What the Modem Was Saying

AOL ended its dial-up internet service on September 30, 2025. The company posted a short notice saying it routinely evaluates its products and had decided to discontinue dial-up, and that the AOL Dialer software and AOL Shield browser would go with it. The service had run for over three decades, and it taught America what connecting sounded like. Two sounds went with it: the handshake screech and a recorded voice saying you had mail.

The Census Bureau counted 163,401 American households that used dial-up alone to get online in 2023. That is table B28002 of the American Community Survey, with a margin of error of about 5,900 households, and it works out to just over 0.13 percent of all homes with an internet subscription. People in that count heard the screech the week the notice went up. Some of them heard it on the last night.

The company’s own annual report counted about 26.5 million U.S. subscribers at the end of 2002. Every dial-up session on every one of them began the same way. One phone line served one machine at a time, and a household learned to listen for the hiss before making a call. Every session began with thirty seconds of noise, and everyone in the house knew what the noise meant.

The noise had a grammar. Almost nobody who heard it nightly knew the words. I learned some of the words because of a job. In the CompuServe years I was Toshiba’s Modems and Communications guy, and my days were built from failed connections. Initialization strings. Noisy rural lines. A hotel switchboard that hung up the call at forty seconds could eat a whole afternoon. When you do that work, the screech stops being noise. It becomes a transcript, and the whole job was finding where the negotiation broke.

Oona Räisänen published the score for it in 2012. She is a Finnish engineer who writes about radio signals, and she took a recording of a dial-up handshake, made by William Termini on an iMac G3, and turned it into an annotated spectrogram on her site. Time runs left to right, pitch runs bottom to top, loudness shows as brightness. Her labels walk through the picture second by second, and the sound you remember resolves into sentences.

Those first sounds belong to the telephone network itself. The calling modem listens for a dial tone, the same hum you hear when you pick up a landline, and then it dials the number in dual-tone multi-frequency signaling, two pure pitches per digit. Your touch-tone phone used the same code. The modem begins by speaking the network’s oldest language, because nothing data-specific exists until a second machine picks up.

The answering modem speaks first, and its first word is a single long note at 2100 hertz. That note is aimed at the network itself. The telephone system was built around a theory of human conversation: one person talks while the other listens. To keep you from hearing your own voice echo back, the network’s echo suppressors mute the return path while you speak. Modems cannot live with that arrangement, because modems talk at the same time. The 2100-hertz note is the answering machine’s opening demand: switch off the suppressors. Under the standards, the suppressors quit after about 400 milliseconds of the tone.

The note carries a second demand folded inside the first. Every 450 milliseconds the tone snaps 180 degrees out of phase, a click you can hear if you know to listen. Those snaps target the echo cancellers, the newer machinery that subtracts the echo and keeps both directions open. A steady tone defeats one device and leaves the other running, so the modem breaks its own tone, on schedule, until both kinds of equipment stand down. The cancellers are more stubborn; they hold out for a full second of the broken tone before they quit. On Räisänen’s chart the snaps show as thin vertical seams in the band of the note.

The two modems introduce themselves once the network’s echo machinery stands down. They trade short bursts of slow frequency-shift data, 300 bits per second, on two different pitches so each can tell the other’s voice from its own. Räisänen’s labels call this exchange the V.8bis transaction; the V.8 family is the menu protocol. The caller reads out the modulation schemes it can speak: V.34, V.32bis, V.22, the older and slower dialects. The answerer reads back the overlap. Those fuzzy buzzes in the middle of the recording are two machines settling on a shared dialect.

The modems then stop describing themselves and start measuring the line. Under the V.34 standard, each modem sends a comb of test tones, 21 of them spaced 150 hertz apart from 150 hertz to 3,750 hertz, skipping four frequencies reserved for other signals. A short loud burst goes first, then a longer quieter one. Each receiver measures what survived the copper, tone by tone: how much of it came through, and how much noise came with it. The two machines swap their lab results and settle on a speed the wire can carry. These probing tones are the low, mournful boings in the middle of the sound you remember.

The last loud stretch is rehearsal. Each modem sends scrambled data, pseudorandom bits with no pattern, so the receiver can tune its equalizer against the measured line and train its own echo canceller on the real channel. The rushing noise at the end of the handshake is that practice running at full power. Then the calling modem detects carrier, the speaker clicks off, and the session begins. The hiss that stayed on the line for the whole call was data.

None of it needed to be audible. The command ATM0, four characters typed before dialing, turns the speaker off for good. Almost nobody typed it. The factory default did the opposite and kept the speaker live until the connection was established. Räisänen has a theory about why. Telephone lines were audio territory first. The earliest modems were acoustic couplers, and you placed the handset into rubber cups and listened to your own call like any other. A modem that dialed in silence would have felt broken, and an audible handshake let you catch a busy signal or a human voice saying hello in time to pick up and apologize.

The sound had a second use. In the forum I could ask a stranger a thousand miles away what the noise did, and the answer located the failure. Most of them had never once been asked to listen to it. A screech that quit before the boings meant the modems had failed to agree on a language. Boings with no final rush meant the line could not carry the chosen speed, and the fix was to force a slower one. Retrains replayed the back half of the handshake. It was a diagnostic instrument sitting in every den in America, and we treated it as an annoyance.

That audible negotiation was the last one the public got to hear. Every device you own still negotiates each time it joins a network. Your laptop bargains with the router for an address. The phone proves to the access point that it knows the password, without ever saying the password aloud. A browser argues encryption with every site it touches, and the handset asks each cell tower for permission to join. All of it happens silently. The negotiation still happens every time a device joins a network, and it now happens where the people paying for it cannot hear a word.

The silence has taught us to forget that connecting is an agreement. Forgetting has a price. When the terms of a connection become inaudible, they start to feel like weather. Throttling feels like a slow evening. A captive portal presents itself as a door that was always locked. Carrier address sharing passes for the nature of the internet. The modem era made one small correction to that power imbalance: for thirty seconds per session, the bargaining played out loud in the room where the bill was paid. You could hear the other side answer. You could hear when it refused.

The screech said that a machine was asking for something in your name, and here, in plain air, was the other side’s reply. We called it noise. We imitated it as a joke. Quieter modems and faster lines buried the whole arrangement under silence. It was the last time the network showed its work.

The other sound AOL retired had a human being inside it. Elwood Edwards recorded four phrases in his living room in 1989, into a cassette recorder, for $200. He was 39 years old, and the $200 was the whole fee. The phrases were “Welcome,” “You’ve got mail,” “Files done,” and “Goodbye.” His wife worked for the company then, and the company liked his tape and never asked him into a booth. Edwards died on November 5, 2024, one day short of his 75th birthday. The service he greeted outlived him by eleven months, less a week.

A few smaller providers still sell dial-up, so the handshake survives in scattered houses and on farm lines at the far end of the copper. The stage where America learned the sound has gone dark. Räisänen’s poster is a score now: the tones and timings of a performance the network no longer gives for an audience. Almost no line will carry it aloud.

The ritual ran the way it always ran on the last night of September. Dial tone, digits, the long note with its scheduled snaps. The two buzzes, the measured boings, the rush of rehearsal, and then the click of the speaker switching itself off at carrier, exactly as designed. The negotiation closed with its terms accepted. Everyone in the room heard every word of it, the way they always had. The words meant nothing to them, the way they always had. Then it was October. The negotiation kept happening in silence, and nobody was left who remembered that the silence had once been speech.