lundi 3 juillet 2023

Feeding the Beast

Image de couverture de l’article

There is a scene from the acclaimed animated series BoJack Horseman that always stuck in my mind. The media mogul Stefani Stilton repeatedly pushes the journalist Diane Nguyen to churn out new articles for "Girl Croosh", a viral-pop-feminist website à la Jezebel or Bustle. She screams on the phone something along the lines of “Give me my juicy content Diane!” like a demented, voracious beast.

Stilton is a hilarious parody of 2010s media owners: the ones who went for trashy, attention-grabbing titles, suddenly “pivoted to video” or favoured silly personality tests to drive up engagement. Put differently, the ones who rode the algorithmic wave to optimise their visibility on search engines and social media; and the ones who are now in dire straits, if not already dead. 

It’s easy to see the grotesque in Stefani Stilton’s yearning for fresh internet fodder, but as they say: “hate the game, not the players”. And these media were top-notch players we (myself included) all admired, in a game alas rigged with complex rules and fickle arbiters.

One way to see viral journalism’s immense cultural influence is to ask ChatGPT to come up with its own funny list articles (or listicles) — for example, “15 Butterlicious Reasons to have a Butter-only diet” (I just prompted this one, which includes “reason 8: Skin Health: Butter makes your skin as smooth as, well, butter. Say goodbye to dull skin and hello to a buttery glow that will make heads turn.”)

The results are impressive and sometimes even funnier than original human-made listicles. They show how this particular format is codified and easy to duplicate ad nauseam. And in a way, it’s a testament to how mediocre listicles have come to be. Don’t get me wrong: the original idea was genius and many articles are truly hilarious and thoughtful, but most of them are now pure mush destined for Search Engine Optimisation. In a recent interview with the Financial Times, the Sci-Fi writer Ted Chiang recognised we may have put the bar too low with Web journalism and copywriting: “AI-generated text is not delightful, but it could perhaps be useful in those certain areas (…) but that’s not exactly an endorsement of their abilities (…) more a statement about how much bullshit we are required to generate and deal with daily.”

In other words, Generative A.I. tools are excellent at producing flavourless stuff we take for excellent, because our expectations are a reflection of the content we train them on. Like insatiable beasts, they swallow billions of crappy web pages and promptly regurgitate them. The formula you are what you eat applies to A.I. — if you want to influence its production, you must work on its diet.

YOU ARE WHAT YOU EAT

Now, there are two ways to look at the diet we give to A.I. tools.

The first one is resistance, with a restrictive diet. Robots can have basic content — administrative lingo, recipes, minutes from meetings, weather reports — but proper literature or art should remain the prerogative of humans. Expect some creators or publishers to forbid A.I. companies from training their tools on their work to prevent their know-how from being absorbed, metabolised, and ultimately made redundant by technology. Some countries will make it easy to opt out from A.I. scraping, while others will make it even more complex than it already is, for instance by considering copyright protection does not fully apply to content used to train A.I. However, resistance will always be possible for authors and artists. Offline and deliberately disconnected media could even enjoy a revival among them.

The opposite approach is upgrading the diet, making it richer and more diverse. Because A.I. Copywriting and decision-making is soon to be inevitable, you want it to be built on the best training set of data. Therefore, it is about feeding even more content to A.I. models to teach them subtlety, quirkiness and vulnerability, Humour and Beauty and Justice. Granted, that approach can sound depressing — enslaved creators catering to the monster they unleashed years ago when they first tried to woe the algorithms, like a modern-day version of punishment from the Gods. However, it also means we could all have a greater role to play when we create content to share on the Internet.

Every now and then, when I start thinking even the slightest bit about posting something online, a deep feeling of absurdity strikes me. « A quoi bon ? », “What’s the point?” is the question resonating in my head and keeping me away from writing 95% of the time. Indeed, what is the point of sharing some thought that statistically must have been already articulated 10 times and 10 times better?  What is the point in adding to the online cacophony? Oftentimes the Internet feels like an open-air market with only sellers yelling at each other without a single customer in sight. What’s the point in putting so much effort into words with so little impact? 

I guess the same thing happens to anyone who wants to publish a photo, edit a video, or work on some code. It can feel overwhelming and useless.However, tomorrow, with A.I. scraping the Web to learn our ways, every word, every picture will count. Human-Generated Content will become twice as precious — as a testament to our humanity, and as learning material to teach A.I. something other than age-old clichés, corporate blurbs and bland stock images. If the Beast’s appetite won’t be tamed anytime soon, we still have agency to influence it for the better.