Plea for more AI competence

Rea­ding time: 7 minu­tes

Gene­ra­tive AI can sup­port rese­ar­chers and crea­ti­ves in many areas. Howe­ver, AI com­pe­tence is essen­tial for its pro­fi­ta­ble pro­fes­sio­nal use. This article deals with the ques­tion: What is AI com­pe­tence and how can it be acquired?

Our last blog article from Sep­tem­ber 2023 on “AI in mar­ket rese­arch and inno­va­tion” was still titled “expe­ri­ment”. The term “inte­gra­tion” would now be more accu­rate, as gene­ra­tive AI has become a natu­ral tool for us in both the rese­arch and crea­tion pro­cess: ChatGPT, Whisper, Mid­jour­ney, Fire­fly, Flux, Run­way, Krea, Luma, Suno, and cer­tainly a few more. At the same time, we see that there are still many myths, exag­ge­ra­ted expec­ta­ti­ons and a lot of bull­shit cir­cu­la­ting around the topic. We are the­r­e­fore cal­ling for a little more AI com­pe­tence – and in this article we want to ask what con­sti­tu­tes AI exper­tise and how to achieve it.

The term “AI com­pe­tence” is a direct ana­logy to the long-estab­lished con­cept of “media com­pe­tence”. Media com­pe­tence is what we demand when deal­ing with social media and fake news, what we want to teach our child­ren and what we some­ti­mes admire – or some­ti­mes cri­ti­cize – in young peo­ple. This does not mean skep­ti­cism, fear or with­dra­wal from social media, for exam­ple, but – on the con­trary – an adult, reflec­tive, cri­ti­cal atti­tude towards the media, ques­tio­ning sources, thin­king before spre­a­ding dubious posts or fin­ding out about how algo­rithms work or the inten­ti­ons and inte­rests of the sen­ders of cer­tain messages.

AI com­pe­tence can be seen as ana­log­ous to this: Cri­ti­cally exami­ning which tools can be used in which places and which are useful and effec­tive — and which are not, or an extra dose of cau­tion is requi­red. It makes it pos­si­ble to use the new tools pro­fi­ta­bly – and it pro­tects against uncri­ti­cally buy­ing into nar­ra­ti­ves about AI. Howe­ver, it has not yet been deve­lo­ped to the same ext­ent as we think we have it in media com­pe­tence. For many, AI is still unchar­ted ter­ri­tory, and this makes us (still) sus­cep­ti­ble to sto­ries and pro­mi­ses that do not stand up to rea­lity and are sold to us partly out of tan­gi­ble finan­cial inte­rests, partly out of inex­pe­ri­ence or partly for under­stan­da­ble psy­cho­lo­gi­cal reasons. We would like to call the oppo­site atti­tude to AI com­pe­tence the AI hype.

So what con­sti­tu­tes AI com­pe­tence? From our expe­ri­ence — that is the per­spec­tive of psy­cho­lo­gi­cal rese­arch and crea­tion — there are essen­ti­ally three simple things that equip us for the pro­fes­sio­nal use of gene­ra­tive AI:

Basic knowledge of how generative AI works and the psychology of dealing with AI

Of course we have all dealt with it. We have all heard or read about the “sto­cha­stic par­rot”. Howe­ver, it is easy to lose sight of exactly what this means. The so-cal­led “hal­lu­ci­n­a­tion” in ChatGPT or the pro­duc­tion of ste­reo­ty­pes in image-gene­ra­ting pro­grams are not unwan­ted side effects that will even­tually dis­ap­pear, but a con­sti­tu­ent part of the design of gene­ra­tive AI. The out­put is never based on any kind of under­stan­ding of the task, but is always the result of a sel­ec­tion pro­cess of tokens based on pro­ba­bi­li­ties deri­ved from the trai­ning data. Wit­hout this con­trol­led (more or less varia­ble depen­ding on the “tem­pe­ra­ture”), i.e. pro­ba­bi­lity-based rand­om­ness, it would not be “gene­ra­tive” AI, the ans­wers would not be fluid and plau­si­ble and the model would be use­l­ess for our purposes

This shows the strengths and weak­ne­s­ses of using gene­ra­tive AI quite pre­cis­ely. It is a pri­ce­l­ess tool for wri­ting a social media post because the strengths of con­trol­led rand­om­ness can be fully exploi­ted here. In the field of rese­arch, this also applies (albeit to a les­ser ext­ent) to tran­scrip­tion, sum­ma­ri­zing report volu­mes and many other tasks. Howe­ver, it is less sui­ta­ble for the eva­lua­tion of an in-depth inter­view because the focus here is on under­stan­ding and not on pro­ba­bi­li­ties and appa­rent plau­si­bi­lity, and atmo­sphe­ric and inter­per­so­nal dyna­mics are more important than the texts pro­du­ced in the inter­view. It is excel­lent for pro­du­cing test mate­rial — images or claims — as long as you don’t have very spe­ci­fic and com­plex image motifs in mind and you have the image idea yourself.

It is also important to rea­lize that our evo­lu­tio­nary back­ground means that we humans are wired in such a way that we are almost auto­ma­ti­cally foo­led by this prin­ci­ple. Because we are human, we instinc­tively expe­ri­ence text pro­duc­tions by AI as “expres­si­ons“ of a quasi-living being, espe­ci­ally if they come across as flu­ent and plausible.

AI com­pe­tence then pro­tects against uncri­ti­cally “belie­ving” ever­y­thing from ChatGPT, for exam­ple. It makes us cau­tious, for exam­ple, when using the con­trol­led ran­dom prin­ci­ple as a sub­sti­tute for inter­vie­w­ing peo­ple. “Syn­the­tic data” is the key­word here. Even if the syn­the­tic “test sub­jects” inter­viewed are fed with mate­rial from real sur­veys, the prin­ci­ple of con­trol­led rand­om­ness and pro­ba­bi­lity again applies when “exten­ding” this basis. As a result, what emer­ges seems highly plau­si­ble to us, and we humans like to fall for the plau­si­ble and obvious (also because we are human and as such love meaningful sto­ries, which is what sto­rytel­ling lives on). But isn’t it exactly in rese­arch that the sur­prise and thus the excee­ding of the obvious is what makes it so exci­ting, and not actually its jus­ti­fi­ca­tion for exis­tence? In psy­cho­lo­gi­cal rese­arch in par­ti­cu­lar, the gold of know­ledge usually lies in the devia­tion and the unpre­dic­ted. It’s not about the last 5% of addi­tio­nal know­ledge, but usually about the core of the matter.

The same applies to inter­ac­tive AI ava­tars or AI per­so­nas if they are used in rese­arch as a sub­sti­tute for gene­ra­ting or enri­ching insights or for insights-based idea deve­lo­p­ment: if they are to be more than a “tal­king data­base” of exis­ting stu­dies, i.e. if they are to play to the strengths of con­trol­led rand­om­ness (e.g. through the con­nec­tion with a trai­ned LLM), they will sim­ply “enrich” the data from the trai­ning mate­rial in a plau­si­ble way (by the way, plau­si­bi­lity is also the basis of the Turing test: does what the model out­puts sound “like a human”?). You can do this, but you should be aware that it will be dif­fi­cult to judge how well they simu­late rea­lity or not, even – or espe­ci­ally – if they sound plau­si­ble. In any case, it is bet­ter not to base any important stra­te­gic decis­i­ons on them.

Part of the logic of the hype is the con­stant refe­rence to the future. “In the future” or “soon“ this or that will be pos­si­ble (AI pro­ba­bly has this in com­mon with nuclear fusion). In any case, there is curr­ently not­hing in sight to sug­gest that com­ple­tely new forms of gene­ra­tive AI will over­come the fun­da­men­tal limi­ta­ti­ons based on the prin­ci­ple of pro­ba­bi­li­ties. The impro­ve­ments in the first year were rapid. After that, things leve­led off, and it is impos­si­ble to pre­dict whe­ther the threat of AI incest (AI incre­asingly lear­ning from AI-gene­ra­ted con­tent) will inten­sify the qua­lity pro­blem that some indus­tries are curr­ently crea­ting for themselves.

Critical evaluation of the added value of generative AI from the perspective of a specific profession

If I want to assess how well gene­ra­tive AI can sup­port me in my pro­fes­sio­nal pro­ces­ses, I don’t just need to fami­lia­rize mys­elf with how AI works and its strengths and weak­ne­s­ses. It is almost more important that I am fami­liar with the pro­ces­ses that AI is sup­po­sed to replace or sup­port! That sounds banal, but it’s not. If I’m not a psy­cho­lo­gi­cal rese­ar­cher, I can’t assess how and where it can add value in psy­cho­lo­gi­cal rese­arch. If I’m not a pro­fes­sio­nal crea­tive, I can’t assess whe­ther and how AI can help me create an ad or a commercial.

This point is also not tri­vial because quite a few AI experts pro­pa­gate this, even though they under­stand not­hing about the pro­ces­ses for which they want to sell AI as a savior. For exam­ple, repre­sen­ta­ti­ves from mar­ke­ting or mar­ket rese­arch use the prin­ci­ple of con­trol­led rand­om­ness to create arbi­trary adver­ti­se­ments that they con­sider somehow chic and “crea­tive” – wit­hout kno­wing what is important in the crea­tion pro­cess. Desi­gners or film pro­du­cers usually want to pro­duce a very spe­ci­fic result, a very spe­ci­fic mood, a very spe­ci­fic sur­face com­po­si­tion, a spe­ci­fic facial expres­sion, etc., or they have design pro­blems to solve, such as fit­ting three important mes­sa­ges into a sin­gle image, which should nevert­hel­ess be coher­ent and not over­loa­ded. An adver­ti­sing medium should have an effect on the right tar­get group, make the company’s cor­po­rate design reco­gnizable, achieve stra­te­gic goals and be well recei­ved by cus­to­mers. In any case, AI exper­tise, as out­lined here, pro­tects against such over­rea­ching fan­ta­sies of omni­po­tence. We would­n’t even think of explai­ning to a che­mist how she should use AI.

Practical use of generative AI in concrete and real projects

The third and pro­ba­bly most important aspect of AI com­pe­tence is prac­tice. Just do it, and not just try it out or “play around” with it, but actually use AI in real pro­jects in your own pro­fes­sion (see point 2). Of course, you first have to test and try out a lot of things and, if neces­sary, get trai­ning in one tool or ano­ther – and given the large num­ber of tools, this can also cost money in the long run. Howe­ver, prac­ti­cal and pro­fes­sio­nal use quickly sepa­ra­tes the wheat from the chaff.

It would go bey­ond the scope of this blog post to list all the expe­ri­en­ces we have had with AI tools in rese­arch and deve­lo­p­ment. We would like to use the exam­ple of film pro­duc­tion to illus­trate this — and since we have alre­ady given our rea­ders enough text at this point and because it is cus­to­mary in the field of crea­tion to show results ins­tead of wri­ting about them, we refer to “Eddi’s Jour­ney” (switch on the sound!):

The film pres­ents some of our use cases for AI, and at the same time “shows” what is pos­si­ble with AI in film pro­duc­tion – and where “con­ven­tio­nal” pro­fes­sio­nal pro­grams deli­ver bet­ter results. Eddi hims­elf was crea­ted with a 3D pro­gram. All AI tools were unable to cope with the simple geo­me­tric shape. Pro­ces­sing the three sphe­ri­cal seg­ments that make up Eddi would require some kind of design reco­gni­tion. The Luma soft­ware, on the other hand, is unbeata­ble when it comes to mor­phing a figure. Here, the prin­ci­ple of pro­ba­bi­lity, accor­ding to which indi­vi­dual pixels fol­low one ano­ther, can create beau­tiful effects.

Other spe­cial effects could be gene­ra­ted fas­ter and bet­ter with Adobe After Effects. The music is partly AI-gene­ra­ted. From minute 01:08, Suno takes over and “extends” an own track. Some real sce­nes are AI-gene­ra­ted, for others it was easier to use a scene from a stock database.

In any case, we are opti­mi­stic that AI com­pe­tence will con­ti­nue to grow and will soon finally replace the hype phase. Not out of dis­il­lu­sionment, but on the con­trary: only with a com­pe­tent atti­tude can the won­derful pos­si­bi­li­ties of gene­ra­tive AI be fully exploi­ted and used pro­fi­ta­bly in companies. 🙂

Share Post :

weitere Beiträge