AI Insight: Creation or Mechanization?

(Time to Read: 6 mins.)

In this post, I’ll cov­er main ideas and issues that have both dri­ven me away from and pulled me towards using AI-based tools for con­tent cre­ation. (If you’re more inter­est­ed in prac­ti­cal uses than my philo­soph­i­cal ram­blings, check out my oth­er post where I doc­u­ment prac­ti­cal appli­ca­tions of AI-based tools to improve the qual­i­ty of dig­i­tal con­tent in elearn­ing projects.)

Ever since I start­ed read­ing about Arti­fi­cial Intel­li­gence in my sec­ond year of art col­lege, I’ve been both fas­ci­nat­ed by and skep­ti­cal of it. Back in the mid-1980s, the idea of AI exist­ing in the real world felt like sci­ence fic­tion, like HAL-9000 in that big famous space movie by Stan­ley Kubrick. It all seemed rather mag­i­cal, and poten­tial­ly scary.

A Keyhole View on the Evolution of AI

In the real world, we’ve come a long way from the deter­min­ism of the Expert Sys­tems of the 1980s and from the first neur­al net­works I read about back in the 1990s. Even with the insane amount of tech­no­log­i­cal advance­ment that’s occurred in the past 30–40 years, I main­tain a healthy skep­ti­cism about the whole AI endeav­our. The term “Arti­fi­cial Intel­li­gence” itself still both­ers me.

This video from IBM does a real­ly good job of describ­ing the evo­lu­tion of AI, and its relat­ed sub-dis­ci­plines and appli­ca­tions:

AI can mean a lot of things

Recent­ly, with all the hype and excite­ment in the con­sumer mar­ket­place over AI-pow­ered ser­vices and devices, I have won­dered if many peo­ple parse the term as if the com­put­er medi­um is the “arti­fi­cial” part but the “intel­li­gence” part is actu­al­ly kind of gen­uine. It’s worth remem­ber­ing that the word “arti­fi­cial” is the main mod­i­fi­er. It’s not arti­fi­cial just because it’s being done by a machine. The intel­li­gence itself is still an arti­fice. In my opin­ion, the so-called “intel­li­gence” in AI is just a new form of mechan­i­cal fak­ery, like those 18th Cen­tu­ry clock­work automa­tons that could sim­u­late writ­ing in script with a quill pen.

Back in my col­lege days, one of my tech savvy media instruc­tors refused to even use the term “arti­fi­cial intel­li­gence”, pre­fer­ring instead to call it “arti­fi­cial ratio­nale”. He was being pre­cise with his lan­guage, and I still appre­ci­ate the care­ful dis­tinc­tion he made. If I was allowed to rename AI, I might con­sid­er using the term “sim­u­lat­ed intel­li­gence”.

AI research and devel­op­ment has many com­pet­ing philoso­phies and areas of spe­cial­ty. One philo­soph­i­cal axis is “Hard AI” ver­sus “Soft AI”. Hard AI seeks to syn­the­size a mod­el of intel­li­gence based as close­ly as pos­si­ble on the func­tion and phys­i­ol­o­gy of the human brain (i.e. an arti­fi­cial mod­el of human cog­ni­tion). By com­par­i­son, Soft AI is less rigid in its goals, aim­ing for accept­ably human-like end results, regard­less of the method­ol­o­gy or design.

None of what we call AI is tru­ly intel­li­gent: it’s just much bet­ter and faster at sim­u­lat­ing intel­li­gence using rapid cal­cu­la­tions per­formed on mass­es of data. The increased horse­pow­er of today’s soft­ware and hard­ware plat­forms makes it pos­si­ble for AI to per­form tril­lions of cal­cu­la­tions per sec­ond. But instead of enabling a more intel­li­gent mod­el, isn’t it real­ly just enabling a more rapid ver­sion of stu­pid­i­ty? Aren’t we just talk­ing about the quan­ti­ty and speed of infor­ma­tion being processed, ver­sus the qual­i­ty of some­thing resem­bling cog­ni­tion?

To use a sim­plis­tic com­par­i­son, I think that the recent tech­no­log­i­cal improve­ments in GenAI per­for­mance could prob­a­bly be equat­ed with the kinds of tech­ni­cal improve­ments we’ve seen in dig­i­tal graph­ics and video. Over the past thir­ty years, as the spa­tial res­o­lu­tion of dig­i­tal imagery has improved from the 1980s up to the present day, the qual­i­ty of the end-prod­uct has been ever-more com­pared to visu­al real­ism, or tak­en as being “real­is­tic”. In graph­ics, con­ve­nient pho­to-real­is­tic rep­re­sen­ta­tion has often been used as a bench­mark of suc­cess, sim­i­lar to how pho­tog­ra­phy has large­ly replaced illus­tra­tion for doc­u­ment­ing news events in mass media. It’s fast, and looks as good (or even a bit bet­ter).

But the out­ward appear­ance of real­ism real­ly just helps us to accept and under­stand the role that the a new medi­um can play. We tend to put our trust in what we think the truth has his­tor­i­cal­ly looked like. Back in the 1960s, Mar­shall McLuhan wrote about how we’ve tend­ed to accept new media as exten­sions of our­selves and in rela­tion to how they extend upon the role and use­ful­ness of old­er famil­iar media. But maybe that’s just a con­ve­nient short­cut for user accep­tance. Sur­face sim­i­lar­i­ties can mask a false equiv­a­lence between the new and old tools. The new tool could be very capa­ble of cre­at­ing sim­i­lar-look­ing prod­ucts, but behind the famil­iar sur­face effects there are deep­er dif­fer­ences and emer­gent by-prod­ucts that the new medi­um may be bring­ing to the table. We’re back to “the medi­um is the mes­sage”, once again, per Dr. McLuhan.

Even so-called “AI learn­ing mod­els” that incor­po­rate data from mas­sive sets (like the Inter­net) are still based on rules, para­me­ters, and train­ing data medi­at­ed by human design­ers, and we know how fool­proof our human designs can be (yeah, I’m look­ing at you, HAL-9000). AI devel­op­ers already admit that their mod­els come to sur­pris­ing con­clu­sions and exhib­it emer­gent behav­iours that they did not orig­i­nal­ly pre­dict, and for which the under­ly­ing caus­es remain unknown. While unfore­seen results can be a healthy part of sci­en­tif­ic exper­i­men­ta­tion (real, excit­ing evi­dence), it can also be wor­ri­some or even dan­ger­ous if it neg­a­tive­ly affects real users in the pub­lic mar­ket­place of open-source or com­mer­cial soft­ware. How much beta-test­ing should the gen­er­al pub­lic be exposed to?

What constitutes experience or deep knowledge?

My cyn­i­cal expec­ta­tion is that our robots and AI will even­tu­al­ly run amok and turn on their cre­ators. That’s based on some well-worn expec­ta­tions about the fail­ures and lim­i­ta­tions of human cre­ators, like we saw in the sequel to that big space movie, that came out many years lat­er (the one with Roy Schei­der, that was­n’t direct­ed by Kubrick).

What we seem to have today are neur­al net­works that devel­op “con­fi­dence” from per­form­ing mil­lions of com­par­isons between con­cepts and usage, devel­op­ing sta­tis­ti­cal mod­els of like­li­hood, sim­i­lar in con­cept to what they do for eco­nom­ic fore­cast­ing. Those sys­tems don’t devel­op expe­ri­ence so much as they devel­op ever-more accu­rate sta­tis­ti­cal mod­els to pre­dict accept­able answers to ques­tions, based on evi­dence of pop­u­lar­i­ty from oth­ers’ expe­ri­ences (real or fic­tion­al). GenAI does­n’t know what is true or objec­tive­ly believ­able. It only seems to deter­mine what is sta­tis­ti­cal­ly like­ly. Look­ing at it that way, maybe GenAI and Large Lan­guage Mod­els are just show­ing us a dis­tort­ed, “fun­house mir­ror” reflec­tion of our own expe­ri­ence.

GenAI does­n’t know what is true or objec­tive­ly real­is­tic. It deter­mines what seems sta­tis­ti­cal­ly like­ly.

For me, this all brings to mind the ques­tion of what con­sti­tutes expe­ri­ence or deep knowl­edge. I don’t think we under­stand the mechan­ics of human cog­ni­tion well enough to say if today’s neur­al net­works are even a close equiv­a­lent to how we learn or under­stand things. How­ev­er, from a sur­face glance of today’s com­mon­ly-seen results, the gap between man and machine ratio­nale does appear to be nar­row­ing (at least if you believe all the breath­less hyper­bole out there in the mar­ket­place). The Tur­ing Test, Alan Tur­ing’s famous post-WWII acid test to deter­mine if an AI’s per­for­mance was good enough to fool a human observ­er, has like­ly now become obso­lete. In many cas­es, today’s GenAI sys­tems are capa­ble of fool­ing many of us too much of the time, unless we know what to look for. Six fin­gers on one hand may just be a new anom­aly, not the new nor­mal, but it’s still dan­ger­ous to trust what you read and watch online. Media lit­er­a­cy and crit­i­cal think­ing skills are now more impor­tant than ever.

The Downside of AI Judgement for Content Creation

So, by now it must be pret­ty obvi­ous that I have reser­va­tions about using GenAI for some aspects of con­tent cre­ation. Even though the accu­ra­cy of ChapG­PT has improved ver­sion-after-ver­sion, AI hal­lu­ci­na­tions can still pose a big risk to the integri­ty of your AI-gen­er­at­ed con­tent (unless you’re cre­at­ing sur­re­al­ist art or you don’t care about your audi­ence that much).

In AI-gen­er­at­ed text, hal­lu­ci­na­tions can cause unre­lat­ed state­ments to be con­flat­ed into con­vinc­ing alter­nate “facts” about a his­to­ry that nev­er hap­pened, or they may encour­age you to use soft­ware fea­tures that sim­ply don’t exist. In AI-gen­er­at­ed imagery and video, real­is­tic-look­ing body parts have been har­vest­ed from a wealth of unknown sources and assem­bled into sur­re­al mon­strosi­ties with sev­en fin­gers and limbs that blend into the sur­round­ing fur­ni­ture.

Personal Conclusions

Hav­ing demon­strat­ed all my reser­va­tions on using AI, I am still becom­ing a con­sumer of AI ser­vices. The AI-dri­ven con­tent gen­er­a­tion tools I’ve seen are help­ful and pret­ty amaz­ing a lot of the time, and the more I use them, the more of my pre­cious time they may free up. But in terms of raw deci­sion-mak­ing, idea gen­er­a­tion, or judge­ment, I’m deter­mined to keep my own hand on the tiller of ratio­nale.

Maybe at some point, if the investor and mar­ket-dri­ven AI hype and psy­che­del­ic hal­lu­ci­na­tions die down, we’ll be left look­ing at our­selves in the mir­ror, won­der­ing how we can do things bet­ter for our real future.