Two letters, two philosophies, and a sixty-year-old fork in the road that almost nobody on stage at the AI conferences will name out loud. In 1962, Douglas Engelbart published Augmenting Human Intellect — a research program whose whole point was to make a person more capable. In the same era, John McCarthy and Marvin Minsky were christening Artificial Intelligence — a program whose point was to build a mind that does the task instead of you. Same machine. Opposite purpose. The industry took the AI fork and never looked back. We think it took the wrong one.
The difference is not a matter of degree, and it is not a matter of taste. It is a disagreement about what the machine is for. An AI succeeds when the human is no longer needed. An IA — an Intelligence Augmentor — succeeds when the human is more powerful, and more themselves. Flip the letters and you flip the entire ambition.
The same word, pointed two ways
Hold the two side by side and the stakes get clear fast.
- AI builds a mind that acts instead of you. The human is the thing to be automated away. The measure of success is that the human can step out of the loop entirely.
- IA builds a tool that makes you more capable. The human is the thing to be amplified. The measure of success is that the human is sharper, and that the sharpness survives the tool being switched off.
Both are honest engineering goals. They are just not the same goal, and conflating them is how we got a decade of products that are marketed as empowerment and built as replacement. The telling detail: every major venue now bars an AI from being listed as an author on a paper, for a single reason — authorship is accountability, and software cannot be accountable. The field already knows the human is irreducible. It just builds as though the human were optional.
The test
Here is the one question that cuts cleanly through the noise, and it is stolen, gratefully, from the Santa Fe Institute's David Krakauer. He divides the tools of thought into two kinds. A complementary cognitive artifact — the abacus, the slide rule, the piano — transfers its structure into you. After enough practice the physical object can be set aside, because you carry a working model of it in your head. You keep the gain when the tool is gone. A competitive cognitive artifact — the GPS, the calculator — executes for you. Take it away and, in Krakauer's words, you are “no better than when we started.” One kind teaches you to fish. The other delivers the fish and quietly lets you forget how.
Does using this leave you more capable when it's gone, or less? — the only question that matters
That is the whole test. Run any tool through it. Run any feature through it. If the answer is “less,” you are holding an AI move — a GPS for the mind — no matter what the box says. If the answer is “more,” you are holding an IA. The abacus leaves a smarter person behind when you take it away. The GPS leaves a stranded one. The difference is not the task it performs; it is what it does to the person performing it.
Why the field keeps shipping the GPS
Not out of malice. The competitive tool feels better in the demo. It is smoother, it asks less of you, it produces the answer while the complementary tool is still asking you to understand the question. McLuhan had the warning sixty years ago: every extension is an amputation. Reach far enough through a tool and you go numb in the faculty it extended. Krakauer supplies the mechanism McLuhan was missing — why some tools amputate and some don't. It turns on whether the tool transfers its structure into you, or merely acts on your behalf. A competitive artifact is one whose amplification has already flipped into atrophy; you just haven't noticed yet because it still works.
Krakauer's own fear, for the record, is not a HAL-style superintelligence seizing the controls. It is “an aggressive App, imperfect and partial, that diminishes human autonomy” — the recommender that saps your ability to explore and to judge, until discernment itself quietly becomes a thing you used to do. That is the near, boring, already-happening danger, and it is the one the competitive design makes inevitable.
Engineering to the other pole
If the test is real, it is not a slogan — it is a design gate you can hold every feature against. A few principles fall straight out of it. Coach, not serf: make the method legible — the reasoning, the playbook, the why — so the person internalizes it instead of just collecting outputs. A silent oracle that only emits conclusions is the GPS failure mode wearing a nicer interface. Leave a mental simulacrum: favor named patterns and repeatable procedures a person comes to know by heart, so the gain outlives the tool. Keep the option-set open: present the choices and the tradeoffs; let the human make the discerning call, especially where it counts; don't auto-pick and collapse the decision down to nothing. Expose the essence sometimes: don't always pre-chew the problem — let the person meet the raw version now and then, so skill actually compounds.
None of that is anti-technology. It is the opposite — it is technology held to a harder standard than “did it produce the answer.” The harder standard is: did it produce a better person, one who no longer strictly needs it. A slide rule leaves its structure behind in the mind that used it. Remove it and you reveal someone smarter. Remove the GPS and you reveal someone lost in a city they've driven through a hundred times. We would rather build the slide rule.
So: not AI. IA. Same two letters, pointed the other way. The reversal is a small joke and a whole thesis at once — which is about right for an idea still being worked out in the open. Augment the human. Don't automate the human away. And if a tool can't pass the one test — if it leaves you weaker the day it disappears — then whatever it is, it isn't on our side.