Wishful Mnemonics

Opening

Wishful mnemonics is Drew McDermott's 1976 phrase, from his paper "Artificial Intelligence Meets Natural Stupidity," for the failure mode where AI researchers give their program components intuitive, anthropomorphic names (UNDERSTAND, GOAL, BELIEVE, KNOW) without evidence that the named component implements the corresponding human-level concept. The phrase re-entered the AI discourse in 2025-2026 through Melanie Mitchell's writing on large reasoning models, where it describes how the field talks about reasoning-trace tokens, embedding geometry, and emergent capabilities.

Mechanism

The mechanism McDermott describes is easy to fall into. A researcher builds a system whose internal state has some structure. Rather than naming the structure by what it does ("G0034"), they name it by what they want it to do ("UNDERSTAND" or "REASONING"). The convenient name shapes the researcher's intuitions and the discourse around the system, creating a feedback loop: the system "understands" because that is what its main loop is called, and the apparent reasonableness of the name substitutes for evidence that the underlying function implements understanding. McDermott's prescription was deliberately orwellian newspeak: replace the intuitive name with a meaningless identifier and see if you can still argue that the component does what the intuitive name claims.

Tradeoffs

The trade-off in using wishful mnemonics is between two forms of cognitive economy. The intuitive name makes the system easier to talk about, easier to teach, easier to build intuition around, and easier to motivate funding for. The misleading name does the same things while substituting aspiration for evidence. Researchers and the public both come to believe the system has the named capability because the system is described as having it. McDermott's critique is that this substitution is a central disease of AI research: the field's intuitions about what its systems are doing are often wishful rather than evidence-based, with the convenient name doing the work of proof.

Open Questions

Whether the wishful-mnemonics critique applies with equal force to contemporary LLM discourse, where terms like "reasoning," "thinking," "understanding," and "belief" are routinely applied to transformer-based systems. Whether the critique can be operationalized, that is, whether there is a research methodology that distinguishes wishful mnemonic use from evidence of the named capability. Whether McDermott's prescription (use meaningless identifiers) is useful or produces worse research by making systems harder to think about. Whether the LRM discourse repeats the mistake McDermott identified in 1976 symbolic-AI research, or whether mechanistic interpretability and faithfulness studies have given the field new tools to escape the trap.