Current AI models may be superhuman or close to superhuman at constrained writing, such as writing a page without the letter E.

This is a hedged claim ('might be'), judged as of 2023-08-08. True: evaluations or well-documented demonstrations from around then show that frontier models such as GPT-4 or Claude 2 could reliably produce constrained text, such as a page with no letter E, at or above the level of skilled humans. Mostly true: the models performed at a level close to strong human performance, with occasional constraint violations. Mixed: performance was inconsistent, succeeding in some constrained-writing tasks and failing often in others. Mostly false: the models usually broke such constraints (for example letter-level lipograms, where tokenization made them struggle), with only isolated successes. False: the evidence shows the models clearly performed well below ordinary human ability at these tasks. Given the hedging, a plausible but unproven case leans towards mixed rather than false.

Set before the research began.

1. E1 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (authors affiliated with Oracle Corporation and University of Oregon)

   “We show that language models, which are fine-tuned even on the simple lexical constraint of omitting the letter “e”, still occasionally ignore their constraints. In fact, even when these models are overtrained to an absurd degree, complete adherence to these constraints is unlikely.”

2. E2 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (authors affiliated with Oracle Corporation and University of Oregon)

   “We find that most language models generate compelling text even under significant constraints.”

3. E3 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (authors affiliated with Oracle Corporation and University of Oregon)

   “Among other things, we observe that language models can remain coherent even with a remarkable amount of filters applied to their vocabulary.”

4. E4 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (Allen Roush, Sanjay Basu, Akshay Moorthy, Dmitry Dubovoy; Oracle Corporation and University of Oregon)

   “We show that language models, which are fine-tuned even on the simple lexical constraint of omitting the letter “e”, still occasionally ignore their constraints.”

5. E5 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (Allen Roush, Sanjay Basu, Akshay Moorthy, Dmitry Dubovoy; Oracle Corporation and University of Oregon)

   “applying the constraint to ban the letter “e” from the vocabulary of these models results in both improved perplexity, as well as zero ignored constraint errors.”

6. E6 Most Language Models can be Poets too: An AI Writing Assistant and Constrained Text Generation Studio arXiv (Allen Roush, Sanjay Basu, Akshay Moorthy, Dmitry Dubovoy; Oracle Corporation and University of Oregon)

   “We find that most language models generate compelling text even under significant constraints.”

7. E7 The human-authorship halo: attribution bias in literary style evaluation by humans and AI (arXiv 2510.08831) · AI research paper excerpts and notes inquiringlines.com (summary of arXiv 2510.08831) · 9 October 2025

   “GPT-4’s ‘Lipogram’ contained 9 instances of the forbidden letter, a failure that aligns with well-documented limitations of LLMs in character-level tasks [61, 62].”

8. E8 Evaluating, Understanding, and Improving Constrained Text Generation for Large Language Models arXiv (academic paper by external researchers)

   “In the context of Order constraint, Table 3 shows that GPT-4 achieves an accuracy of 92.30%, and GPT-3.5 achieves a competitive performance of 91.40%.”

9. E9 Evaluating, Understanding, and Improving Constrained Text Generation for Large Language Models arXiv (academic paper by external researchers)

   “In summary, we contend that the incorporation of relation constraints remains notably challenging for current state-of-the-art LLMs.”

The only direct evidence on a frontier model's lipogram performance is a single GPT-4 output reported in a 2025 study [E8], and the Oracle/Oregon paper reports only occasional constraint violations by unnamed fine-tuned models [E1, E5]; this does not show that models around 2023-08-08 usually broke such constraints. Other items show compelling constrained output and high GPT-4 accuracy on some constraint types [E2, E9], so the evidence does not settle the claim under the criteria.