AT&T (Bell System)

Contribution to AI

A regulated monopoly turns out to be an odd but effective way to fund basic science. The licence fees the operating companies paid into headquarters — usually around one per cent of their revenues, recovered through telephone rates rather than justified against expected returns — kept a laboratory alive through decades in which most of its work had no market. The 1956 consent decree that settled the government's antitrust suit then determined how the results escaped: barred from any business other than common carriage, the company could not sell software, so it distributed its operating system with full source to universities for little more than the cost of the tape. Computer science departments spent the 1970s building on a complete system they were free to read and rewrite, and the language it was written in became the one nearly every later numerical library assumed underneath it. AT&T also ran the largest machine anyone had assembled, which let its pattern-recognition work be judged against real traffic instead of benchmark corpora. Voice Recognition Call Processing, deployed in 1992, automated the handling of collect and calling-card requests from unrehearsed callers over ordinary telephone lines. Through NCR, which it owned from 1991 to 1996, the convolutional networks developed at Holmdel went into cheque-reading machines that by the late 1990s processed something like a tenth of American cheques — a trained network working as operational equipment rather than as a published result.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).