It matters because we die
AI and media creation
I have experimented a lot with using AI in writing. Reviewing my papers and essays. Suggesting revisions. Tried to have it write arguments and stories from simple prompts. It is an extremely helpful tool. But, at least for now – and maybe forever – it cannot replace the writing of a human being.
AI is a great tool for reviewing one’s writing, particularly if you have less practice, are less talented, or, as in my case, write in a language you’re not native to. I like asking for alternative words and phrasings for sentences I’m not happy with, to help me improve rhythm, clarity, brevity, and variety. It’s like a “super thesaurus” that spits out infinite variants until I find the one that suits my fancy. Your friendly bot never tires of providing options and helpful recommendations, much beyond the point that would exhaust the patience of any human editor.
In the first few weeks after I got into generative AI, I would take a lot of its phrasings verbatim, feeling that it improved my writing. But as I look back at those texts, some of the AI-assisted passages stick out to me. It makes me unhappy to remember that certain turns of phrase did not come from my mind, but rather from GPT’s – even when they are good; particularly when they are good…
So, I try not to take too much. I cycle through options until I find one that sounds like something I could have come up with, that I could hear myself say or think; even then, I often use only parts of a sentence it suggests and modify the rest. And I certainly do not have it write full paragraphs. This balance works for me right now, but I understand and appreciate the purism of refusing to take anything from the machine.
Beyond line editing, AI is great for endlessly discussing and analyzing the finished product, from any conceivable angle. For me, writing is mainly about sharing my ideas with others, with the hope of sparking additional discussions. A follow-up conversation or online comment exchange makes the time and effort I poured into a piece of writing worthwhile. Unfortunately, writing when one is not a well-established author tends to result in limited engagement. Relatives and good friends will read your stuff and engage for a while, but you don’t want to exhaust their patience with continuous discussion of your essay or story. Anonymous readers on the internet have so many options that the likelihood they’ll actually engage, comment thoughtfully, do more than give an up-vote on any random piece of writing is vanishingly low – be it Reddit, Substack, or any other platform. I know because I’m like that too. Sometimes I have short bursts of trying to be the change I’d like to see in the world and engage with others’ writing in the way I wish others would for me. But the lack of reciprocity quickly curtails my motivation, and I go back to quiet lurking. AI never tires of commenting on your story or essay. And the comments are clever, if sycophantic. This greatly enhances the replay value of one’s writing. AI praise should not be taken as a sign of true merit in our writing. But it feels good to have an outlet that endlessly engages with your ideas, discusses their implications and subtleties… it feels good to have the conversations you wish you could have with another human.
I have experimented with creating full AI stories from simple prompts as well as from detailed outlines. Like others, I find the results disappointing. LLMs have no passions or desires to share with the world. When prompts are simple, outputs tend to be vanilla variants of the idea. When they are more specific, chatbots may venture into less-well-trodden territory. But, even then, the outputs generally stay close to the prompt, with little originality. And the longer one tries to extend the story, the more the repetition sets in – unless one intervenes, taking back the reins of the creative process. It also doesn’t take much for bots to forget details due to their memory limitations. And, all along, the prose has all the tells of AI writing. When brainstorming, it is possible to get AI to suggest truly bonkers ideas – I liked when it once suggested to me, “man cannot escape the gravitational field of his lover… literally.” But even when starting from its own wild premises, AI fiction quickly defaults to the same uncreative patterns as when starting from human prompts.
AI writing will likely overcome these limitations one day. Models trained to focus on creative writing may be able to innovate within their stories. Larger context windows would allow models to remember plot details and maintain immersion. Maybe one day I won’t be able to tell if a story is AI or human-made. But even then, I bet that knowing something was written by AI would still make it less interesting – even if the prose is fluid, the examples vivid, the metaphors biting… Because AI is cheap. Outputs can be tweaked again and again, in search of the variant that hits just the right notes. But this unlimited potential for optimization devalues art. No version is truly special if another one is just a click away.
We smile, we cry, we bleed, we die. Our time on earth is limited. Da Vinci won’t be back, nor will Shakespeare. Our stories embed our finite time and ideas, the visions we want to share with the world. You cannot rewind time and have Dostoyevsky change Raskolnikov’s fate. We are finite, and what we do is scarce, valuable. I suspect that in a world of infinite AI content, even if it is no longer “slop,” we will still cherish connecting to each other through reading and writing. Even when human content is less technically proficient or intellectually sophisticated. It’s in our nature to care about other people, not infinite content slot machines.
…………………………..
AI writing is more useful for learning. LLMs are great at making content accessible and can be very accurate when there are many reliable sources to draw from; still, never forget to ask for citations if it matters to be right!
On the other hand, interacting with LLMs feels unsatisfying when it comes to matters of opinion, be it politics, economics, philosophy, ethics, or other social matters. The initial answer to a normative prompt often sounds well-reasoned. But the illusion that we are interacting with something that can think is shattered by two weaknesses: the sensitivity of answers to small changes in the prompt and the extreme sycophancy when the user pushes back.
Answers that can change significantly upon small tweaks to prompts feel random, diminishing their normative weight. I realize this issue can be mitigated by reducing the “temperature” of models – the degree of randomness LLMs use in providing their answers. But while this works when repeating the exact same prompt under exactly the same conditions, my experience is that small changes to prompts often lead to wildly different answers, even with temperature set to zero. A model that answered more stably would be more interesting to interact with, like it had a personality. There’s already an inkling of this with the differences between ChatGPT, Gemini, Claude, and Grok, but I’m imagining more consistent and distinctive sets of views. Such views should not be sharp just for entertainment value. Reality is complicated, and often the most rational take on an issue is nuanced and acknowledges uncertainty. But that is very different from views that change based on where a comma is placed in the chat prompt.
Relating to this is the problem of sycophancy. It’s uninteresting to engage in intellectual discussions with an entity that changes its opinions to match ours upon minimal pushback. Ideally, AIs would start by providing the best answer given their training data, and it would be challenging, but not impossible, to persuade them to change their minds on a topic. The AI should be able to reflect on the user’s arguments and concede when they are truly superior to its initial position. But it should not constantly change its views on a dime.
These limitations do not mean that AI models are not useful when considering normative questions. They are a good sounding board for considering different perspectives and stress-testing one’s views. But the infinite adaptability of current models means they cannot offer satisfying closure to any argument.
…………………………..
Similar potential and challenges apply in using AI to generate video. As an amateur storyteller, grappling with my first attempts at fiction, I dream of turning my stories into film. It is nice to dream of real actors and a director making a movie out of my machinations. I know that will never happen. Good AI video generation is the only plausible way any of the stories in my mind will ever become visual.
But just like for writing, AI’s seamless infinity devalues the visual media it creates. If any schmuck with a 50-page Word file can churn out a movie, then the market will be oversaturated. Some will be good, most will not. And the gems may be drowned by oceans of slop.
Also, I think we’ll continue to crave seeing real people act. You can make a million movies with Brad Pitt’s likenesses. Tweak his words and fighting stances. But there is only one Brad Pitt, and the movies he chooses to use his talents in will continue to matter.
I can imagine ways to make these artificial creations more valuable. One is to make an AI mind scarce by making it unique. Perhaps only one AI is made with certain specs. There is already a flavor of this in the personalities that an LLM adopts within a context window, which disappear when one deletes the conversation. But this scarcity would be artificial – one could always program another AI bot with the same specs, and the memory of an AI bot can be saved and implanted into another one. Not the same with us beings of flesh and bone.
Another approach will be through human curation or the wisdom of the crowds. Some AI media will be genuinely good and climb the ranks in recommendation systems. Looking at my YouTube feed, this mechanism does not seem promising. But maybe better algorithms will identify quality within the infinity of options. And some humans may take the time and energy to find what is worth watching among the torrents of slop – a job I do not envy.
Maybe the younger generations will just feel differently. Growing up in the midst of infinite slop, constant scrolling, flashy dopamine mainlined into their brains, may dull the appreciation of each other’s expression. I hope not. Media is a means for connection, and connection means two, not one and a silicon mind that never sleeps.
(The writing of this essay was assisted by Grammarly. I discussed it along the way with Claude, Grok, and Gemini, but mostly did not take their advice.)