How Path's AI works, and where it falls short

1. Which AI we use

Path uses Claude, made by Anthropic, through Anthropic's service for businesses. We don't train AI models, and we don't use your plans or chats to train any AI model.

• Main model: Claude Opus 5 writes every part of the app by default.
• Smaller models: to keep Path free, some parts may run on Claude Sonnet 5 or Claude Haiku 4.5 instead. [operator: list any part that uses a different model at launch]
• Backup model: if Claude Opus 5 declines a request, Anthropic's service may pass it to another Claude model automatically.
• Web search: the Edge and Map sections may search the web for current information (a few searches per request), through Anthropic's search tool.
• Our own computers (optional): some parts can instead run on an open-weight model, gpt-oss-120b by OpenAI, on computers we run ourselves. Those requests don't go to Anthropic, unless our computers are busy or down, when Claude answers instead. Reading photos, Edge and Map always use Claude. [operator: say which parts, if any, run on our own computers at launch; delete this line if none]

Our server writes the instructions the AI follows. The app only sends your details, never its own instructions, so the AI always gets Path's safety rules.

2. What the AI does in each part of the app

3. It's an AI, not a person

Nobody at Path writes your answers or reads your chats as they happen. To make that clear:

• The app reminds you that you're talking with an AI when you start, and again every hour if you're 13–17 (with a suggestion to take a short break), or every 3 hours if you're 18 or older.
• Chat coaches like this may not be suitable for some people under 18. If you're 13–17, we suggest going through your plan with a parent or guardian.
• The chat doesn't remember past conversations. It only sees your plan and the last few messages of the conversation you're in.
• The AI follows safety rules on every request. It must not recommend anything unsafe, illegal or unhealthy, such as crash diets, skipping sleep, performance-enhancing drugs, dangerous training or faking documents.
• If a chat message suggests you might be in crisis, the app shows help lines (in the United States, such as 988) before the AI replies (see below).
• For people aged 13–17, the AI keeps everything age-appropriate, stays away from dating, sexual content, drugs, alcohol, vaping and gambling, and suggests involving a parent or guardian for decisions about money, travel, meeting people or physical safety.

If a message sounds like a crisis

• What is checked. Every chat message, on our server, before the AI answers. We look for signs of suicide, self-harm, abuse or immediate danger.
• What you see. A short, caring message and help lines you can call or text with one tap. In the United States: the 988 Suicide & Crisis Lifeline, the Crisis Text Line, the Childhelp National Child Abuse Hotline and 911; if your plan is in Oregon, YouthLine too; if you're 13–17, also Teen Line and The Trevor Project. In other countries: that country's own lines. The AI itself is told to put the plan aside and point you to the same help.
• What we keep. Only a daily count of how often help lines were shown, split by age band and US state. It has no link to you and none of your words.
• What we don't do. Nobody watches chats as they happen, and we don't contact anyone. Path is not an emergency service. In the United States, if you are in danger, call 911. Outside the United States, if you are in danger right now, call your local emergency number.

4. What the odds mean

How they are made

1. The AI lays out each route as a series of stages, for example "degree", then "application", then "selection". Each stage gets a pass rate, a number of years and a number of tries.
2. Where a published figure exists (for example "10 picked from more than 8,000 applicants"), the pass rate is based on it and links to the source. Otherwise it is marked as an estimate.
3. Your device then adjusts each stage's pass rate for your current grades, habits and experience, and works out the chance of getting through every stage, counting the tries each stage allows. That chance is your odds.
4. It also runs each route thousands of times (between 4,000 and 48,000 simulated tries) to see when people get there. The ages at which they reach the goal give a typical age and a range (the middle 80% of results).

This runs on your device, so when you move a habit slider the odds update straight away.

How to read them

• Use them to compare. They are best for comparing routes and seeing which habits move the number most.
• They are not about you as a person. They don't know about luck, health, money, family, discrimination, rule changes, or the many things you haven't entered.
• They are rough. Many pass rates are estimates, and small differences (say 4% against 5%) don't mean much.
• They haven't been checked against real outcomes yet. We plan to measure how well they match what really happens and publish the results.
• A low number is not a "no". Hard goals have low odds for almost everyone. The plan is there to show what would raise them.

5. How facts are checked

The facts pack

Path keeps a "facts pack" of sourced facts for 22 popular goals, such as astronaut, nurse, chef, lawyer and game developer. Each fact is a short claim in our own words, with the publisher, the year and a link to the page that states it.

When the facts pack is refreshed:

1. An AI researcher re-checks each goal against live sources, preferring official ones: government agencies, regulators, official statistics, the governing body of the sport or profession, and peer-reviewed research.
2. A separate AI fact-checker opens every cited page and marks each fact as supported, corrected or unsupported.
3. Unsupported or unchecked facts are dropped. If a goal ends up with fewer than 10 facts, it keeps its previous version.
4. A new version goes live only after automatic checks pass: every goal present, at least 10 facts each, and every link secure. [operator: describe any human spot checks]

Public data

Job skills and education levels come from O*NET, wages and job outlook from the U.S. Bureau of Labor Statistics, and college admission rates, prices and graduation rates from the U.S. Department of Education's College Scorecard. We load these from the official files. See Data sources.

How to tell them apart in the app

• A source tag links to the page a fact comes from.
• "estimate" marks a number the AI worked out without a direct source.
• College Scorecard records are marked as data.

Pages found by web search (in Edge and Map) are linked, but they are not checked the way the facts pack is. When no fact covers something, the AI uses its general knowledge, which may be out of date.

6. Limits

• The AI can be wrong, out of date, or sound more certain than it should.
• It can misread a photo. Check the grades or text it reads before you rely on them.
• It is built for the United States and English. Rules in other countries may be missing or wrong.
• It only knows what you tell it, and it can't check it.
• It may reflect biases in the data it learned from and in its sources. If you spot this, please report it.
• It is not a doctor, counselor, lawyer or financial adviser, and it is not an emergency service.
• There are daily limits on AI use to keep Path free, so it may pause until tomorrow.

7. How to report an answer

1. Tap Report this answer next to the answer.
2. Pick a reason: wrong, harmful, unsafe, offensive or other.
3. Add a note if you like (up to 500 characters), then send.

A person on our team reviews reports every week and uses them to fix prompts, facts and safety rules. What we keep with a report is explained in the privacy policy. You can also write to us at legal@feynmanlabs.com.

Reports are not watched in real time. In the United States, if you or someone else is in danger, call 911, and for suicide or self-harm, call or text 988. Outside the United States, if you or someone else is in danger right now, call your local emergency number.

8. Your data and the AI

Your plan details, chat messages and any photo you add are sent to the AI (Anthropic, or our own computers where we use them) only to get the answer you asked for. Your account number is not sent. We don't save chat messages or photos. For the details, see the privacy policy and service providers.