← back to NothingBox
AI · Transparency
How AI works
in NothingBox.
NB uses AI in three specific places. Here's exactly what it does in each one, what it never does anywhere, and what happens when AI isn't available.
Where AI is used
Exercise matching
Triggered when you select a mood
20% of sessions
When you select a mood, NB occasionally uses Claude Haiku to select the exercise most relevant to your specific combination of mood, time of day, and session context. 80% of the time, rule-based matching is used instead — no AI involved.
✓ Selects from a fixed set of pre-written exercises
✓ Uses only your mood selection and time of day
✗ Does not generate new content
✗ Does not read anything you've written
Name It reflection
Triggered after the 4-question Name It flow
On request only
After you answer 4 questions about how you're feeling, Claude Haiku generates one short sentence that witnesses what you described. Nothing more.
✓ Generates exactly one sentence
✓ Always starts with "It sounds like" or "It feels like"
✓ Under 25 words
✗ Never gives advice
✗ Never asks questions
✗ Never offers a silver lining or solution
Permission statements
In the Pressure Permission exercise
Occasionally
For exercises focused on expectation and pressure, Claude Haiku occasionally personalises the permission statement shown to you — based only on what you wrote in the exercise prompt.
✓ Selects from a curated pool of permission statements
✓ Warm, not clinical
✗ Does not store what you wrote
✗ Does not reference previous sessions
The rules AI follows
Every AI call in NothingBox follows these rules
◎ Witness only. AI describes what it hears — it never tells you what to do, what to feel, or how to interpret your experience.
○ No diagnosis. NB's AI does not identify conditions, disorders, or clinical states. "Blank" is not depression. "Scattered" is not ADHD. We don't go there.
◌ No memory between sessions. Each AI call starts fresh. The AI has no knowledge of your previous visits, moods, or responses.
〰 Timeout of 1 second. If AI takes longer than 1 second to respond, a static fallback is shown. You'll never know the difference.
✦ Under 25 words per response. The AI is constrained to short, warm, specific output. It cannot write an essay. It cannot spiral. It says one thing and stops.
What AI never does in NothingBox
✗
Diagnoses you. NB's AI does not identify conditions, suggest you have a disorder, or make clinical assessments of any kind.
✗
Gives advice. "You should..." does not exist in NB's AI output. Ever. Not even helpful advice. The role of AI here is to witness, not to direct.
✗
Reads your burn letters. Burn letter content never reaches our server. AI never sees it.
✗
Analyses your camera. The mirror feature uses your camera as a mirror only. No frame is processed by AI or any computer vision system.
✗
Tracks you across sessions. AI in NB has no memory. Each session is completely isolated. It cannot learn your patterns or build a profile of you.
✗
Decides what you should do next. AI suggests features based on mood. It does not make decisions for you or steer you toward specific outcomes.
When AI isn't available
Every AI feature in NothingBox has a static fallback — a carefully written response that appears if AI is unavailable, too slow, or returns something unexpected. NB works completely without AI. The fallbacks are not degraded versions — they're written with the same care as the AI-generated responses.
"It sounds like you're carrying something that hasn't found its shape yet. That's okay. It's still real."
Example: the static fallback for the Name It reflection. This appears whenever AI is unavailable.
The AI model
NothingBox uses Claude Haiku by Anthropic — a small, fast model designed for focused tasks. We chose it because it follows constrained instructions reliably, operates within a 1-second window, and is made by a company (Anthropic) with a stated commitment to AI safety.

Anthropic processes your inputs per their own privacy policy. They do not retain your data from API calls.

NothingBox does not use GPT, Gemini, or any other AI model. One model. Constrained. Transparent.
← back to NothingBox