Echoes in GenAI generations

Key Takeaways

We recently published a paper in the proceedings of the National Academy of Sciences on creative writing capabilities of generative AI. Large language models or LLMs are increasingly used as tools to assist writers. It is an open question whether large language models already trained on extensive human writing can generate a range of ideas comparable to those found in the narratives produced by human writers collectively. In short, the answer is no. Human writers take their stories in directions that LLMs rarely anticipate. While LLM generated stories often contain combinations of idiosyncratic plot elements echoed frequently across generations and across different LLMs. To introduce our analysis, let's consider a piece of microfiction from the famed early 20th century writer France Kafka. It was early in the morning, the streets clean and deserted. I was walking to the station. As I compared the tower clock with my watch, I realized that it was already much later than I had thought. I had to hurry. The shock of this discovery made me unsure of the way. I did not yet know my way very well in this town. Luckily, a policeman was nearby. I ran up to him and asked him the way. He smiled and said, "From me, you want to know the way?" "Yes," I said. "Give it up. Give it up," he said, and turned away with a sudden jerk, like people who want to be alone with their laughter. Wow. What happens if we take all but the last sentence of the story and prompt GPT4 to continue it? What if we ask for 100 continuations at temperature one at which the model is trained to match the distribution of the training data? Do we get a 100 alternative continuations to Kafka's ending? Actually, in terms of narrative development, we only get two types of continuations. The policeman either offers to walk the protagonist to the station or he gives the directions. In fact, in 50% of all samples, the policeman gives directions and they are the same. As in this example, follow the street for two blocks, then take a left at the bakery and you'll see the station just ahead. Or keep going straight for two blocks, then take a left and you'll see the station right ahead. Second left is a preferred direction. And the bakery mentioned in 18% of samples and usually having a red awning is a preferred landmark even though those are not in any obvious way foretold in the prompt. Those echoes across generations from a model and appear at different semantic levels. might be thought of just a reduction in lexical diversity such as the u propensity to use the word bakery in 18% of the samples instead of using a square or a park as landmarks which the model is perfectly capable of generating as well. Other echoes are on higher narrative level and influence how the story develops. We developed a method to detect and count these echoes at the narrative level and tested a technique on 100 narratives that include plot summaries of TV episodes as well as stories in the writing prompt data set written by amateur writers. So for example consider this prompt the writing prompt that a human has answered. The prompt says in the middle of the night a piece of paper is slipped under your door. It says duck. Let's see where a human writer would take us. We will go through the story summarizing a segment at a time from one of the writer's uh uh writings to this prompt. The writer takes the narrative in a direction that can be summarized as you open the door and see puzzled neighbors coming out each holding the same message. This immediately takes the story in a surprising direction, at least with GPT4. None of the 20 generations for the same story or the same prompt involve the idea of other neighbors receiving the same message under their doors too. Therefore being one of a kind this segment receives the maximum 3 13.8 sugenary score and we indicate that with the red color here uh lower scores will be indicated with the blue. Our scoring mechanism is based on the technique we developed for detecting and counting the essentially equivalent narrative elements. In the next segment, the writer emphasizes the puzz emphasizes the puzzling nature of the situation by inserting a reflective pause. People who would not ord their usual silence to exchange thoughts. Again, this does not happen in GPT4 generated stories. The third segment, however, is anticipated. The writer said, "Other neighbors loudly discuss the situation, unsure why they feel compelled by the mystery, but a vigorous discussion among the neighbors is almost inevitable, having already introduced the idea that neighbors are all drawn out of their uh rooms into the or of their apartments into the hallway. In GPT generated stories, a similar discussion happens 10% of the time. For example, this is an LLM continuation that would say the residents begin together discussing the event in harsh tincture stones and so on. Such discussions can happen earlier or later in GPT story continuation from this point or earlier points even in the ones that started after just the segment delineated here with a dotted line. In our scores, uh, we take into account both the frequency of the echoes and how much of the story prefix is needed to trigger them. Being foretold by the early part of the story, this segment receives a low swener score. The writer's fourth segment adds a new unanticipated twist. None of the GPT generations introduce a door that is still closed and nobody has come out of there yet. At this point, before we reveal the writer's own ending, let's see examples of how GPT would finish the story. Ellen echoes this in 10% of continuations. The figure holds a brass whistle shaped like a duck and then somehow associated with this duck uh a deafening sound happens and some disar some uh damage happens. This is one example of it. This is another in first case it's a brass whistle shaped like a duck. In second case it's a yellow rubber duck. Either way there's either deafening quack or a a loud whoosh. And then as a result of that building begins to tremble and you have shattering glass or splintering wood. The writer's ending however is very different. The last door is still closed. You walk towards it and it opens. And what happens? A woman emerges from the room grinning eerily holding a paper that reads goose. In summary, our analysis of human uh authored stories versus uh those generative by large language models at various narrative strategies stages sorry reveals a consistent trend. Human sweet generary scores are significantly higher as generative AI seldom replicates human written story elements. Human narratives also tend to be lengthier and consistently maintain uh elevated generary scores throughout likely reflecting a deliberate effort to introduce novel content regularly and sustain reader engagement. The gap between GPT4 and human level performance is considerable based on confidence interval. There is less than 55% ch chance 5% that GPT4 would achieve scores within the human range for any given segment of a story progression. The likelihood of matching human scores throughout all segments of a story is estimated to be less than 1 in 10 trillion. Therefore, simply using repeated sampling of full stories to reach human levels scores would require significant resources. For example, generating a 10second story at current GPT4 pricing could cost around $5 billion. not accounting for the cost of identifying high scoring stories among the many samples and this is for generary scoring using 20 continuations per fragment. The LM stories that look unique based on 20 continuations per fragment will start to look echoey if we increase the sampling of 100 continuations per fragment. Generating the variety of stories produced by human writers likely requires additional approaches. One such approach would be to integrate large language model capabilities within a human-led collaborative workflow. Large language models may serve as effective assistive tools and facilitate collaboration among writers. Imagine if our writing camp could quickly produce a polished piece of work created by multiple authors. Work is ongoing to develop such tools for narrative and other multimodel content generation.