The systems you did not notice

A useful first observation is that artificial intelligence is already part of many ordinary systems you use.

The most direct measurement of this available comes from a survey of 3,975 American adults with a published margin of error of 2.6 points. Asked whether they had used an artificial intelligence product in the previous seven days, 36 per cent said yes, 50 per cent said no and 14 per cent were unsure. Asked separately about six specific categories of product, 99 per cent had used at least one of them in that week: weather applications 87 per cent, streaming 83 per cent, online shopping 82 per cent, social media 81 per cent, navigation 81 per cent, voice assistants 50 per cent (Gallup, 2024).

Sixty-four per cent of people did not recognise something they had done in the previous seven days.

That figure needs one honest qualification. Gallup decided what counts as an artificial intelligence-enabled product, and a person could reasonably argue about whether a weather application belongs on the list. What is not arguable is the direction and the size.

A second measurement tests knowledge rather than self-report. Eleven thousand and four American adults were asked, for six ordinary applications, whether artificial intelligence was involved. The mean score was 3.7 out of 6. Fitness trackers were identified correctly by 68 per cent, customer service chatbots by 65 per cent, product recommendations by 64 per cent, face recognition on security cameras by 62 per cent, music playlists by 57 per cent, and email spam filtering by 51 per cent. Thirty per cent got all six; 31 per cent got two or fewer (Pew Research Center, 2023).

Read the bottom of that list. Spam filtering is the oldest and most widely deployed consumer machine learning system in existence, it runs on every message every reader of this course receives, and it is the one people were least able to identify. Familiarity is not recognition. A system that has worked quietly for twenty years is precisely the system nobody attributes to anything.

There is a matching gap on the other side. Seventy-nine per cent of a sample of 1,013 artificial intelligence researchers said people in the United States interact with these systems almost constantly or several times a day. Twenty-seven per cent of the public thought they did (Pew Research Center, 2025). The experts and the public are describing the same week.

Every complaint mechanism in ordinary life needs a subject. You can dispute a bank charge because you can point at the charge. You can raise a grievance about a supervisor because the supervisor has a name. When a system reorders what you see, filters what reaches you, or scores you against a threshold, and you experience the result as simply how things are, there is nothing to point at, and the absence of a complaint is read by everyone involved as satisfaction.

This has a second consequence that is worse. The systems that most people cannot identify are the ones with the longest history and the widest deployment, and those are also the ones most likely to be working well. Spam filtering is genuinely excellent. Navigation is genuinely useful. So the population of systems people do notice is skewed toward the new, the badly built and the annoying, which produces a general impression of this technology that is drawn from an unrepresentative sample of it.

The inventory in Exercise 0.A is designed to correct that sample. Your own record may reveal both how many systems were present and which useful ones you would not want removed.