Week 9 - Visual Analysis Task and AI Training Sets

Visual Analysis Task and AI Training Sets

Briefing on Visual Analysis Task

The visual analysis task involves applying the tools and concepts learned in the first part of the semester (ideology, representation, semiotics, social semiotics) to an image to respond to one of the provided questions. The due date has been extended to June 4. The task is a written assignment of 1,200 words (+/- 10%).

Task Details

  • Word Count: 1,200 words (+/- 10%)

  • Due Date: June 4

Questions to Choose From:

  1. Visual texts offer superficial and simplistic reflections of issues in our society.

  2. Even the most mundane forms of visual communication are implicated in the creation and circulation of ideology.

  3. Images have the opportunity to contest and overturn dominant discourses.

  4. All visual information shared on social media is disinformation.

Pick one statement to engage with and formulate an argument about whether you agree or disagree. The key terms associated with each question should guide your choice (e.g., discourses for question 3, ideology for question 2, etc.). Use analytical tools from the unit, such as semiotics and representation, to support your argument.

Image Selection

Choose one of the provided images or find your own with the instructor's approval. Images include memes around climate activism, examples of disinformation, and AI-generated images. To use your own image, get approval via the forum.

Research Requirements

  • Engage with a minimum of four required or recommended readings.

  • Include one academic, peer-reviewed article found independently from media journals.

  • Use APA 7th style referencing.

Assessment Criteria

The assessment criteria are the same as the presentation, focusing on:

  1. Responding to the chosen prompt.

  2. Referencing.

  3. Analysis

  4. Argument

Tutorial Survey

  1. A questionnaire will be available on iLearn to determine what students want to practice before completing the visual analysis.

  2. Based on the survey results, tutorial time will be used to address areas such as visual analysis structure, evaluating paragraphs, brainstorming, and APA referencing.

Support Resources

The university offers resources such as:

  • Ask (student study coaches)

  • StudyWise

  • Writing Center

  • English as a Second Language resources

  • Studio City (for grammar and writing feedback)

Recap of Visual Mis/Disinformation

Last week's lecture focused on visual mis/disinformation.

  • Visual text that has been manipulated is classified as visual disinformation, especially when there is intent.

  • There is a research gap in visual disinformation compared to textual mis/disinformation.

  • Visual content has a bigger reach, is stickier, and can be more emotive and persuasive.

Peng's Framework for Analyzing Visual Disinformation:

  1. Format: Different formats (photographs, memes, data visualizations) have different attributes.

  2. Features: Consider perceived features such as credibility, vividness, and aesthetic appeal.

  3. Meaning: Analyze how visuals make meaning.

AI-Generated Images

AI-generated images are proliferating, used in advertising, journalism, and more. Examples include:

  • Shutterstock AI-generated images.

  • Liberal Party's AI-generated electoral ads.

  • Coca-Cola's AI-generated advertising (criticized as soulless).

  • Michael Christopher Brown's AI-generated photojournalism.

  • AI optimization of images in journalism.

Case Study: Georgie Purcell Image

An image of Georgie Purcell was optimized by AI in a way that added a midriff and fuller bust, reinforcing biases present in society.

AI Training Sets

AI image generators are not neutral; they draw on training sets that can be highly ideological, biased, and informed by dominant discourses around gender, race, and sexuality. When AI generates images it is drawing from stock images and videos. Therefore:

  • If original images reproduce biases, those biases are amplified in the AI-generated image.

  • Initial assumptions that AI would be more objective have proven incorrect.

Crawford and Paglen on AI and Data

Despite the common mythos that AI and the data it draws on are objectively classifying the world, Crawford and Paglen argue that "everywhere, there is politics, ideology, prejudices, and all of the subjective stuff of history."

AI Training Set Layers

  1. Taxonomy of overarching classifications and categories.

  2. Subcategories (e.g., apple, apple butter, apple).

  3. Individual images labeled with judgments (e.g., good apple, bad apple, suspicious).

ImageNet Example

ImageNet contains more than 14,000,000 labeled images. There are 20,000 categories and 2,833 subcategories under the category of "person":

  • Categories include race, nationality, profession, economics, behavior, character, and morality.

  • Examples of categories: bad person, call girl, drug addict, closet queen etc. These are political.

ImageNet→PoliticsImageNet \rightarrow Politics

Problems with Image Sets

AI training sets are problematic because the data used to train AI tools is not neutral, natural, or apolitical. Images are collected, categorized, and generated based on dominant discourses around gender, race, class, and politics. These image training sets reinforce dominant colonial and Western ways of understanding the world.

AI and Ideological Archives

AI images can be read as ideological archives, showing dominant discourses and ideologies around race and gender that are reproduced in visual culture.

Recap of Key Ideas

  • The relationship between representations and reality is ambiguous.

  • Images are interpreted within social and cultural contexts (constructivist model).

  • The relationship between a signifier and signified is arbitrary.

  • Discourses are the socially organized processes of talking about a particular subject matter.

  • Ideologies are widely shared social assumptions about the way things are and should be.

Examples of AI Image Generation

  • AI has generated the campaign for Mango but used unrealistic bodies. These images are not representative but the product of a dominant, imagined, mostly westernized esthetic.

  • Nature investigated AI imagery by feeding in a number of prompts. The prompt for the image generation was: black African doctor is helping poor and sick white children, photojournalism. This is an example of the discourse of the white savior going to Africa in order to assist children.

  • Amnesty International have also been critiqued for drawing on AI or using AI to generate their campaign materials.

Conclusion

It is necessary to study visual disinformation and the way AI fits within that discussion. It is equally important to move away from seeing AI as objective or scientific, and recognize it as highly political and ideological. AI amplifies the sorts of visual violence, as well as the dominant esthetics, present in our culture. In conclusion:

AI≠ObjectiveAI \neq Objective
AI→HighlyPoliticalAI \rightarrow HighlyPolitical

Summarisation:

Visual Analysis Task and AI Training Sets
Briefing on Visual Analysis Task

The visual analysis task applies course concepts (ideology, representation, semiotics, social semiotics) to an image, responding to a provided question. The due date is June 4, and the assignment should be 1,200 words (+/- 10%).

Task Details

  • Word Count: 1,200 words (+/- 10%)

  • Due Date: June 4

Questions to Choose From:

Pick one statement to engage with, arguing whether you agree or disagree. Use key terms and analytical tools (e.g., semiotics, representation) to support your argument.

Image Selection

Choose a provided image or find your own with instructor approval.

Research Requirements

  • Engage with a minimum of four required or recommended readings.

  • Include one academic, peer-reviewed article from media journals.

  • Use APA 7th style referencing.

Assessment Criteria

Focus on:

  1. Responding to the chosen prompt.

  2. Referencing.

  3. Analysis

  4. Argument

Tutorial Survey

A questionnaire on iLearn will determine student practice needs (visual analysis structure, paragraphs, brainstorming, APA referencing).

Support Resources

The university offers resources like Ask, StudyWise, Writing Center, ESL resources, and Studio City.

Recap of Visual Mis/Disinformation

Last week's lecture covered visual mis/disinformation.

  • Visual disinformation involves manipulated visual text, especially with intent.

  • Research on visual disinformation is limited compared to textual disinformation.

  • Visual content has a broader reach and is more emotive and persuasive.

Peng's Framework for Analyzing Visual Disinformation:

  1. Format: Different formats have different attributes.

  2. Features: Consider credibility, vividness, and aesthetic appeal.

  3. Meaning: Analyze how visuals make meaning.

AI-Generated Images

AI-generated images are widely used in advertising, journalism, etc.

Case Study: Georgie Purcell Image

An AI-optimized image of Georgie Purcell reinforced societal biases.

AI Training Sets

AI image generators are biased, drawing on ideological training sets. If original images reproduce biases, AI amplifies them. Initial objectivity assumptions were incorrect.

Crawford and Paglen on AI and Data

AI and its data are not objective; they contain politics, ideology, and prejudices.

AI Training Set Layers

  1. Taxonomy of overarching classifications.

  2. Subcategories.

  3. Individual images labeled with judgments.

ImageNet Example

ImageNet contains over 14,000,000 labeled images with categories including race, nationality, and morality.

ImageNet→PoliticsImageNet \rightarrow Politics

Problems with Image Sets

AI training sets are not neutral, natural, or apolitical, reinforcing colonial and Western understandings.

AI and Ideological Archives

AI images reflect dominant discourses and ideologies around race and gender.

Recap of Key Ideas

  • Representations and reality are ambiguous.

  • Images are interpreted within social and cultural contexts.

  • Signifier and signified relationship is arbitrary.

  • Discourses are socially organized processes of talking about a subject.

  • Ideologies are widely shared social assumptions.

Examples of AI Image Generation

  • AI-generated Mango campaign used unrealistic bodies.

  • Nature's AI imagery test highlighted the "white savior" discourse.

  • Amnesty International was critiqued for using AI in campaign materials.

Conclusion

Study visual disinformation and AI, recognizing AI as political and ideological, amplifying visual violence and dominant aesthetics.

AI≠ObjectiveAI \neq Objective

AI→HighlyPoliticalAI \rightarrow HighlyPolitical