How it works
Horus doesn’t have an opinion. It has two datasets.
One describes what every job in America is made of. The other measures how people are really using AI right now. Neither was built by us, both are public, and you can go and read them yourself.
Source one
O*NET
U.S. Department of Labor
A public catalogue of American work. For every occupation it lists the skills that matter, and breaks each one down into the specific tasks people perform. It has been maintained for over twenty years, and it says nothing whatsoever about AI.
Think of it as an ingredients list for every job in the country.
Reference: onetonline.org
Source two
The Anthropic Economic Index
Anthropic · published free under an open licence
Anthropic sorts millions of anonymised Claude conversations onto that same list of tasks and publishes the totals. It records what people actually did, measured between 1 Apr 2026 – 1 May 2026. Nobody is asked to guess, and it is not a forecast.
Think of it as a meter reading, not a weather forecast.
Reference: the Index · the raw data
And here is the join
They share one word: task
Your courses decide which skills get looked at, and how central each one is to your degree. The verdict for a skill is the same for everyone who studies it — what changes from degree to degree is which skills show up, and how much of your course list leans on them.
The answer you get
Every skill lands in one of three places
- Automating
- Machines increasingly do this. Your role shifts to directing and checking the work.
- Transforming
- You will do this alongside AI. The skill worth building is the augmented version.
- Untouched
- A durable human edge AI is not reaching. Worth protecting.
The questions you should be asking
Can you trust it?
Does an AI decide whether my degree is safe?
No. Verdicts are arithmetic on the two datasets above. Claude does exactly two jobs here: it reads a course’s published outcomes to work out which skills that course teaches, and it writes the sentence of advice you see. By the time it writes anything, the verdict is already fixed and handed to it as a given.
How accurate is it, really?
We measure it and print the number, including when it isn’t flattering. A second reviewer sees one course’s outcomes and one skill we matched to it — with no access to our reasoning and no idea what the verdict was — and judges whether that match actually holds. On a random sample of 1,000 pairings, 56% held up (52–59% at 95% confidence). It runs with a fixed seed, so anyone can repeat it and get the same answer.
Has a person checked any of this?
Two things, on 16 August 2026. All 34 skill verdicts were reviewed and none were changed — those are the foundation every forecast rests on, which is why they get the human hours. And the course ranking was checked across 27 skills, confirming in each case that the course we rank highest really does involve more of the work AI is already doing than the one we rank lowest. Beyond that, the tagging is covered by the audit above rather than by review.
Is any of this a prediction?
No, and that distinction matters more than anything else on this page. The Index reports how AI was used in a measured window. It cannot see jobs that don’t exist yet, and it doesn’t model the next decade. Horus is a meter reading.
What do you collect about me?
No account, no uploads, no transcript — you pick a major from a public catalog and the page you get was built ahead of time and sits on a server as a file. What we do collect is usage analytics, through Google Analytics, and it is worth naming precisely rather than calling it “anonymous”: the pages you visit, what you type into the search box, which degree you open, which courses you uncheck on the confirm step, which skills you expand, and whether you play the video. Google sets a cookie and assigns your browser an identifier so returning visits can be recognised, and infers your approximate location from your IP address.
Why does that include the courses I uncheck?
Because it tells us which catalog entries are wrong, which is the single most useful thing we could learn from real use. It is also the most personal thing this site sends, since the set of courses you did not take is a partial academic record — so we would rather say so here than bury it. It is never tied to a name, an email or an account, because we hold none of those. If that is not a trade you want, a tracker blocker stops all of it and the site works exactly the same.
Honestly
What this can’t tell you
Every limit below is a real property of the data, not a legal disclaimer.
The list of skills is narrow. Horus reads courses against O*NET’s 34 work skills, and 31 of them land in the middle category. That is why most degrees end up looking similar to one another, and why a “Yes” currently rests on a single skill. A finer list would separate degrees far better, and it is the next thing we are changing.
It is one AI platform, not all of them. The usage measurements come from Claude conversations. Other tools get used differently, and none of that is in here.
Not every course publishes what it teaches. Of 11,320 course entries across the degree plans we read, 7,604 have published learning outcomes. The rest carry no skills at all — we do not guess from a course title, so they appear in your plan marked as unread. An unread course is one we have no evidence about, not one that teaches nothing AI-relevant.
A degree is not a job. Horus reads what your courses teach. What you go on to do with it isn’t in any dataset.
One cohort, one ship
Every summer, our students build a new app
Summer 2025
CapMetro
Live bus tracking for Austin, and which routes actually run on time.
Summer 2026
Horus
Reads a degree course by course and shows which skills are changing.
Summer 2027
This one’s up to you
Next summer’s cohort runs on sponsorship. That’s the whole model.
Back the next appApps for Good is run by Teamup, a registered 501(c)(3) — EIN 92-0649381. Gifts are tax-deductible in the U.S., and many employers match charitable giving.
Built on the U.S. Department of Labor’s O*NET framework and the Anthropic Economic Index, used under CC-BY.