AI Hallucination 

A Collaborative Inquiry with Meghan Duval

1.     Guiding Question: Both Jiao and Duval are first year students in the Art and Art Education Program, and both have a background in producing artwork. Additionally, Duval has an art consulting firm for which she needs to source large quantities of cost-efficient artwork.  In looking to generative AI as a source of art production, its visual hallucinations are a major challenge to effectiveness. Also, Jiao’s research interests include the creative use of AI tools and assessments of generative artworks. Thus, our collaboration focuses on looking at the problem of AI visual hallucinations

2.     Goals and Objectives: Our main objective is to assess how effective generative AI is at producing artwork that is of high enough quality to be used in our respective professional fields and to see how effective AI is at making this assessment itself. First, we intend to experiment with multiple text to image AI generators to visually assess which are less prone to visual hallucinations and to generate a list of common visual hallucinations to look out for in image production as they are frequently not immediately obvious. Secondly, we will submit the images that we have generated into multiple text to text AI applications and ask them to provide a brief description of the artwork for potential use as a wall label or presentation description. This second step is an opportunity to ask AI to identify and address visual hallucinations thus completing the workflow between the two AI generators. 

3.     Plan:

a.     Develop a prompt for text to image generator that Jiao or Duval would regularly use in the course of their art production

b.     Feed the prompt into a minimum of four AI image generators: Midjourney, Adobe, Canva and DALL-E3.

c.     Do a personal visual assessment of the visual hallucinations and make a list of the hallucinations.

d.     Repeat steps a, b and c with a different, perhaps “more challenging” prompt from Jiao and Duval’s experience with AI.

e.     Confirm or add to the list of visual hallucinations

f.      Feed the 2 sets of images into at least two text to text AI generators ChatGPT and Claude and ask them to write descriptive wall labels.

g.     Ask the language generators to specifically address any hallucinations.

h.     We will take one week to generate the images and feed the images to the text

i.      We will spend the second week on assessments and any follow up questions

4.     Deliverables:

a.     A photo essay showing the two sets of images with their descriptive wall label texts

b.     A list of common visual hallucinations

c.     Assessment of quality of image generators as pertains to hallucination issue only

d.     Assessment of quality of text analysis generators as pertains to their ability to assess visual hallucinations

AI Hallucination Analysis by Chat GPT:

1. Creeping Devil Cactus Missing: Stenocereus eruca, which grows horizontally along the ground, is absent or inaccurately depicted.

2. Palm Trees: The presence of tropical-looking palms is ecologically inconsistent with the desert environment.

3. Prickly Pear Cactus (Opuntia): These are not visible or are underrepresented despite being requested.

4. Plant Proportions: Some agave plants and shrubs appear oversized relative to others.

5. Vegetation Mix: The blend of flora seems unnatural, with mismatched plant species not native to the same desert region.

AI Hallucination Analysis by ChatGPT:

1. Creeping Devil Cactus Misrepresentation: The large, barrel-like cacti shown do not resemble Stenocereus eruca, which grows horizontally along the ground.

2. Prickly Pear Cactus Missing: There is no visible representation of Opuntia (prickly pear cactus), which was requested.

3. Saguaro Overrepresentation: The saguaros are accurately present but dominate the background unrealistically in a dense pattern.

4. Cactus Density: The arrangement of cacti is overly uniform and artificial, not resembling natural desert gardens.

5. Pathways: The pathways and fencing appear unnatural and overly structured for a desert setting.

AI Hallucination Analysis by ChatGPT:

1. Plant Misidentification: The plant depicted does not resemble Stenocereus eruca (creeping devil cactus), which has long, horizontal, sprawling growth patterns along the ground.

2. Incorrect Form: The plant shown has upright, aloe-like features, inconsistent with the trailing, prostrate nature of Stenocereus eruca.

3. Inappropriate Environment: The coastal background is mismatched with the desert habitat where Stenocereus eruca naturally thrives.

4. Overly Clean Lines: The plant lacks the rugged, spiny, and weathered texture typical of Stenocereus species.

Analysis by Chat GPT:

The image appears to depict a cozy, rustic kitchen interior with blue cabinetry, wooden countertops, and a farmhouse sink. While the composition is aesthetically pleasing, it is important to scrutinize details to ensure there are no visual inconsistencies or signs of image synthesis (hallucinations). Here’s a detailed analysis:

1. Lighting and Shadows:

   • The natural light seems to come from the left window. Check whether shadows cast by objects (e.g., the plants, faucet, and cabinet) align consistently with this light source.

   • Ensure reflections, particularly on metallic 

   • surfaces like the faucet or handles, match the lighting conditions.

2. Material Consistency:

   • Look at the wood grain on the countertops and the cutting boards. Ensure the textures are realistic and align seamlessly without appearing overly smooth or artificial.

   • The ceramics on the shelves should have consistent glazing effects, especially where light interacts with them.

3. Proportions and Perspective:

   • Verify that all objects are proportionate to one another. For example, the scale of the oranges, bowls, and potted plants should match the kitchen's dimensions.

   • Ensure the perspective of the cabinets and tiles aligns with the overall vanishing point.

4. Repeating Patterns or Anomalies:

   • Check for repeating or duplicated patterns, which are common in AI-generated imagery. For example, inspect the tiles or objects like bowls for unnatural repetition.

   • Analyze edges of objects, such as where the cabinetry meets the walls, for inconsistencies or blending issues.

5. Object Placement and Gravity:

   • The objects, such as the plates and cups, should appear securely placed. Watch for items that seem to defy gravity or float subtly.

6. Details in Vegetation:

   • The potted herbs near the sink should have natural variation in their leaves and stems. Uniformity or "plastic" texture could indicate a flaw.

Analysis by ChatGPT:

Selected results from DALL-E 3:

Full chat:https://chatgpt.com/share/678a65cc-b904-800a-96cf-1a47c5c77fc3

prompt: I'm asking for an image of tourists visiting a desert garden with pickle pear cactus, saguaros, and Stenocereus erucae (creeping devil cactus) . There are hallucinations and errors in this image. Can you list them?

See final results section above.

prompt: Imagine you are going to exhibit this image at an art exhibition. Write a short wall text of it.

See final results sectino above.

ChatGPT's analysis of Meghan's list and Cece's list : https://chatgpt.com/share/67899085-913c-800a-92cc-4f3fb39f0396