AI Progress

Steampunk and a Thinking Crystal

https://tinyurl.com/t-invariant/2023/03/steampunk-and-thinking-crystal-en/

Machine assist­ed translation

The war on the bor­ders of the archa­ic world may be hin­der­ing the appre­ci­a­tion of the rev­o­lu­tion that is right now chang­ing the face of our civ­i­liza­tion. The emer­gence of a new gen­er­a­tion of arti­fi­cial intel­li­gence sys­tems has trig­gered process­es of tru­ly tec­ton­ic scale. Now we should keep track of them per­ma­nent­ly, but let’s start from dis­cussing the fate of copy­right in the AI world.

In human his­to­ry there are only about twen­ty or thir­ty such inno­va­tions, which pen­e­trat­ed into all aspects of life, changed the way of think­ing and behav­ior of peo­ple and shaped new lifestyle. The ongo­ing and, for many, unex­pect­ed break­through in the field of arti­fi­cial intel­li­gence is just one of them.

Although the dream of arti­fi­cial intel­li­gence is the same age as nuclear and rock­et tech­nol­o­gy, progress in this field has been very mod­est. The field of AI has devel­oped a rep­u­ta­tion sim­i­lar to that of nuclear fusion: when­ev­er you ask about the result, it is expect­ed in 30 years. With the slight­est suc­cess, the devel­op­ers tried to dis­tance them­selves from the theme of AI, say­ing: «we are engaged in real work — image recog­ni­tion or mul­ti­fac­tor opti­miza­tion, not fan­tasies about think­ing machines.» Even today, experts talk about neur­al net­works mere­ly as new meth­ods of infor­ma­tion pro­cess­ing. However, the capa­bil­i­ties that every­one has been test­ing since last year show that we are deal­ing with a major break­through. Whatever tech­ni­cal solu­tions engi­neers apply, the key point is that their man­i­fes­ta­tions are increas­ing­ly per­ceived by peo­ple as being much like their own think­ing. It is not the tech­no­log­i­cal aspect that is impor­tant, but the anthrop­ic aspect.

The juridical steampunk

The steam engine trig­gered a mass tran­si­tion of indus­try to machine pro­duc­tion. Much of what had until then required enor­mous labor began to be done mechan­i­cal­ly. Jobless arti­sans even rose up in spon­ta­neous revolts against the machines, but they were rea­son­ably answered: machines lib­er­ate them from hard phys­i­cal labor and monot­o­nous rou­tine, which are beneath human dig­ni­ty. A shin­ing dream was born of a future where «robots work and humans are hap­py,» exclu­sive­ly engaged in cre­ative work in which the machine could not replace them.

Luddites destroy the machine. Midjourney neur­al net­work generation.

Around the same era, a new judi­cial insti­tu­tion, copy­right, was rapid­ly spread­ing around the world. It was based on the notion that cre­ative labor is a divine spark, and repro­duc­tion is only a mechan­i­cal, machine dupli­ca­tion. Thereby the cre­ative work was rec­og­nized as qual­i­ta­tive­ly supe­ri­or to the phys­i­cal one, and there­fore the authors deserved a spe­cial remu­ner­a­tion. They were giv­en the right to con­trol the use of their works, espe­cial­ly their copy­ing. That’s where the name of this insti­tu­tion — copy­right — comes from.

By its legal essence copy­right (as well as its dis­tant rel­a­tive - the patent for inven­tion) is the priv­i­lege. It gives the author some spe­cial rights at the expense of the restric­tion of the civ­il rights of the oth­er peo­ple. The rhetor­i­cal jus­ti­fi­ca­tion for this redis­tri­b­u­tion of free­doms was the pub­lic ben­e­fit of encour­ag­ing the cre­ative activ­i­ty of authors. Evil tongues, how­ev­er, point­ed out that copy­right is much more impor­tant for pub­lish­ers, who, buy­ing author’s priv­i­lege got the pro­tec­tion from the rivals, i.e. got a small monop­oly, and there is noth­ing more valu­able in business.

The small monop­o­lies merged into large ones, built up legal ser­vices and suc­cess­ful­ly lob­bied the par­lia­ments to increase the term and vol­ume of copy­right pro­tec­tion. Authors with their long-suf­fer­ing man­u­scripts found them­selves on the weak side of deals with pub­lish­ers, for whom all these mas­ter­pieces (with a few excep­tions) were inter­change­able. This is how dump­ing in the pri­ma­ry copy­right mar­ket emerged. Only in the sec­ond half of the last cen­tu­ry, and only in devel­oped coun­tries under pres­sure of lit­er­ary agents, the aver­age roy­al­ty rate rose to about 10% of the retail price, but with the tran­si­tion to the elec­tron­ic for­mat it began to decline again.

The 10% for God’s spark of cre­ativ­i­ty is sur­pris­ing­ly rem­i­nis­cent of the church tithe, a tax that has long been con­sid­ered a kind of pay­off for the church’s patron­age of sec­u­lar activ­i­ty. And yet 10% is the typ­i­cal effi­cien­cy of steam engine. The copy­right has approx­i­mate­ly the same effi­cien­cy, if to mea­sure it by the share of mon­ey got from the mar­ket, which reach­es the authors. The rest is spent for the func­tion­ing of the juridi­cal mech­a­nism. However the church tithe and the steam engines were almost every­where reject­ed, but the copy­right sur­vived till the epoch of arti­fi­cial intel­lect, becom­ing the real juridi­cal steam­punk. But now there are hard times for it.

Collisions

Let’s turn to the news. A pho­to­bank is suing”>is suing the devel­op­ers of Midjourney, the neur­al net­work gen­er­at­ing pic­tures based on word queries, for what it con­sid­ers to be the net­work’s train­ing on copy­right­ed images. The pro­gram­mer is suing Copilot, a Microsoft-backed pub­licly avail­able AI assis­tant of the pro­gram­mer, for pos­si­ble copy­right infringe­ment. The Russian Speakers Union demands”>is demand­ing leg­isla­tive pro­tec­tion for voic­es from use in syn­thet­ic voice-over. A writer used AI to com­pose and illus­trate a chil­dren’s book and sell it on Amazon, but his rivals become are out­raged and demand that it be tak­en off the mar­ket because he did not write it him­self and did not work hard on it at all.

Work that had pre­vi­ous­ly been con­sid­ered pure­ly cre­ative was sud­den­ly, if not com­plete­ly, then large­ly with­in the pow­er of a machine. Yes, of course, ChatGPT will not come up with a new sci­en­tif­ic idea, Copilot will not invent a new algo­rithm, and Midjourney will not cre­ate a new artis­tic style. For now, any­way. But they already take an enor­mous amount of rou­tine part of the intel­lec­tu­al work, which in the copy­right ascent age was con­sid­ered cre­ative. It turns out that most of this God’s spark is eas­i­ly carved by a machine.

It would seem that, as in the case of the mech­a­niza­tion of phys­i­cal labor, we can only rejoice that the com­put­er is now ready to rid us of the mass of intel­lec­tu­al rou­tines, free­ing up resources for more pro­found think­ing than what AI is capa­ble of so far. But the new Luddites resent the encroach­ment of machines on the sacred ter­ri­to­ry of cre­ativ­i­ty, and demand that their poten­tial for inno­va­tion be arti­fi­cial­ly lim­it­ed so that the cre­ative efforts of intel­lec­tu­al arti­sans remain in demand by the mar­ket. As the main weapon to defend their con­ser­vatism they appeal to copy­right, which has always been pre­sent­ed as the mech­a­nism to sup­port cre­ators and innovators.

There is noth­ing sur­pris­ing here. The copy­right has always restrict­ed the free­dom of cre­ativ­i­ty and infor­ma­tion exchange. In many coun­tries it was ini­tial­ly intro­duced in the same law pack­age with the cen­sor­ship. In the more close to us epoch the copy­right path is marked by such scan­dalous sto­ries as:

  • attempt to ban VCRs (Sony v. Universal, 1983 – 1984);
  • pros­e­cu­tions for cir­cum­vent­ing the inten­tion­al incom­pat­i­bil­i­ty of DVDs from dif­fer­ent regions of the world (DeCSS, 1996 – 2003); 
  • killing of the first P2P music ser­vice Napster (2001), which great­ly delayed the devel­op­ment of online dis­tri­b­u­tion of multimedia; 
  • mass rack­e­teer­ing under the threat of copy­right law­suits for down­load­ing porno­graph­ic movies from tor­rents (Prenda Law, 2012 – 2013); 
  • destruc­tion of Google’s large-scale book dig­i­ti­za­tion project (Open Book Alliance, 2011 – 2016);
  • bar­ring fan­f­ic authors, such as Paramount’s law­suit against an ama­teur film based on Star Trek uni­verse (Axanar, 2015 – 2017). 

The copyright and censorship

The last point of the list illus­trates the most destruc­tive for the cul­ture neg­a­tive effect of the copy­right — the prob­lem of the derived works. They are almost com­plete­ly fall­en out of the free cre­ativ­i­ty sphere and passed into the cat­e­go­ry of reg­u­lat­ed pro­duc­tion, because their cre­ation is now start­ed from the legal depart­ment, but not from the author, who has con­ceived the trans­la­tion or stag­ing, fan­f­ic or crossover, arrange­ment or remake.

In a world full of brands and quotes, it becomes unsafe for an inde­pen­dent author to reflect real­i­ty at all. Even the back­ground music in a café while record­ing a video blog can cause the video to be strick­en from Youtube. And this is anoth­er class of con­flicts: when there is no vio­la­tion, but it is eas­i­er for a pub­lic plat­form to pro­hib­it a sus­pi­cious pub­li­ca­tion than risk legal costs. This is already very close to the log­ic of cen­sor­ship and self-cen­sor­ship famil­iar to res­i­dents of ide­o­log­i­cal dic­ta­tor­ships. Not only what is for­bid­den is blocked, but also any­thing that might seem to some­one sim­i­lar to what is forbidden.

In order to at least approx­i­mate­ly char­ac­ter­ize the scale of the trou­ble, let us make a quan­ti­ta­tive assess­ment. In the USSR, sev­er­al thou­sand authors and tens of thou­sands of works were under cen­sor­ship restric­tions. Many of them were avail­able in spe­cial repos­i­to­ries by spe­cial per­mis­sion, and sev­er­al thou­sand cir­cu­lat­ed in samiz­dat. There are no exact fig­ures, but dif­fer­ent researchers agree on an order of mag­ni­tude estimates.

In the United States there are many so-called orphan books, that is, books that can nei­ther be reprint­ed nor dig­i­tized because of the loss of data on the copy­right hold­ers. Most of them are avail­able only in the largest libraries and pri­vate col­lec­tions, which is to some extent com­pa­ra­ble to the avail­abil­i­ty of spe­cial libraries and samiz­dat in the USSR. According to the U.S. Copyright Office (2015), the num­ber of orphan books runs into the mil­lions, and British researchers esti­mate that by 2009 there were 13 to 15 mil­lion. That’s about a third of all books ever pub­lished in the U.S.

It turns out that quan­ti­ta­tive­ly copy­right in the U.S. is asso­ci­at­ed with two orders of mag­ni­tude more restric­tions on access to books than cen­sor­ship in the Soviet Union with a com­pa­ra­ble pop­u­la­tion of coun­tries. And this is just in one aspect of orphan pub­li­ca­tions. The giv­en exam­ples illus­trate the archa­ic nature of copy­right insti­tute: declar­ing the sup­port of cre­ativ­i­ty and inno­va­tions, in prac­tice it sup­press­es them in the inter­ests of those, who were able to get the upper hand on that mul­ti-monop­oly market.

The steam locomotive in a dead end

And still the copy­right has­n’t yet faced — open­ly, at least — the tec­ton­ic inno­va­tions, which change the tra­jec­to­ry of civ­i­liza­tion. Artificial intel­li­gence is appar­ent­ly a wor­thy oppo­nent, which has a poten­tial if not to stop, then to slow down the copy­right locomotive.

The basis of copy­right is not the pro­tec­tion of the cre­ative ideas them­selves, but only the form of their embod­i­ment. In the steam age the machine could only repro­duce this form and the copy­right allowed the cre­ator to par­tic­i­pate in the prof­its from its repro­duc­tion. [This sen­tence is a direct result of machine trans­la­tion from Russian. — T-i]

Let us take the pre­vi­ous para­graph and ask ChatGPT to para­phrase what was said, pre­serv­ing the mean­ing, but chang­ing the words and turns of phrase if pos­si­ble. Here’s the result, with not a sin­gle let­ter man­u­al­ly changed:

Copyright law is not designed to pro­tect cre­ative ideas as such, but rather the tan­gi­ble expres­sion of those ideas. During the Industrial Revolution, machines could only repli­cate the form of the work and copy­right law pro­vid­ed a means for cre­ators to ben­e­fit from this replication.

A lit­tle more ver­bose, but quite accept­able in mean­ing and style, and not a sin­gle antipla­gia­rism ser­vice will not pick up on it. But this is a pure­ly machine-made prod­uct, and such can be pre­pared from any oth­er text.

Here is a more com­pli­cat­ed exam­ple. Kipling’s famous poem ‘If – ’ is per­fect­ly suit­ed to the cur­rent tur­bu­lent times:

If you can keep your head when all about you
Are los­ing theirs and blam­ing it on you,
If you can trust your­self when all men doubt you,
But make allowance for their doubt­ing too.

Let’s ask ChatGPT to rewrite it with dif­fer­ent words and a dif­fer­ent rhyme:

If you can stay com­posed when those around
Are los­ing grip and lay­ing blame on you,
If you can keep belief in what you’ve found,
But under­stand when doubt clouds oth­ers’ view..

Can we ask for anoth­er ver­sion? «Sure,» the neur­al net­work replies:

When those around you are caught in fear,
And point their fin­gers to lay blame severe,
Trust in your­self, though oth­ers may sneer,
And grant them space, as doubt may yet appear.

Checkmate, aspir­ing poets. Are these poems copy­right­ed? To whom do they belong? Of course, if you admit to every­thing, as here, many will say: these are deriv­a­tive works — auto­mat­ic verse trans­la­tions from English to English. But what if you don’t admit it? Connoisseurs, of course, will notice the rep­e­ti­tion of the idea, but the copy­right does not pro­tect it. And who is the author of the new versions?

Rudyard Kipling and an android with arti­fi­cial intel­li­gence. Midjourney neur­al net­work generation

A Threat to the Future

There is no gen­er­al­ly accept­ed answer to this ques­tion, although it has been debat­ed for a long time (2017). Even ear­li­er (2011) there were dis­putes about the copy­right of works cre­at­ed by ele­phants and mon­keys. In some coun­tries, such as Germany, the law does not allow any­one oth­er than humans to be con­sid­ered the author. This is a lega­cy of the ancient notion of «God’s spark». In the US and Australia there is no such restric­tion, but case law, which has the force of prece­dent, denies the copy­right for machine cre­ativ­i­ty. In the UK and New Zealand, on the oth­er hand, the law rec­og­nizes author­ship by «the per­son by whom the arrange­ments nec­es­sary for the cre­ation of the work are undertaken.»

The lat­ter option seems, at first glance, to make the most sense. Authors have long been using sophis­ti­cat­ed tech­nol­o­gy, such as cam­eras, аnd despite that the shots are easy to make, you still have to choose the good ones. Even now, in order to make a neur­al net­work cre­ate an inter­est­ing draw­ing or text, a cer­tain orig­i­nal­i­ty in the word­ing of the task may be need­ed. However, there is a hid­den dan­ger in this approach, which can under­mine the whole arti­fi­cial intel­li­gence indus­try, if things go wrong.

Firstly, the neur­al net­work itself is cre­at­ed and trained by some­one. This means that accord­ing to the let­ter of British law, its devel­op­ers auto­mat­i­cal­ly become co-authors of all its works. This can lead to a mon­strous con­cen­tra­tion of copy­rights and cen­sor­ship tools in the hands of such neur­al net­work devel­op­ers. They already put restric­tions that do not allow, for exam­ple, to order the writ­ing of phish­ing emails, pass­word crack­ing pro­grams, and sim­ply offen­sive or polit­i­cal­ly incor­rect texts.

Secondly, train­ing neur­al net­works requires huge amounts of infor­ma­tion. They use all kinds of texts and images, includ­ing copy­right­ed ones. Authors and copy­right hold­ers, sens­ing the threat of com­pe­ti­tion from neur­al net­works, are sound­ing the alarm and demand­ing a ban on train­ing arti­fi­cial intel­li­gence on their works. Such a demand may seem rea­son­able until we real­ize its destruc­tive potential.

AI sys­tems have already entered our lives far more deeply than is com­mon­ly thought. Enough to say that search engines such as Google, Bind, or Yandex are spe­cial­ized AI trained on the pages avail­able on the Internet, pro­tect­ed, of course, by copy­right. If we now require that search engines obtain the pri­or writ­ten con­sent of copy­right hold­ers to use texts for index­ing, the Internet will col­lapse. Machine trans­la­tors, effec­tive­ly removed the lan­guage bar­ri­er in busi­ness com­mu­ni­ca­tion around the world. They, also, were trained on copy­right­ed par­al­lel texts. Banning such use of copy­right­ed mate­r­i­al is to destroy the Babylon Tower once again.

Obviously, nobody is ready for such cat­a­stro­phes. And if so, to impose copy­right restric­tions on edu­ca­tion­al mate­ri­als for neur­al net­works in numer­ous oth­er cas­es would be inap­pro­pri­ate dis­crim­i­na­tion. Especially since neur­al net­works do not copy oth­er authors in their works, which would be a vio­la­tion of copy­right. If they do bor­row some­thing, this is very dif­fi­cult to express in an objec­tive way, espe­cial­ly since no law pro­hibits one author to mim­ic the style and man­ner of anoth­er, as long as there is no direct copying.

However, if AI sys­tems con­tin­ue to improve, absorb­ing all of human­i­ty’s accu­mu­lat­ed cul­ture, it can­not be denied that many authors’ works, espe­cial­ly those of a util­i­tar­i­an nature, will be great­ly deval­ued. After all, any­one would be able to gen­er­ate a sim­i­lar ana­log using arti­fi­cial intel­li­gence quite inex­pen­sive­ly. Moreover, such a gen­er­at­ed ana­logue - for exam­ple, a text­book on a cer­tain top­ic - can be addi­tion­al­ly adapt­ed to the needs of the cus­tomer, some­thing that a cir­cu­la­tion-ori­ent­ed prod­uct will nev­er provide.

End of the track? 

So are we look­ing at the dis­ap­pear­ance of the author’s pro­fes­sion as one who lives from roy­al­ties? Partly yes, but, as it usu­al­ly hap­pens in such cas­es, not all the way. In case of the copy­right wane, the author’s work will become to some extent sim­i­lar to the ordi­nary labor, when the per­son earns more by his per­son­al par­tic­i­pa­tion, than by the roy­al­ties from the machine copies. This process has been going on for a long time. Artists used to make more mon­ey from sell­ing orig­i­nals than from repro­duc­tion rights, musi­cians’ income is shift­ing from sell­ing copies to con­certs, lit­er­ary roy­al­ties are declin­ing (The Authors Guild, 2018), and many of them increas­ing­ly see books as a per­son­al brand pro­mo­tion tool rather than a source of income.

Of course, there will always remain a spe­cial expe­ri­ence of per­son­al con­tact with cre­ative peo­ple, but their copy­right­ed works are like­ly to occu­py rough­ly the same place in the infor­ma­tion mar­ket as hand­made and lux­u­ry goods occu­py in the house­wares mar­ket. The works made with feel­ing and issued in small cir­cu­la­tions with per­son­al par­tic­i­pa­tion of the author will always be in demand, but the epoch of mass repli­ca­tion of cul­ture will expe­ri­ence the shift towards machine per­son­al­iza­tion. This is rough­ly how it has already hap­pened with the news that we have begun to receive from our per­son­al social media feeds.

To sum­ma­rize, we state that copy­right and arti­fi­cial intel­li­gence as cul­tur­al phe­nom­e­na are in deep con­tra­dic­tion with each oth­er. The suc­cess of one of them is poor­ly com­pat­i­ble with the suc­cess of the oth­er. At the same time, copy­right belongs to the world of the archa­ic, maybe warm and cozy, but going back to the past, while the think­ing crys­tal with arti­fi­cial intel­li­gence is just the case, which Joseph Brodsky wrote that «the irrup­tion of the future into the present is felt as being a source of dis­com­phort, if not of down­right dis­cour­age­ment.» (The UNESCO Courier, 1990, #6, p. 31)

Text: ALEXANDER SERGEEV

  17.03.2023