Can most adults write prompts good enough for AI?

Jakob Nielsen, writing on UX Tigers in 2023, argues that prompt-based generative-AI interfaces trade one usability problem for another. They remove the "myriad of tedious commands" of command-based UI, but they require users to be "highly articulate to write the required prose text for the prompts." Citing OECD PIAAC literacy data, he notes that roughly half the adult population in rich countries like the US and Germany tests at low literacy levels, and concludes that "less than 20% of the population is sufficiently articulate in written prose to make advanced use of prompt-driven generative AI systems," adding that "10% is actually my maximum-likelihood estimate."

His reasoning runs through PIAAC's five-level reading scale: level 3 is "the first level to represent the ability to truly read and work with text," while only levels 4-5 (about 1% of the population at level 5) can perform the "complex inferences" needed to integrate and synthesize information. Because producing new prose is harder than reading someone else's, he suspects the share of "low-articulation users" exceeds the share of low-literacy readers, though he is upfront that this is a hypothesis: he has "been unable to find large-scale international studies of writing skills." As his one piece of supporting evidence he points to the existence of "prompt engineer" as a job title — if ordinary articulation were sufficient, a specialist to translate intent into working prompts wouldn't be a hireable role — and he likens the problem to the familiar failure of enterprise IT departments to build what a department head's written spec actually asked for.

This literacy-gradient framing sits apart from Does AI literacy reduce the damage from AI disclosure?, which locates its moderating skill in AI-specific savvy; Nielsen's bottleneck is one level earlier, in the general prose literacy a user needs before they can produce a workable prompt at all. It also cuts against Does polished writing actually signal better quality work? and Does polished AI output trick audiences into trusting it?: both of those locate AI's usability failure in how polished output deceives evaluators after generation, while Nielsen locates a prior failure at the input stage, where most users cannot articulate the prompt needed to get good output in the first place.

The excerpt does not establish its central number. PIAAC measures reading, not writing or prompt articulation, so the 10-20% estimate is explicitly "my early, coarse analysis," and Nielsen's own call for "detailed qualitative studies" concedes that no one has yet tested how literacy level maps onto actual prompting success. His single piece of evidence — that "prompt engineer" became a job — is anecdotal, not measured. If the general direction holds, it implies that prompt-only AI interfaces may be structurally inaccessible to a majority of adults in exactly the countries where AI adoption is being pushed hardest, an access problem distinct from, and prior to, the output-quality problems other notes describe.