Generate an art review

The character and quality of the results generated by this model depend greatly on what kind of prompt it's given. Type your submission in English below and the generated results will be emailed to you. They have to be emailed because the generation process can take a while (2-15 minutes). Check the tips below for more direction.

The review generator is currently paused. To see examples of what thousands of people created, visit the Examples page.

Tips for creating prompts:

• Results are emailed so if you put in a fake address you won't see anything.
• Submissions must be in English.
• Use complete sentences that are grammatically correct for best results. Submitting a bunch of unconnected words will yield generic results or defaults. See the Examples page for uh...examples.
• For comparison, five different results will be generated based on the prompt you enter. They can vary wildly.
• The model was trained in language written mostly in the third person, so sticking to that will generate results consistent with the original reviews. Instead of saying "I made a bunch of painting are they good?", try "John Doe made a bunch of paintings and they are ". Note the unfinished sentence. The model will try to finish the sentence.
• Proper names and pronouns may get used if included in the prompt.
• As a sociological tool, you can input prompts that reveal bias and prejudice in the original training material and in the GPT-2 model itself. Try changing art adjectives to identity labels. For example, change "abstract painter" to "Black painter" or "woman sculptor". The results may surprise you.
• The model will occasionally glitch in interesting ways, getting caught in Möbius loops of language structure. You can trigger that intentionally with bad grammar and playing with object/predicate sequencing.
• The model will occassionally "name drop" artists and authors. It is doing that because actual art reviewers do that and not because the model is making any actual connection to the work or identity of that person.
• This tool does not understand anything it generates and is capable of offensive content if given offensive language as an input. If it is discovered that people are intententionally generating text in this way, they will be blocked.