What our AI refuses to say (and why that helps you)

Pathfinder's AI refuses three things: it will not invent a deadline or a cost, it will not tell you your chances of getting in, and it does not decide scholarship eligibility by opinion. Each refusal is built into the product, not left to a disclaimer. Dated figures it prints with the year attached: $11,950 average published tuition and fees, public four-year in-state (College Board, 2025-26).

This is written for a student or a parent who has typed a question into a chancing tool or a general chatbot, got a confident answer back, and wants to know what an honest tool would say instead. Pathfinder treats every path after high school as equally valid, so nothing below assumes college.

What does "refuse" mean for an AI counsellor?

An AI refusal is a behaviour Pathfinder's AI is built not to do: invent a deadline or cost, state your chances of getting in, or decide scholarship eligibility by opinion. It is a property of the system, not a tone setting. The rule applies in every wording of the question, and there is no advanced mode where the refusals come off.

The mechanism that makes a refusal possible is the claim label. Every claim in a plan or an AI counsellor answer carries one of three public labels, and a claim that cannot be grounded is labelled modelled or advisory — it is never promoted to grounded:

• Grounded. A claim Pathfinder can point to a specific, dated source for, checked for the exact fact and the cycle in view.
• Modelled. A figure Pathfinder computed from available data rather than read from a source: a reasoned estimate, not a fact.
• Advisory. General guidance that is not specific to this student or this exact context.

Notice what is missing. There is no label for "probably", and none for a number the model produced because a number was expected. A date or a figure with no sourced record behind it is not printed as a fact; it comes back as unknown, with the place to check.

Those rows are definitional, not sourced figures: they describe how the shipped product behaves today. The rest of this page shows each one working.

Why won't it invent a deadline or a cost?

Because a date and a price are both decisive claims — a claim, such as an application deadline, an admission requirement or a course requirement, where being wrong changes the decision itself. Pathfinder holds a decisive claim to a stricter freshness bar and never lets it soften into a lesser label. A deadline that is off by a week is not a small error; it is the whole outcome.

So the test is not "does this sound right?" It is "can I point at the page, and does the page say which year it means?"

A dated cost passes. The College Board reports average published tuition and fees for 2025-26 as "Public four-year in-state: $11,950, $340 higher than in 2024-25" (College Board, 2025-26). That figure names its sector, its year and its measure — a published price, before any grant aid — so it can be printed as grounded, for exactly what it says and no further.

A dated cycle passes too. NACAC reports the share of colleges rating each factor of "considerable importance" in the admission decision for the Fall 2023 cycle (NACAC, 2023). The cycle is stated on the page, so the figure travels with a year attached.

An undated figure fails. The U.S. Department of Labor's page on registered apprenticeship programmes carries a percentage about apprentices continuing in employment after they finish. It is a federal page and the claim may well be sound. But no reference year and no citation is attached to that figure anywhere on the page, so Pathfinder does not print the number. It can tell you the page exists and that the figure carries no year.

What you get in place of a missing number is a sentence and a link: it tells you it has no sourced data for that, and points you to the page that would settle it. For a deadline, "where to check" is the provider's or the institution's own page, because that is the only page that can be right. A source reference is exactly that narrow thing — a record that one specific fact was checked against one official publication for one cycle. It is never a quality score for a website and never a statement about a student.

If you want to see the labelled version at full length, our apprenticeship and college comparison prints every figure with its publisher, its year, and the mismatch between measures spelled out.

Why will Pathfinder not state your chances of getting in?

Acceptance odds are a predicted probability of admission. Pathfinder never states one, in any wording, because it cannot be computed honestly from a profile. Not as a percentage, not as a reach-match-safety label, not as an encouraging aside.

The reason is visible in what admissions offices say they weigh. In NACAC's Fall 2023 data, "High school grades in college prep courses" was rated of considerable importance by 76.8% of colleges and "Total high school grades (all courses)" by 74.1% (NACAC, 2023). Further down the same list sit factors no profile can hold: "Positive character attributes" at 28.3%, "Essay or writing sample" at 18.9%, "Student's interest in attending" at 15.7%, and "Counselor recommendation" at 11.9% (NACAC, 2023).

The problem is not that a model has too little data about you. It is that several of the weighted inputs are judgements made by people who have not met you, about an essay you have not written, in a cycle whose applicant pool does not exist yet.

A vendor that does publish a probability says something similar about its own limits. CollegeVine writes that "there are some things that we don't yet consider, such as essays, rec letters, high school context, early decision, and legacy", and that "we do not claim you will or will not get in. Rather, the engine is calculating your probability of acceptance" (CollegeVine, 2021). Those are CollegeVine's own words about CollegeVine's own product, quoted with the link, and this page says nothing further about it.

Pathfinder's position is that a number a reader will treat as a forecast should not be printed as one. So the counsellor answers the question underneath the question: what does this institution say it weighs, which of those things can you still change this year, and what is your second route if this one does not come through.

Why isn't scholarship eligibility decided by opinion?

Because eligibility is arithmetic against somebody else's published rules, and an opinion about it is worthless. Scholarship matching checks every approved award against your profile and sorts it into what you qualify for, what rules you out and why, and what cannot be determined yet. The catalogue is curated and human-approved, and each award is scored on ten eligibility dimensions: citizenship, residence, study country, field of study, pathway, grade level, GPA, demographic, financial need and deadline window.

Each dimension comes out open to everyone, restricted to certain people, or simply not stated by the provider. "Not stated" is never quietly rounded up to "open". The AI can explain a verdict and cite the award; it cannot overrule the check, and it cannot talk you into an award you do not qualify for.

Here is what that looks like against one real provider page: the merit scholarships published by The Ohio State University (Ohio State, 2026).

That page lists several separate awards, and Pathfinder scores each one on its own rules rather than blending them. The table below is a page-wide reading, so each row names the award the rule comes from; no row applies to every award on the page.

Three of the ten rows come out unknown, and that is the useful part. An unknown row names the one detail that would settle the question — here, whether a grade threshold applies, whether studying in the United States is a condition, and whether any demographic rule sits behind an award — so you know what to ask the provider instead of guessing. A tool that filled those rows in with plausible-sounding rules would look more helpful and be less useful.

Read as a single award, this table would mislead: the financial-need row belongs to the Land Grant Opportunity Scholarship alone, and no reader should take it to mean Ohio State's merit awards are restricted to Pell-eligible students. That is exactly why the check runs award by award and stores the rule it read.

What do you get instead of a guess?

Four concrete things, all of which you can check:

1. Every claim carries its label. Grounded, modelled or advisory, printed next to the claim, not buried in a footer. You know whether you are reading a sourced fact, an estimate or general guidance.
2. Every grounded figure carries a source and a year. Not "recent data shows" — the publisher, the cycle and the link, the way the College Board and NACAC figures appear above.
3. An unknown is named, not defaulted. When a rule is not stated or a date cannot be sourced, the gap is written down along with the place to resolve it. Silence is a worse answer than "I do not know — ask here."
4. A primary path and an explicit backup. Not a single verdict about your future, but two routes you have actually planned.

One thing you do not get: a claim that any of this produces better outcomes. Pathfinder does not claim to pick your best path, and it has not been shown to improve anyone's results. Its claim is narrower and testable — that it is honest about what it knows, and that you can check every sentence it prints.

What can't these numbers, or this list, tell you?

A refusal list is not an accuracy promise, and the figures above describe populations, not you.

• A grounded claim can go stale. It was checked against one publication for one cycle. Providers move deadlines and publishers revise series. That is why the year is printed beside the claim, and why decisive claims are re-checked.
• A modelled figure is still an estimate. Marking a number as modelled makes its status honest. It does not make it right.
• Pathfinder reads the sources it can reach. Not every page on the internet. Some official sites block automated readers outright, and a page nobody can open cannot become a source reference, however authoritative it is.
• National figures are not local ones. The College Board figure is a published national average across a sector (College Board, 2025-26), and the NACAC percentages describe colleges as a group in one cycle (NACAC, 2023). Neither describes the institution you are applying to.
• This describes today's product. The refusals change only by changing code, and the public glossary is the record of what each label means. If a refusal is ever relaxed, it changes there first.

How do you check any AI answer yourself?

These work on any tool, including this one. Run them the next time a chatbot hands you a confident date or figure:

1. Ask for the source and the year. A tool that cannot name both has given you a sentence, not a fact.
2. Open the link. Then find the exact figure on the page. A link that merely lands on the right website is not a citation.
3. Check the cycle. A deadline or a price from last year, quoted as this year's, is the single most common failure.
4. Check whether the number is a mean or a median. They answer different questions, and swapping one for the other quietly changes the story.
5. Ask the tool what it will not do. If the answer is "nothing", you have learned something about the tool.
6. Keep a backup path. No answer, however well sourced, should leave you with one route and no second plan.

If it helps to know what a tool costs before you start, Pathfinder's what is free and what is not page sets that out. The free plan includes five AI counsellor messages a day and one plan generation, and it needs no card.

Want a plan that says what it doesn't know?

Pathfinder builds you a plan for whichever path you pick — trade, apprenticeship, work, military, certificates, community college, university, or "still deciding" — and labels every claim as grounded, modelled or advisory. It will not invent a deadline, a cost, or your chances of getting in.

Start free — no credit card required.

All articles