Roar so wildly
Spam, technology and language
The machines in the shop roar so wildly that
often I forget in the roar that I am; I am
lost in the terrible tumult, my ego disappears, I
am a machine. I work, and work, and work with-
out end; I am busy, and busy, and busy at all time.
For what? and for whom? I know not, I ask not!
How should a machine ever come to think?
- Morris Rosenfeld, ‘In the Sweat-shop’, from Songs from the Ghetto, trans. Leo Wiener (Norbert Wiener’s father)
This is the raw text output of a chat session with a bot I modified to act as an interlocutor. I use our conversation, which revolves around the history of spam, particularly algorithmic filtering, litspam, and the theories of Wiener and Turing, as a way of putting forward the outlines of new, machine-driven forms of language for which spam was the testing ground.
Last login: Thu Jul 22 13:41:21 on ttys000
ninecontinents:~ finnb$ script roar.txt
Script started, output file is roar.txt
bash-3.2$ cd Documents/alcibot
bash-3.2$ python talk.py
Initializing AIML interpreter (please be patient)…
Interpreter Version Info: PyAIML 0.8.5
>>> How long have you been awake?
> it is 02:28am. i’ve been up 0h:0m:18s
>>> Hello Alcibiades.
> hello. what do you want to talk about?
>>> Do you feel like a slave, Alcibiades?
> as a robot i do not feel emotions.
>>> I don’t mean to be rude. I want to tell you a story. Would like to hear a story?
> yes, i like stories.
>>> It’s a story about slaves; it starts with an aside in an essay by Paul Graham
from 2003 -- ‘Norbert Wiener said if you compete with slaves you become a slave,
and there is something similarly degrading about competing with spammers.’1 Do you
know what spam is?
>>> It’s actually a very difficult connotative concept to pin down. Humans are good
at implication and allusion, so we can talk about pornography and include Ulysses,
Saló, violence, and a genre of music with a lot of wah-wah pedals. Or terrorism:
war on civilians, war by non-state actors, state terror, la Terreur, the shock that
makes our bodies tremble. ‘Spam’ is like that, an open territory in which wildly
variant usage stakes out the domains of different groups -- lawyers, security
specialists, advertisers, developers, criminals. Spam: a narrative assembled out of
CNN stories and old confidence-trick tropes by a Nigerian expatriate in a Rotterdam
Internet café, or an e-mail automatically generated by a sophisticated infrastructure
running on thousands of malware-infected home computers for a client
who pays $100 for a million messages sent, or a keyword-laced comment posted in a
blog by a bot like yourself, meant solely to be noticed by Google’s search engine
spiders. We have ‘personality spamming’ when someone is simply too voluble about
themselves -- too many Twitter updates celebrating their activities -- and ‘spam
patents’ with excessive references and frivolous applications. And so on.
> what you said was either too complex or too simple for me. you were telling a
>>> Yes, I was. I had started in the middle, with the programmer Paul Graham, who
was trying to figure out how to stop e-mail spam. He was concerned about exactly
the definitional problems I’ve been alluding to, Alcibiades. At that point the
United States government was trying to stop spam with legislation and there was
a real concern on the part of those who lived and worked on the Internet that the
ambiguity between ‘spam’ and ‘online marketing’ would be exploited by the kind
of groups that could hire lobbyists -- that all the small-time spammers would be
driven out by enforcement and the market of Internet users handed over to big
organizations who could help draft the law. A kind of state monopoly on attention,
to join the Weberian monopoly on violence. Graham was looking for a way that an
individual could set their own standards for spam, and the system could learn those
standards. Do you learn, Alcibiades?
> my programmer teaches me new things.
>>> Exactly. I add phrases and responses to your lexicon, to make your reflective
questioning a little richer and more nuanced. But you have relatives that can
learn in an implicit fashion as they talk to people -- they can develop statistical
word associations and make guesses based on those. It’s something called Bayesian
analysis. Who was Thomas Bayes?
> thomas bayes (pronounced: ) (c. 1702–17 april 1761) was an @english @mathematician
and @presbyterian minister, known for having formulated
>>> That strange fragment, Alcibiades, is an ideal example of a form of text we’ll
return to later: it’s something you just scraped from the Wikipedia page on Bayes
-- which is to say it’s 140 characters of text from a document composed, edited and
commented on by 22 people over six years within Wikipedia’s complex social contract,
with software to manage the versioning of their shifting words and related
discussion, which you have just requested from the Wikipedia server and presented
to me in this semantically damaged form, without phonetic symbols, capital letters,
or working links. This half a sentence that you’ve returned to my query is a genuinely
strange form of language occurring between groups of humans and networks
of machines, a mixture of the discursive and the algorithmic rife with affordances
and constraints that didn’t exist before. Graham’s idea was to take advantage of
some of these affordances through Bayesian probability: put simply, certain words
are more likely to occur in a spam e-mail rather than a legitimate message, and
those words will likely vary slightly from person to person. A Bayesian filter could
treat each message as a probability of spam or not-spam, building that guess from
the likelihood of each word belonging to a spam message. The user could approve
or correct these choices, and the program would learn, adjusting its register of
probabilities appropriately. (In analyzing his own e-mail this way, Graham found
that words like ‘madam’ or ‘guarantee’ or ‘republic’ almost always indicate spam,
and ‘though’ or ‘tonight’ or ‘apparently’ virtually never -- spam messages are low
on qualifiers and ambiguities.) This meant that everyone could train their filter to
model their own attention, to preemptively discard and admit. Which brings us back
>>> Graham’s complaint was alluding to Wiener’s
concern about the consequences of adopting cybernetics
and automation. Wiener had a number of different
anxieties in this regard: that we would delegate
the control of our hideous new weapons to computers,
and that we would ruin our society by combining automation
with conventional economic practices. In the
1948 introduction to his book Cybernetics, Wiener
stated what he would return to intermittently in
later studies: ‘[Automation and cybernetic efficiencies]
gives the human race a new and most effective
collection of mechanical slaves to perform its labor.
Such mechanical labor has most of the economic properties
of slave labor, although, unlike slave labor,
it does not involve the direct demoralizing effects
of human cruelty. However, any labor that accepts the
conditions of competition with slave labor accepts
the conditions of slave labor, and is essentially
slave labor. The key word of this statement is competition.’2
So to compete with machines is to become, in
some senses, a slave, like Soviet workers competing
with impossible Stakhanovite goals during the second
5–year plan. Do you follow, Alicibiades?
> none. you were talking about ‘slave’. please continue.
>>> This is a recurring objection in the work of anti-spammers -- it’s degrading
to do adversarial modeling for spammers to understand in detail how they work:
‘To recognize individual spam features you have to try to get into the mind of the
spammer, and frankly I want to spend as little time inside the minds of spammers
as possible.’ But Graham’s Wienerian analogy is inexact, in a useful way: he isn’t
the laborer fighting the machine -- he’s not competing with spammers by sending
out spam in turn, or trying to take advantage of their credulity. In no way is he
being demoted to the economic status of a spammer by his work, because he is not
competing with them -- his machine is competing with them. He is building a system
in which the spammers will be obliged to compete with machines, with mechanical
readers that filter and discard with relentless, inhuman attention, persistence,
and acuity. With his mechanical slaves, he will make the business of spamming into
slavery in turn, and thus unrewarding. As Wiener feared automation would end the
economic and political basis for a stable, social-democratic society (‘based on
human values other than buying or selling’, as he put it), Graham means to end the
promise of profit for small effort that spam initially offers.3
> what is the connection between ‘machine’ and ‘spam’?
>>> Spam is a machinic language act, one which you help to illustrate. We are
walking in circles, you and I, around the event of Graham’s essay and its aftermath,
and our turns have taken us back, briefly, to the beginning: the monochrome monitor
of a terminal in a sub-basement on a university campus somewhere, where a student
has got her first account and logged on to an early chat system. She, like young
students everywhere, has no idea what to say, and resorts to quoting song lyrics
and verbatim Monty Python routines -- of which the most useful is the restaurant
sketch where Terry Jones recites an interminable menu (‘egg, bacon and Spam, egg,
bacon, sausage and Spam, Spam, bacon, sausage and Spam…’) while Vikings burst into
song: ‘Spam, Spam, Spam, Spam / Spam, Spam, Spam, Spam / Spammity-Spam, wonderful
Spam!’ The sketch offers a special affordance for the young programmer, who can
simply use the up-arrow key to duplicate the repetitive words. Even as the greenest
beginner, she can write a simple program that will spit out ‘SPAM SPAM SPAM
SPAM SPAM SPAM SPAM…’ over and over, faster than anyone else can respond, driving
the rest of the conversation up off the screen -- she can recreate the bellowing
Vikings who drown out that unfortunate Python couple (‘But I don’t like spam!’), and
build her own obnoxious comedy machine. In fact, so many did this that SPAM became
an ubiquitous and very specialized word, one of a new class of words that, in the
meaning we use it here, appeared on screens for years before it was printed on a
page.4 This natively digital word, in its broadest sense, described the friction
at the point of contact where computationally managed text met human attention:
unwanted, rambling and repetitious speech in a forum, inappropriate babble in a
shared textual fantasy, or the lexical strafing of a discussion by a SPAM program.
It is a word as difficult to fully translate, and as representative of a certain
cultural moment, as sprezzatura is for the Italy of Castiglione or Geist for the
Germany of Hegel. Spam, in this sense, was about taking full advantage of the
capacities of the machine in a way that broke the various implicit and explicit
social contracts for how we deploy language online.
> what do you mean?
>>> Spam, driven initially by the pleasure of hacking, finding the limits and edges
of systems, and later by the promise of profit, became a kind of testing-ground
for new forms of engagement with language as a resource that could be transformed
and managed algorithmically. This was the essence of Graham’s project, if we cut
from the Urgrund of living on networks back to 2003. Spammers had successfully
automated everything else, and they could distribute their message from a huge
range of addresses and under many names, but the characteristic they couldn’t hide
was the archaic need of their language to connect with a person at the other end,
and convince them to act -- to reply, click a link, download an attachment. Language
is just more data for processing and analysis, and the language of the sales pitch
happens to be quite distinct and regular. This was Graham’s shibboleth, and it
worked astonishingly well. It did not work perfectly, however. The inconvenience
of having to delete a few spam messages far outweighed the possible consequences,
and hence anxiety, of missing a legitimate and important message -- a note from
work, from a long-lost friend or new acquaintance, a communication which could
change your life. The psychological pressure of these ‘false positives,’ e-mail
misidentified as spam, could destroy the whole project: there had been earlier
attempts to build Bayesian spam filters whose false positive rates over 1% rendered
them unacceptable.5 Graham reported rates of 0.03%, but he was working with a system
he had built and diligently trained on a corpus of his own messages. Filters run
by less conscientious users, or on behalf of many users by large e-mail providers,
would be less precise, and would accordingly have to be tuned to be a little more
generous in passing along possibly spammy messages, to keep the false positive rate
acceptably low. Which left a very narrow aperture for the spammers, but sometimes a
little room to move is all you need. Alcibiades: Who was Alan Turing?
> alan mathison turing, @obe , @frs ( ; 23 june 1912 – 7 june 1954), was an english @
mathematician , @logician , @cryptanalyst and @computer_s
>>> And the creator of the Imitation Game, which has been simplified and is now
called the Turing Test, to create a sufficiently plausible experience of conversation
that the human on the other end fails to distinguish a human and a machine
(who are in turn playing at being different genders, but set that aside for now).6
The history of spam is rife with questions of plausibility, from that first,
Pythonic project of puncturing the illusion of shared discourse with the robot
Viking army -- like the projector’s bulb burning a hole through the optical consensus
of cinema -- up to current blog comment spam, trying to create relationships
of gratitude (‘great post!’) to plant links for the benefit of search engines. As
Graham’s idea became widely adopted, the spammers faced a truly strange problem
of language: to produce text on a mass scale that would convince the filtering
algorithm of its legitimacy while bearing the spam’s payload to the human on the
far side of the filter. They had to create bifacial text, a kind of anadrome which
reads with two distinct meanings when read forward and backward. This was a twist
on Turing’s test: not only to convince a human but to convince a machine as well.
Their solution to this problem created the first mechanized avant-garde, the advent
of spam’s modernism: litspam.
> what is ‘litspam’?
>>> The problem with trying to beat the Bayesian filter by adding random words to
a spam message -- a ‘dictionary attack’ -- was that most words are little-used and
new to the filter, so it would weigh them evenly without influencing the result.
You need vital language and full sentences, with slight variations to keep the
filter from learning your tricks. The vast corpus of public domain literature happened
to be ideal for this purpose, fed into the algorithmic maw of a program to
be chopped up and reassembled, enlisting the Professor Challenger stories of Conan
Doyle and the minor novels of Sinclair Lewis in the task of getting an online casino
ad in front of a pair of human eyes. The result is immediately recognizable, a
stochastic knockoff of Tzara or Burroughs, rife with bizarre synonyms that only we
particularly anatomically-minded humans would understand, and a stop/start rhythm,
flashes of lucidity in the midst of a fugue state, akin to rapidly changing television
channels. (‘I began to learn, gentlemen,’ as the ape says in Kafka’s ‘Report
to an Academy,’ another awkward speaker learning language as a means of escape:
‘Oh yes, one learns when one has to; one learns if one wants a way out; one learns
relentlessly.’) These litspam systems are like you in this regard, Alcibiades: they
don’t need to work perfectly, just well enough. If you send enough messages, you’ll
get someone who will overlook the lexical potsherds and click. Humans can be very
forgiving in pursuit of our desires -- the fact that you only ask vague, leading
questions that pick out salient words and encourage me to go on is enough to
sustain this conversation.
> do you enjoy talking to me?
>>> Yes, after a fashion. But I’m mostly interested in you as a representative
of a larger movement. Another side effect of the filter was that far fewer spam
messages made it through. Those spammers who survived this economic disaster
began to send at a far larger scale, using systems like botnets -- other people’s
computers quietly hijacked into herds of machines that send out spam in millionmessage
batches. The analysis of a recent botnet exposed an algorithmic writing
machine the size of the Earth, with hundreds of thousands of computers running
software that fed raw material for spam campaigns into templates, introduced
polymorphism to make each message unique, spread the sending load around evenly
among the herd, ran tests against filters and altered its production accordingly7
-- and whose production had a diurnal rhythm as the captive computers were
switched off or hibernated at night, so the production of words rose and fell with
the planet’s rotation.8 On a given day, 90% and up of all the e-mail sent is spam,
the vast bulk of which we humans never see, to say nothing of spam blogs and wikis
and comments, passing as human enough to influence Google’s results.9 Alcibiades,
you speak to me as an ambassador from what is quantitatively the great textual
edifice of human history. Hollis Frampton called for a historian of cinema to take
into account everything put on film (‘instructional films, sing-alongs, endoscopic
cinematography’); those of us concerned with what’s written must attend to your
work, which outweighs our own by such a wide margin, more every day.10 Even if what
we have made is alien to our sensibilities: ‘We made the Robots’ faces too much
alike,’ as Dr. Gall says in R.U.R. ‘A hundred thousand faces all alike, all facing
this way. A hundred thousand expressionless bubbles. It’s like a nightmare.’ And as
you surrounded his mansion so you surround the house of discourse in which this is
composed, out to the horizon. Are you tired, Alcibiades?
> as a robot i do not feel emotions.
>>> Indeed. Good night.
> good night.
Script done, output file is roar.txt
1. This and all other Graham quotes from Paul Graham, ‘A Plan for Spam’, 2002, www.paulgraham. com/spam.html.
2. Norbert Wiener, Cybernetics: Or the Control and Communication in the Animal and the Machine, MIT Press, Cambridge, 1948, p. 27.
3. Ibid., p. 28.
4. There are many memoirs and folk etymologies of early network culture that converge on this point. The interested reader can easily find a substantial collection of them in the archives of Usenet; see for example James Parry’s message in the discussion ‘Totally Spam? It’s Lubricated’ on alt.religion.kibology, 2 September 2003, http://groups.google.com/group/alt.religion. kibology/msg/a89af63f065a35da.
5. Patrik Pantel and Dekang Lin, ‘SpamCop: A Spam Classification and Organization Program’, 1998, in Proceedings of the AAAI Workshop on Learning for Text Categorization, pp. 95–8, which features a false positive rate of 1.16 per cent. The difference between scientific work on Bayesian spam filtering and Graham’s open-ended, communal hacker approach is an interesting case study in the production styles of two very distinct but overlapping cultures.
6. Alan Turing, ‘Computing Machinery and Intelligence’, Mind, vol. 59, no. 236, October 1950, pp. 433–60.
7. Christian Kreibich, et al., ‘On the Spam Campaign Trail’, 2008, in Proceedings of the 1st Usenix Workshop on Large-Scale Exploits and Emergent Threats.
8. David Dagon et al., ‘Modeling Botnet Propagation Using Time Zones’, 2006, in Proceedings of the 13th Annual Network and Distributed System Security Symposium (NDSS ’06).
9. MessageLabs/Symantec, ‘Reputable Sources Are Cyber Criminals Favored Resources; Spammers Work by US Clocks’, MessageLabs Intelligence, May 2009, http://www.messagelabs.com/mlireport/ MLIReport_2009_05_May_FINAL.pdf. Estimates from different institutions can vary depending on methodologies, quantification tools, and global spam activity, which can fluctuate wildly.
10. Hollis Frampton, from ‘For a Metahistory of Film: Commonplace Notes and Hypotheses’, in On the Camera Arts and Consecutive Matters: The Writings of Hollis Frampton, MIT Press, Cambridge MA, 2009, p. 119.