evaluation

Override combines teen forum research, usability studies with 18-year-old boys, clinical-advisor rubrics, and safety evals, while treating early behavior shifts as directional.

Metrics: risk classification quality · tone quality · directional attitude shift · longitudinal behavior change

“I did a scrape of the last 2,000 posts in each Reddit or in each subreddit, and classified them high-level into how many of these scenarios that people are asking about”transcript, stated

“at least directionally after using the product, people took more of a cautious next step as opposed to a like confident next step”transcript, stated

Outcomes

Early prototypes produced directional shifts toward more cautious next steps.
The privacy-first design supports low-friction use without accounts or persistent profiles.

The project is nonprofit/prevention-oriented; the source emphasizes careful safety and evaluation design rather than production-scale usage numbers.

“So, he helped me design like high level what that study is. Of course, like again with the whole caveat that this is purely directional”transcript, stated