The phrase "artificial intelligence meets natural stupidity" is the title of a 1976 paper by Drew McDermott of the MIT Artificial Intelligence Laboratory, published in the ACM SIGART Newsletter, issue 57, in April 1976. The "stupidity" in that paper belongs to AI researchers themselves: McDermott catalogs three habits of sloppy thinking in his own field and proposes fixes. The paper is not a measurement of how foolish people are, and no study cited here compares machine performance with human error.
Key takeaways
- McDermott's paper (SIGART Newsletter 57, April 1976, pages 4 to 9) is a self-criticism of AI research, written by a researcher who says he commits the faults himself.
- Its three targets are wishful names for program parts, treating natural language as the natural way to state problems, and reporting a program as finished when only a first version was built.
- The proposed test for the first fault is to rename a component with a meaningless label and see whether the claim still holds.
- The popular one-line quip about artificial intelligence and natural stupidity has no source confirmed here, so it should not be attributed to anyone.
What the 1976 paper argues
McDermott opens by describing AI as a field that has always sat close to the edge of respectability, and therefore close to crackpottery, and he cites two critics of the time, Hubert Dreyfus (1972) and James Lighthill (1973). He argues that the willingness to explore odd ideas is necessary in a young field, but that it had combined with the habits of hacker culture to weaken self-discipline. His stated aim is to poke fun at a few repeated mistakes, and he adds that he suffers from the three he describes.
The closing section lists further gripes, such as the belief that every action is a change of world state and the habit of calling a semantic network semantic. It ends with a reminder that the paper criticizes AI researchers harshly, that other fields would show equally bad faults on inspection, and that most AI workers are responsible people who produce good work despite the pitfalls of a difficult field.
Fault one: wishful mnemonics
A mnemonic is a name a coder gives to a procedure or data structure. McDermott's complaint is that AI coders named the main loop of a program UNDERSTAND or a data structure GOAL before showing that the code did anything of the kind. In his view, a researcher who does this assumes the answer to the question the program is supposed to settle, and may mislead himself first.
His remedy is a renaming test: call the main loop G0034, and see whether the researcher can convince anyone that G0034 implements some part of understanding. His own example is the General Problem Solver (GPS), which he says should have been called the Local-Feature-Guided Network Searcher. He also describes a spreading effect: a component with an impressive label makes its neighbors inherit the grandeur, so that a program called THINK soon acquires data structures called THOUGHTS.
| Planner term | Conniver term |
|---|---|
| GOAL | FETCH and TRY-NEXT |
| CONSEQUENT | IF-NEEDED |
| ANTECEDENT | IF-ADDED |
| THEOREM | METHOD |
| ASSERT | ADD |
McDermott co-designed Conniver, and he uses the table to make a point about the Planner column: writing GOAL gives the coder a feeling of enormous power, and he calls that feeling an illusion. He does not exempt his own language, noting that Conniver's choice to call multiple data bases CONTEXTS was also a mistake.
Fault two: unnatural language
The second section targets the assumption that English is a natural way to state problems to a program or to store knowledge inside it. McDermott argues that the purposes of reference in conversation differ from those of naming inside a program. A speaker says "the arm" or "it" to save breath, while a program needs several internal names for the same object to be tracked carefully. He also objects to the habit of illustrating every new representation scheme with English questions that reduce to simple database lookups, because real discourse depends on context.
This section stops short of dismissing the field. He calls understanding natural language the most fascinating and important research goal in the long run, and recommends studying what language is for before attaching conversational interfaces to programs that lack depth.
Fault three: reporting a first version as if a second existed
The third fault concerns research reports. McDermott describes a typical project in which the researcher builds a first version, finds holes in it, and then writes a final report full of allusions to a program that was never completed, with the admission that only a preliminary version was implemented tucked into a footnote. The report impresses readers and discourages later students, who assume the problem is solved.
His proposal is to divide a two-year effort in two. The first half goes to writing the next version of the program, and the second half to the report and to making the program usable by other people, with statistics on how hard it had to work and a clear failure message. He sums up the case for honest negative results in one sentence: "AI as a field is starving for a few carefully documented failures."
A renaming exercise for current claims
The following checklist adapts McDermott's G0034 test. It is an illustrative exercise, not a procedure from the paper or from any standard.
- Take the label in the claim, such as "understands," "reasons," or "decides," and replace it with a neutral code.
- Write down what the system was actually measured doing: the task, the data, the score, and who ran the test.
- Check whether the original claim still follows from that description alone.
- Ask for the failure cases, and for the version that was tested rather than the one described in a talk.
- Note whether the evidence comes from a program others can run or from a report about a program.
Used on any label, the exercise either confirms that the evidence supports the word or shows how much of the claim rested on the word itself.
FAQ
Who coined "artificial intelligence meets natural stupidity"?
The earliest use confirmed here is the title of McDermott's 1976 SIGART Newsletter paper. The more common saying, that artificial intelligence is no match for natural stupidity, circulates with several attributions, and none could be verified from a primary source, so it is left unattributed.
Does the paper say people are stupid?
No. The faults it describes are methodological habits of AI researchers, and McDermott includes himself among the guilty. Human error in general is outside its scope.
Is the paper still relevant to current AI systems?
Its subject is how claims about programs are named and reported, which applies to any field that labels systems with human terms. The specific systems it discusses, such as Planner, Conniver, and GPS, date from the 1960s and 1970s and are not current products.
Where can the paper be read?
A scanned copy is hosted at the address given in the Sources section. The scan has some character errors, so exact wording should be checked against the original newsletter issue.
Related reading
- Where AI inherits human limits
- Natural and machine intelligence compared
- Whether AI can surpass human intelligence
- Human-centered AI
More in AI Basics.
Last verified: October 2026
Sources
- Artificial intelligence meets natural stupidity, Drew McDermott, ACM SIGART Newsletter No. 57, April 1976, pages 4 to 9 (scanned copy)
- Artificial intelligence meets natural stupidity (bibliography entry with author, venue, and pages), Pablo Stafforini bibliography, 2026