An illustration in Ray Kurzweil's 1999 robots-are-coming manifesto The Age of Spiritual Machines depicts a desperate-looking fellow, complete with sweating brow, laboring over a writing pad in a bare, bulb-lit cubicle. Strewn about the room, discarded sheets of paper bear such claims as "Only humans can play master chess" and "Only humans can recognize faces," each heavily crossed out. Among the casualties: "Only humans can improvise jazz." Yet another endeavor once thought to be our own exclusive cognitive province has, it would appear, been colonized by faster, smarter, ever more complex computing systems.

It's the kind of dramatic assertion AI boosters are fond of, and one that couldn't be better calculated to shake one's faith in our unreplicatable specialness. Even more than most creative endeavors, jazz is surrounded by a rhetoric of "intuition" and "inspiration," especially with regard to the central role of improvisation. This is partly historical: Many of the music's heroic figures have been longer on lived experience than formal musical training. But it's also because the very idea of creating moving (or even coherent) music "on the spot" or "in the moment" can seem like some mysterious conjuring trick, both to non-musicians and those schooled in the European classical tradition, with its strict separation between composition and performance.

Some jazz greats have played up the popular image of their activity: "If you haven't lived it," Charlie Parker famously said, "it won't come out of your horn." On the other hand, avant-garde saxophonist (and MacArthur fellow) Anthony Braxton has written witheringly of "the myth of the sweating brow" (the idea that a soloist is only really blowing if he appears to be undergoing inner tumult) and "the myth of the good night" (the idea that a quality of a jazz performance depends more on fate and circumstance than on skill or creativity). For Braxton, these notions -- especially when applied to black musicians by white critics -- devalue both the brainpower and long experience needed to create good improvised music.

Whatever your view of jazz, one thing is clear: Computers don't have "good nights," and their "lived experience" is, shall we say, limited. That's why the idea that one might extemporize with human speed and fluidity is such a fascinating one. Such a program, assuming it passed musical muster, might overcome the long-standing downside to recorded jazz, which "freezes" a single performance into an infinitely repeatable form. Our hypothetical Deep Blue Monk might, on the other hand, produce a fresh rendition of "Round Midnight" with every click of the mouse -- and never show up late for gigs. But don't throw your copy of Brilliant Corners out just yet. Some of the existing projects in the field fail to live up to Kurzweil's billing, while the most interesting don't even try.

What might make a program count as a full-fledged improviser? First, by definition, its output would have to take place in real time. Though only the furthest-out free-improv purists claim to have no preconceptions about what they're going to play ahead of time, a computer (or musician) that thoroughly composes what it's going to perform, and then performs what it has composed, is no improviser. Second, the program would have to allow for interaction between computer-generated output and other musical elements. There's a tradition of solo improvisation within jazz, most obviously among pianists, but the feedback loop among players responding to one another's moment-to-moment decisions supplies much of the music's interest.

The project most-often cited (by Kurzweil and others) as filling the bill is called, naturally enough, Improviser. The brainchild of British saxophonist and programmer Paul Hodgson, Improviser's most public gig to date was on a 1998 BBC segment, trading Parkeresque licks with noted jazzman Courtney Pine. These were hardly Turing Test conditions, and Pine's assessment of this digital duet was vague at best: "You can't replace the human experience, the human spirit." Unfortunately, both the BBC'S account of the session and a brief description of Improviser in a 1996 Discover article on computer creativity, by AI-Life specialist Margaret Boden, are short on details: Does the program's output sound like a saxophone, or is it a MIDI-voiced approximation? Does it replicate Parker's attack and phrasing, or only the timing and selection of notes? Most importantly, does it interact with other musical input -- and if so, how?

David Cope, the composer and programmer behind Experiments in Musical Intelligence (EMI) has heard the fruits of Improviser's labor. "Paul and I have sat down a few times and played each other tapes," he told me, "but I've lost track of him. At first swack, his program is really quite good at producing believable licks in a Parker style. It sounds like an improvised solo, but there's no ending or beginning, and even the most improvised music has points of arrival and direction. He hasn't yet had a chance to pay attention to the way a piece works. I liked what he did, but he agreed that the program needs quite a bit of work."

Cope's assessment carries weight: His own work might be thought of as the classical-music analogue to Hodgson's jazz-based efforts. Since its beginnings in the early 1980s, EMI, which Douglas Hofstadter has called "the most thought-provoking project in artificial intelligence that I have ever come across," has produced thousands of "recompositions" in the styles of everyone from Bach to Rachmaninoff to Cope himself. In three books about his work, each accompanied by CD-ROMs containing simpler versions of the code he works with, as well as audio samples of EMI's output, Cope offers both technical explanations of EMI's architecture and coding (in LISP) and speculations on its philosophical implications. The latest, the Algorithmic Composer (2000), focuses on Alice, a program that acts as a composer's assistant by applying EMI's principles to work in progress.

How does EMI work? First, Cope codes suitably similar music by a single composer as MIDI information. (The rigid structures of Chopin mazurkas or Bach chorales give the most impressive results.) The resulting database is chopped up into beat-long units, which are stored with both their harmonic character and their place in a composition's structure. "You could look at the information each piece contains," Cope says, "This piece has three simultaneous lines, here's where the lines were headed, and it's the second chord of a five phrase section of a four section piece." Additionally, a pattern-recognition subprogram uses music-theory criteria to identify characteristic "signatures" common to the body of work at hand. (It's this feature that lends some of EMI's output its spooky feel of being the handiwork of a single composer with a recognizable style.)

EMI then proceeds to recompose, by first selecting something it recognizes as a beginning, and then choosing subsequent units that make a snug structural and harmonic fit with what has come before. "The best choice is often the one the composer picked, and that will be nullified," Cope says. "It's also got to be sensitive to the fact that this is the sort of thing that the composer would do at the same time in same sort of place. It can't do certain things in the middle of a cadence, for example." At the level of overall form, the program borrows a page from classic AI work in natural language processing (NLP). Just as a sophisticated NLP program can derive the sentence "John was loved by Mary" from "Mary loved John," EMI applies what is called a "recursive transition network" to a piece's underlying grammar, eventually transforming it into a new one. The musically informed analysis upon which EMI bases its transformations is considerably subtler, but the program is ultimately rule-governed in a similar way.

Cope estimates that current versions of EMI can write about "two hundred bars of music a minute," though coding the source data can take considerably longer. (The output can either be played as a MIDI file or printed out as a score for human performance.) But speed is not the main aim of the project, which is firmly rooted in Cope's compositional background: "Very early on, I just said, 'I'm not going to strive to do this in real time.' Still, one can see how EMI could be applied to the goal of producing convincing improvisations in an extant jazz style. The bulk of swing and bebop consists of variations on a set song structure, most often one in thirty-two-bar AABA form, and the melodic content of a solo by, say, Parker is to some degree constrained by the song's harmonic content. Writing an algorithm to play such solos might be a matter of chopping up a body of performances into scales and patterns, and supplying instructions for how and when to apply them in a given harmonic context. Recent versions of EMI compose a piece's phrasing and dynamics along with its pitches and rhythms; there's no theoretical block to doing the same for jazz.

But even with real-time performance (which might be simply a question of computing power), a single solo voice enslaved to a predetermined rhythm section accompaniment would seem to be insufficiently interactive to count as improvisation. (As it stands, EMI actually composes all of a piece's voices at once, making the notion of a "solo" fairly meaningless.) Also, Cope is quick to point out that much of what gives jazz its character would be harder to mimic. "One of the most important things about jazz is its rhythmic fluidity, and one feature of EMI is it rhythmic rigidity. It has to know exactly when the last piece was so it can match the next piece to it." The samples he's heard of Paul Hodgson's Improviser run afoul of the same difficulties: "He formed his database from notated music, and then applied an algorithm that added the phrasing, so it had the rhythmic earmarks of a jazz performance style. I disagreed with that approach. I would want to find a way to chop up non-notated music in such a way that the performance is preserved."

Several projects are higher on Cope's agenda, but EMI may tackle jazz yet: "I've collected about 150 like-styled MIDI files from the Internet to use as a database. It's the worst possible generic slop, in a general piano style. But it does have the stylistic and rhythmic subtleties, the notes 3/1000 of a second after the beat or 4/1000 of a second before the beat. And I've written an in-house document that -- I believe -- sets the stage to write the code for a jazz improviser that would produce the same quality of music as the classical program. But I don't know when I'll have the hundred or two hundred hours it would take to write the code."