Photography and AI · Module 7 · Lesson 1

Deepfakes, Deception, and the Broken Trust of the Image

When a photograph can no longer be trusted as evidence, what does that cost us?

The World Press Photo Foundation announced its 2023 Digital Storytelling winner: a series by Ukrainian photographer Maxim Dondyuk documenting the war in Ukraine. But the surrounding conversation was dominated by a different image — an AI-generated submission by Boris Eldagsen titled "PSEUDOMNESIA: The Electrician," which the jury awarded first prize in the Creative category. Eldagsen then refused the prize, publicly declaring he had submitted the AI image as a test to see whether the contest was "ready for AI." The resulting debate divided the photographic world and forced every major competition to urgently draft AI disclosure policies.

It was not a scandal of malice. It was a scandal of category collapse — the question of what a photograph is had been left unanswered for too long, and AI forced the reckoning.

The Photograph as Witness

Photography's authority has always rested on a single claim: the image was produced by light reflected from a real scene, captured by a camera at a specific moment. This indexical relationship — the idea that a photograph is a physical trace of reality, like a footprint in mud — is why photographs carry legal weight in courtrooms, why they move us to donate to disaster relief, and why photojournalism has shaped geopolitical outcomes from the execution of Nguyễn Văn Lém in 1968 to Abu Ghraib in 2004.

Generative AI breaks the indexical bond. A model like Midjourney or Stable Diffusion produces images by learning statistical patterns across hundreds of millions of photographs and then synthesising new pixel arrangements that match a text prompt. No camera. No light. No moment. The resulting image may be indistinguishable from a documentary photograph, but it is not one. It is a hallucinated average of photographic style.

The danger is not simply that AI images exist. Composite and manipulated images have existed since the 1860s. The danger is scale and accessibility: any person with a free account can now produce photorealistic images of events that never occurred, people who never posed, and atrocities that were never committed — in seconds, at zero marginal cost.

Real Cases: When Fake Images Caused Real Harm

March 2023 — The Pentagon explosion hoax. An AI-generated image depicting an explosion near the Pentagon was shared on Twitter (now X) by accounts including one verified as Bloomberg. The image briefly triggered a dip in U.S. stock markets before the Arlington County Fire Department confirmed no explosion had occurred. The image was detectable as AI-generated on close inspection — the architectural details were inconsistent — but it spread before verification could catch up.

February 2024 — Taylor Swift NCII. Non-consensual intimate AI-generated images of Taylor Swift were shared millions of times on X before the platform took action. The incident accelerated legislation in the United States; multiple states passed laws specifically criminalising AI-generated non-consensual intimate imagery (NCII) within months. Microsoft, whose Designer tool was used to generate some images, subsequently tightened its content filters.

September 2023 — Balenciaga "refugee couture." A viral set of AI-generated images depicted refugees wearing high-fashion Balenciaga clothing in devastated environments. The images were clearly labelled as AI art by their creator, but once stripped of context on social media, they were shared by thousands of accounts as commentary on real luxury brands, raising questions about dignity, representation, and the aestheticisation of suffering without consent from those being depicted.

KEY CONCEPT — INDEXICALITY

Semiotician Charles Sanders Peirce distinguished between icons (images that resemble), symbols (arbitrary signs), and indices (signs causally connected to what they represent). A thermometer is an index of temperature; smoke is an index of fire; a photograph is an index of a real scene. AI-generated images are icons — they resemble photographs — but they are not indices. Ethical photography practice in the AI era requires clearly communicating this distinction to audiences.

Detection, Provenance, and the C2PA Standard

The Content Authenticity Initiative (CAI), co-founded by Adobe, the BBC, and The New York Times in 2019, developed the Coalition for Content Provenance and Authenticity (C2PA) standard. C2PA embeds cryptographically signed metadata into image files, creating a tamper-evident record of the image's origin, edits, and authorship — a "nutrition label" for media.

In 2023, Leica became the first camera manufacturer to embed C2PA credentials in hardware, launching the M11-P. Canon and Nikon announced similar programmes. Adobe Firefly, a generative AI tool, automatically attaches Content Credentials to AI-generated images. The standard is technically sound but faces an adoption problem: credentials are stripped when images are uploaded to most social platforms, and there is no legal requirement to apply them.

AI detection software — tools like Hive Moderation, AI or Not, and Illuminarty — attempts to identify generated images by statistical signatures. Accuracy rates above 90% are reported in controlled tests, but adversarial techniques (lightly post-processing AI images) can defeat most detectors. Detection is a useful layer but not a reliable solution on its own.

PHOTOGRAPHER'S ETHICAL ANCHOR

The National Press Photographers Association (NPPA) updated its Code of Ethics in 2023 to state that members must "clearly label" AI-generated or AI-altered images and must never use AI to fabricate news events. The World Press Photo contest now requires entrants to disclose any use of generative AI, and submissions found to contain undisclosed AI generation are disqualified. As a photographer working with AI tools, voluntary adoption of these standards before they are legally mandated is a mark of professional integrity.

Lesson 1 Quiz

3 questions — free, untracked, retake anytime.

Boris Eldagsen's "PSEUDOMNESIA: The Electrician" sparked debate at World Press Photo 2023 primarily because he did what after winning?

✓ Correct — ✓ Correct. Eldagsen publicly stated he submitted the AI image to expose whether competitions were prepared for AI-generated entries, then declined the award.

✗ Not quite. Eldagsen refused the prize after revealing the winning image was AI-generated, framing it as a test of the contest's readiness.

In semiotics, what does it mean to call a traditional photograph an "index"?

✓ Correct — ✓ Correct. Peirce's concept of indexicality means the sign (the photograph) is causally linked to what it represents — the scene emitted light that formed the image. AI images lack this causal link.

✗ Incorrect. Indexicality refers to a causal, physical connection: light from the actual scene formed the image. AI images are icons (they resemble photographs) but not indices.

What is the C2PA standard designed to do?

✓ Correct — ✓ Correct. C2PA creates tamper-evident "Content Credentials" — a verifiable chain of custody for media, sometimes called a nutrition label for images.

✗ Incorrect. C2PA embeds cryptographic provenance metadata — a verifiable record of origin and edits — not content filtering or consent management.

Lab 1 — Deepfakes & Visual Trust

Explore the ethics of AI-generated imagery and photographic authenticity with your AI lab assistant.

Analysing the Authenticity Crisis

In this lab you will interrogate the ethical stakes of AI-generated photorealistic images. Use the AI assistant to think through real scenarios: What disclosure obligations do photographers have? How do audiences calibrate trust? When does a synthetic image cross from art into deception?

Engage with at least three substantive exchanges to complete this lab.

Try asking: "If I generate a photorealistic image of a protest that never happened and post it without a label, what specific harms could result — and who bears ethical responsibility?"

AI Lab Assistant Ethics · Deepfakes · Visual Trust

Photography and AI · Module 7 · Lesson 2

Consent, Privacy, and the Training Data Problem

Millions of photographers' images trained the models that now compete with them — without permission, payment, or credit.

In January 2023, Getty Images filed suit against Stability AI in the United States District Court for the District of Delaware, alleging that Stability AI had scraped and used more than 12 million photographs from Getty's collection to train Stable Diffusion — without licence, without compensation, and without removing Getty's watermarks (which sometimes appeared, distorted, in generated outputs). The case is ongoing as of 2024 and represents the largest copyright dispute in AI image generation to date.

The suit does not stand alone. In September 2023, a class-action lawsuit brought by artists including Sarah Andersen, Kelly McKernan, and Karla Ortiz against Stability AI, Midjourney, and DeviantArt alleged that the companies had trained on billions of images scraped from the web without consent. The central question was not whether copying occurred — the companies conceded scraping — but whether training on images constitutes copyright infringement under U.S. law.

How Training Data Is (and Isn't) Obtained

The dominant training datasets for image generation models — LAION-5B (5.85 billion image-text pairs), LAION-Aesthetics, and Common Crawl — were assembled by scraping publicly accessible URLs from the internet. Images posted to Flickr, ArtStation, DeviantArt, personal portfolio sites, and stock agencies were included at scale. The LAION team estimated in 2022 that roughly 47% of LAION-5B's images were hosted on just five platforms: Flickr, Wikimedia, WordPress, Imgur, and Facebook.

The legal status of this scraping is contested across jurisdictions. In the United States, the fair use doctrine — tested in cases like Authors Guild v. Google (2015), which permitted Google to index and display snippets of books — may protect transformative uses of copyrighted material for training purposes. However, AI training is distinct from indexing: the model does not store images but learns to reproduce their statistical features, and outputs can be stylistically indistinguishable from specific artists' work. Whether this constitutes infringement remains an open legal question.

In the European Union, the AI Act (formally adopted in 2024) requires providers of general-purpose AI models to publish "sufficiently detailed summaries" of the data used for training, specifically to enable copyright holders to assert their rights. This represents the first significant legislative mandate for AI training data transparency.

The Consent Framework: Who Gets to Say No?

Several mechanisms now exist — or are emerging — for photographers and artists to opt out of AI training data:

Spawning's "Have I Been Trained?" (haveibeentrained.com) allows artists to search LAION-5B for their images and submit opt-out requests. As of 2023, over 80 million images had been opted out. However, opt-out is retroactive — the images have already been used in training runs — and new models built on different datasets are not affected.

Robots.txt extensions (the "ai-crawlers" token proposed by the Spawning API and adopted by some web crawlers) allow website owners to signal that their content should not be scraped for AI training. Major AI companies including Google DeepMind, Common Crawl, and OpenAI have committed to honouring these signals, but compliance is not legally enforceable in most jurisdictions.

Adobe Stock's consent-based model is the clearest positive alternative: Adobe compensates contributors whose images are used to train Adobe Firefly, with bonus payments based on image usage in training. This opt-in, compensated model represents the ethical standard that advocates argue should be industry-wide.

FACIAL RECOGNITION AND STREET PHOTOGRAPHY

Beyond training data, AI facial recognition tools have created a parallel consent crisis in street photography. Clearview AI scraped over 30 billion facial images from the public internet — including social media — to build a facial recognition database sold to law enforcement. In 2022, Clearview was fined €20 million by France's CNIL and ordered to delete French citizens' data. The case established that publicly posted photographs retain privacy protections; publication does not equal consent to facial recognition indexing.

Style, Likeness, and the Limits of Copyright

U.S. copyright law does not protect artistic style — only specific expression. This means that prompting an AI to generate images "in the style of Annie Leibovitz" or "in the style of Steve McCurry" is not, on current legal interpretation, copyright infringement, because no specific copyrighted image is being reproduced. However, this same principle means that photographers cannot copyright their personal visual style, leaving them legally exposed even as their work is commercially exploited.

The right of publicity — which protects individuals' names, likenesses, and personas from commercial exploitation — offers a parallel avenue. In 2023, a Tennessee law (the ELVIS Act) extended right-of-publicity protections specifically to AI-generated vocal imitations, and similar proposals for visual likeness are under discussion in multiple states. For photographers whose recognisable subjects are reproduced in AI outputs, right-of-publicity claims may offer stronger protection than copyright.

PRACTICE STANDARD

When using AI tools in your photographic practice, proactively investigate what data your chosen tool was trained on. Prefer tools with transparent, consent-based training programmes (such as Adobe Firefly). If you generate images incorporating recognisable people's likenesses, obtain explicit consent. Document your workflow decisions — they are increasingly required by galleries, agencies, and competitions.

Lesson 2 Quiz

3 questions — free, untracked, retake anytime.

Getty Images' 2023 lawsuit against Stability AI centred on which primary allegation?

✓ Correct — ✓ Correct. The lawsuit alleged mass unlicensed scraping of Getty's archive, and the appearance of distorted Getty watermarks in AI outputs was cited as direct evidence that Getty images formed part of the training set.

✗ Incorrect. The lawsuit alleged unlicensed scraping of over 12 million Getty photographs for training, with Getty watermarks occasionally surfacing in AI-generated outputs as evidence.

Which AI image generation tool uses an opt-in, compensated model — paying contributors whose images help train it?

✓ Correct — ✓ Correct. Adobe Firefly was trained primarily on Adobe Stock images, and Adobe compensates Stock contributors whose images are used in training — representing the consent-based model that advocates argue should be standard.

✗ Incorrect. Adobe Firefly is the tool with a consent-based, compensated training data model, using Adobe Stock images with contributor payments.

Why does U.S. copyright law currently offer limited protection to photographers against AI tools generating images "in their style"?

✓ Correct — ✓ Correct. U.S. copyright does not protect style, only specific original expression. This gap is why right-of-publicity laws and new legislation are being explored as alternative protections.

✗ Incorrect. The key limitation is that copyright protects specific original expression, not artistic style — generating images stylistically similar to a photographer's work is not the same as copying a specific image.

Lab 2 — Consent, Copyright & Training Data

Interrogate the ethics of AI training data, opt-out systems, and photographer rights with your AI assistant.

Navigating Training Data Ethics

This lab focuses on the consent and copyright dimensions of AI image generation. Explore questions about training data collection, photographers' rights, opt-out mechanisms, and what a fair compensation framework might look like.

Engage with at least three substantive exchanges to complete the lab.

Try asking: "If I post my photographs to Flickr, can I prevent them from being scraped into AI training datasets — and what are my realistic options under current law?"

AI Lab Assistant Consent · Copyright · Training Data

Photography and AI · Module 7 · Lesson 3

Bias, Representation, and Who Gets Erased

AI image systems do not reflect the world equally — they amplify the biases embedded in their training data, with measurable consequences for how people are seen and depicted.

In March 2023, Bloomberg journalists ran a systematic test of five major AI image generators — Stable Diffusion, DALL-E 2, Midjourney, Adobe Firefly, and DreamStudio — prompting each to generate images of people in high-status professions (CEO, lawyer, doctor, judge) and low-status or criminal contexts (fast-food worker, criminal, social-services recipient). The results were striking: all five systems overrepresented white men in high-status roles and darker-skinned individuals in low-status or criminal contexts, in proportions that significantly exceeded even the biases already present in the U.S. workforce.

The study did not indicate deliberate design choices. It reflected a structural reality: the training data — scraped predominantly from English-language Western internet sources — encoded the photographic and representational norms of those sources, which themselves reflect decades of systemic inequity in how people are photographed, published, and distributed.

How Bias Enters AI Image Systems

Bias in AI image generation operates at multiple levels. At the data level, training datasets over-represent certain demographics, geographies, and aesthetics. LAION-5B, for instance, draws heavily from English-language platforms; content from Africa, South Asia, and Latin America is proportionally underrepresented relative to global population. A model trained on this data will generate images that skew toward Western visual norms by default.

At the labelling level, image-text pairs in training data reflect the assumptions of their labellers. If images of women in leadership roles are less commonly captioned "CEO" than equivalent images of men — because fewer such images existed in training corpora, or because labellers used different language — the model learns a skewed association.

At the RLHF level (Reinforcement Learning from Human Feedback), human raters shape which outputs are rewarded. If raters share demographic or aesthetic preferences — as they are statistically likely to, given that many rater pools are recruited from online platforms with their own demographic skews — those preferences become encoded in the model.

The result is a feedback amplification loop: biased photography norms produce biased training data, which produces biased models, which produce biased outputs used in advertising, editorial, and cultural production — reinforcing the original norms at scale.

The Google Gemini Image Controversy, February 2024

Google's Gemini image generation tool launched in February 2024 and was almost immediately suspended for its image generation features following a wave of controversy. Users discovered that prompts for historical images — "Nazi German soldiers," "the Founding Fathers of the United States," "a medieval English knight" — produced historically inaccurate images showing racially diverse groups in contexts where such diversity was anachronistic or distorting. Google's vice president of product for Gemini acknowledged the tool had "missed the mark" and suspended image generation of people entirely while recalibrating.

The episode illustrated the difficulty of correcting for bias without overcorrecting: Google had implemented diversity-promoting interventions in response to the documented whiteness-by-default problem in AI image generation, but without adequate safeguards for historical contexts where diversity injection distorted rather than improved accuracy. Critics across the political spectrum noted that the episode revealed how deeply political the choices embedded in AI training and tuning actually are.

THE STOCHASTIC PARROT PROBLEM

Linguists Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell coined the term "stochastic parrot" (2021) to describe how large language and multimodal models "parrot" patterns from training data without understanding. Applied to image generation: when a model produces a "doctor," it reproduces the statistical average of how doctors appear in its training data — not the reality of who doctors are globally. This is not intelligence; it is very large-scale pattern completion. The ethical weight falls on those who deploy these outputs as representations of reality.

Erasure: When AI Cannot See Certain People

Underrepresentation in training data does not just produce biased outputs — it can produce technical failure for certain populations. In 2015, Google Photos' image classifier notoriously labelled photographs of Black people as "gorillas" — a failure traced to underrepresentation of darker-skinned subjects in training data. Google's solution, reported by Wired in 2023, was to block all searches for "gorilla," "chimp," "chimpanzee," and "monkey" in Google Photos — a workaround that persisted for at least eight years rather than fixing the underlying training data problem.

In facial recognition, a 2018 MIT Media Lab study by Joy Buolamwini and Timnit Gebru ("Gender Shades") found error rates up to 34.7% for darker-skinned women in commercial facial analysis systems, compared to 0.8% for lighter-skinned men. The study directly prompted IBM, Microsoft, and Amazon to audit and retrain their facial recognition products. Amazon halted sales of its Rekognition tool to law enforcement in 2020 after further studies showed error rates that risked wrongful identification of Black individuals in criminal investigations.

ETHICAL PRACTICE FOR PHOTOGRAPHERS

When using AI generation or enhancement tools, actively interrogate the demographic assumptions of the output. Does the tool's default "person" represent the global diversity of human appearances? Are you using AI-generated imagery in commercial or editorial contexts where representational accuracy matters? Consider whether AI-generated images in your work risk perpetuating the erasure or misrepresentation of already-underrepresented communities — and whether photographic fieldwork with real subjects would be more ethical and more accurate.

Lesson 3 Quiz

3 questions — free, untracked, retake anytime.

The 2023 Bloomberg study testing five major AI image generators found a consistent pattern across all of them. What was it?

✓ Correct — ✓ Correct. Bloomberg's systematic testing of Stable Diffusion, DALL-E 2, Midjourney, Adobe Firefly, and DreamStudio found consistent racial and gender bias in depictions of professional and criminal contexts, exceeding even U.S. workforce disparities.

✗ Incorrect. Bloomberg found all five systems overrepresented white men in high-status roles and darker-skinned individuals in low-status or criminal contexts — reflecting and amplifying biases in training data.

The "Gender Shades" study by Joy Buolamwini and Timnit Gebru found that commercial facial analysis systems had dramatically different error rates for different groups. What was the range of error rates?

✓ Correct — ✓ Correct. The 2018 Gender Shades study revealed a 43-fold difference in error rates between lighter-skinned men (0.8%) and darker-skinned women (34.7%) in major commercial facial analysis systems.

✗ Incorrect. Gender Shades found error rates ranging from 0.8% for lighter-skinned men to 34.7% for darker-skinned women — a dramatically unequal performance gap that prompted major product audits.

Google's Gemini image generation tool was suspended in February 2024. What triggered the suspension?

✓ Correct — ✓ Correct. Gemini's diversity-promoting interventions generated historically inaccurate images — e.g., racially diverse Nazi soldiers — leading Google to suspend the feature while recalibrating its bias-correction approach.

✗ Incorrect. The suspension followed user discovery that historical prompts produced racially diverse outputs in anachronistic contexts — an overcorrection for AI whiteness-by-default bias that created its own accuracy problems.

Lab 3 — Bias and Representation in AI Imagery

Examine how training data bias shapes AI visual outputs and what photographers can do about it.

Investigating Representational Harm

This lab explores how AI image generation systems encode and amplify demographic bias — and what ethical obligations photographers have when using these tools. Discuss real cases, interrogate the mechanisms of bias, and consider practical mitigation strategies.

Engage with at least three substantive exchanges to complete the lab.

Try asking: "I'm a commercial photographer using AI-generated imagery for a global advertising campaign. What specific steps should I take to audit my AI tool for demographic bias before publishing?"

AI Lab Assistant Bias · Representation · AI Ethics

Photography and AI · Module 7 · Lesson 4

Disclosure, Labelling, and the Emerging Regulatory Landscape

Voluntary ethics codes are giving way to legal mandates — and photographers who understand the regulatory landscape will be positioned to lead, not scramble.

When the European Parliament voted to adopt the EU AI Act in June 2023 — the world's first comprehensive AI regulatory framework — Article 50 included a specific requirement that AI systems generating synthetic media must ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated." The regulation further required that providers disclose when text, images, audio, or video had been AI-generated, with specific provisions for deepfakes used for satire or artistic purposes needing to carry labels indicating their synthetic nature.

For photographers and visual media professionals, the AI Act established a clear trajectory: disclosure is becoming a legal obligation, not a professional courtesy. The question is no longer whether to label AI-generated imagery, but how — and what the enforcement consequences of non-compliance will be when the Act's provisions fully take effect in 2025 and 2026.

The Disclosure Landscape: Standards, Laws, and Platforms

As of 2024, the disclosure landscape for AI-generated imagery is a patchwork of voluntary standards, platform policies, and emerging legislation:

EU AI Act (2024): Mandates machine-readable watermarking and human-readable disclosure for AI-generated content produced by general-purpose AI systems. High-risk applications (including certain uses of facial recognition) face stricter transparency requirements. Penalties for non-compliance can reach €15 million or 3% of global annual turnover for generative AI providers.

U.S. NO FAKES Act (proposed, 2023): Would create a federal right for individuals to control digital replicas of their voice or likeness, including AI-generated images. As of 2024, the bill had bipartisan support but had not passed.

China's Deep Synthesis Regulations (effective January 2023): Require service providers to obtain user consent before using their likeness or voice, to label AI-generated content conspicuously, and to maintain logs of AI-generated content for 15 days. China is currently the most aggressive jurisdiction in mandating AI content labelling.

Platform policies: Meta announced in February 2024 that it would label AI-generated images on Facebook, Instagram, and Threads using industry-standard signals (including C2PA metadata and Google's SynthID watermarking). YouTube requires creators to disclose AI-generated content in videos, particularly those depicting realistic people or events. TikTok implemented similar requirements in 2023.

Technical Approaches to Watermarking

Google's SynthID, developed by Google DeepMind and launched in 2023, embeds an imperceptible digital watermark directly into the pixel values of AI-generated images. The watermark is designed to survive common post-processing operations — cropping, resizing, JPEG compression — and can be detected by Google's verification tool without access to the original model. Google has opened SynthID to third-party developers via its Vertex AI platform.

Adobe Content Credentials (built on C2PA) take a different approach: rather than altering pixel values, they attach a signed metadata sidecar to the image file that records generation history. This approach is fully transparent and human-readable but depends on platforms preserving metadata — which most social platforms currently do not.

Invisible watermarking vs. metadata represent two philosophies: invisible watermarks are harder to strip but can be defeated by adversarial processing; metadata is transparent and auditable but easily stripped. Most technologists argue both approaches are needed in combination.

THE 2024 U.S. ELECTION AND AI IMAGERY

The 2024 U.S. presidential election cycle produced the first documented large-scale use of AI-generated political imagery in campaign advertising. In January 2024, a robocall using an AI-generated voice imitating President Biden instructed New Hampshire voters not to vote in the Democratic primary — a case that prompted the FCC to ban AI-generated voices in robocalls. On the image side, AI-generated photos of political candidates in fabricated scenarios circulated on social media throughout the campaign cycle, prompting the FEC to consider requiring disclosure of AI content in political advertising.

Photojournalism Standards in Transition

Major photojournalism institutions updated their AI policies substantively in 2023–2024. The Associated Press policy, revised in 2023, prohibits using AI to generate photorealistic images for editorial use but permits AI tools for image organisation, search, and non-editorial tasks. Reuters maintains a similar prohibition. The New York Times states that photojournalists may not alter or generate images of news events using AI.

For documentary and fine-art photographers, the standards are less prescriptive but the professional community is increasingly aligned: images submitted to competitions or published with documentary intent must disclose AI involvement. The Photography Society of America updated its exhibition rules in 2023, creating separate "Creative AI" divisions for AI-generated work, distinct from traditional and digitally manipulated photography categories.

The deeper shift is conceptual: photography is ceasing to be treated as a single medium and is fracturing into disclosure-dependent categories — documentary, editorial, commercial, creative AI — each with its own ethical and legal framework. Photographers who engage with AI tools need to be fluent in which category their work occupies and what obligations attach to it.

YOUR DISCLOSURE CHECKLIST

Before publishing any image with AI involvement: (1) Determine the publication context — documentary, editorial, commercial, or creative/art. (2) Check the platform's current AI labelling requirements. (3) If publishing in the EU, assess whether EU AI Act disclosure obligations apply. (4) Apply C2PA Content Credentials using Adobe's free tools or camera-native solutions. (5) Include a human-readable disclosure in caption or alt text. (6) Retain documentation of your AI workflow — generation prompts, tools used, edit history — for a minimum of 12 months.

Lesson 4 Quiz

3 questions — free, untracked, retake anytime.

What does Article 50 of the EU AI Act require for AI-generated synthetic media?

✓ Correct — ✓ Correct. Article 50 of the EU AI Act mandates machine-readable marking and detectability for AI-generated content, with specific disclosure requirements for deepfakes used outside of clearly labelled satire or art contexts.

✗ Incorrect. Article 50 requires that AI-generated outputs carry machine-readable marks detectable as synthetic, with human-readable disclosure — particularly for deepfakes not clearly labelled as satire or art.

What distinguishes Google's SynthID watermarking approach from Adobe's Content Credentials (C2PA) approach?

✓ Correct — ✓ Correct. SynthID modifies pixel values imperceptibly, surviving post-processing. C2PA attaches signed metadata as a sidecar — transparent and human-readable but dependent on platforms preserving the metadata, which most currently strip.

✗ Incorrect. SynthID works at the pixel level (imperceptible modification); C2PA attaches metadata externally (transparent, auditable, but easily stripped by platforms).

How did the Associated Press update its AI policy in 2023 regarding photorealistic AI-generated images for editorial use?

✓ Correct — ✓ Correct. The AP's 2023 policy draws a clear line: AI-generated photorealistic images may not be used for editorial (news) purposes, though AI tools for workflow, search, and organisation are permitted.

✗ Incorrect. AP's 2023 policy prohibits AI-generated photorealistic images for editorial use entirely, while permitting AI for non-editorial organisational tasks.

Lab 4 — Disclosure, Regulation & Professional Standards

Navigate the real regulatory and professional framework for AI image disclosure with your AI assistant.

Building a Disclosure Practice

This lab focuses on practical compliance and professional ethics around AI image disclosure. Explore how the EU AI Act, platform policies, and photojournalism standards apply to real-world publishing decisions — and how to build a disclosure workflow into your practice.

Engage with at least three substantive exchanges to complete the lab.

Try asking: "I'm a photojournalist who used AI to remove a distracting object from a news photo I'm submitting to a wire service. What are my disclosure obligations under current AP, Reuters, and EU AI Act standards?"

AI Lab Assistant Disclosure · Regulation · Standards

Module 7 Test — The Ethics of AI Photography

15 questions. Score 80% or above to pass the module.

1. Boris Eldagsen's "PSEUDOMNESIA: The Electrician" won a category prize at World Press Photo 2023. What did he then do that sparked global debate?

✓ Correct — ✓ Correct. Eldagsen's deliberate test exposed that major competitions had no clear AI disclosure policies, triggering urgent policy revisions across the industry.

✗ Incorrect. Eldagsen refused the prize after revealing the image was AI-generated, framing the submission as a deliberate test of the competition's readiness.

2. In Peirce's semiotic framework, what distinguishes a photograph (as an index) from an AI-generated image (as an icon)?

✓ Correct — ✓ Correct. Indexicality is the causal bond between sign and referent — light from the real scene formed the image. AI images resemble photographs but have no such causal connection to any real scene.

✗ Incorrect. The index/icon distinction is about causal connection vs. resemblance. A photograph is an index — light from the real scene literally formed it. An AI image merely resembles a photograph.

3. The March 2023 Pentagon explosion hoax involved an AI-generated image that caused measurable real-world harm. What was that harm?

✓ Correct — ✓ Correct. The image spread before verification could catch up, demonstrating that AI-generated disinformation can cause measurable economic harm even when debunked quickly.

✗ Incorrect. The hoax briefly triggered a dip in U.S. stock markets — a concrete demonstration that AI-generated visual disinformation can produce real financial consequences before corrections circulate.

4. Which organisation co-founded the Content Authenticity Initiative (CAI) in 2019, which developed the C2PA standard?

✓ Correct — ✓ Correct. Adobe, BBC, and The New York Times co-founded the CAI in 2019, later developing the C2PA technical standard for content provenance and authenticity.

✗ Incorrect. The CAI was co-founded by Adobe, the BBC, and The New York Times in 2019.

5. Getty Images filed its 2023 lawsuit against Stability AI in which U.S. court, and what was a key piece of visual evidence cited?

✓ Correct — ✓ Correct. The appearance of distorted Getty watermarks in Stable Diffusion outputs was cited as direct evidence that Getty images were included in the training dataset.

✗ Incorrect. Getty sued in U.S. District Court for Delaware; the presence of distorted Getty watermarks in AI outputs was a key piece of evidence that Getty images formed part of the training data.

6. What is the EU AI Act's specific data transparency requirement for providers of general-purpose AI models regarding training data?

✓ Correct — ✓ Correct. The EU AI Act requires general-purpose AI model providers to publish training data summaries detailed enough for copyright holders to identify whether their work was used and pursue claims.

✗ Incorrect. The EU AI Act requires "sufficiently detailed summaries" of training data — the first legislative mandate for AI training data transparency — specifically to enable copyright holders to assert rights.

7. The "Gender Shades" study (Buolamwini & Gebru, 2018) measured which specific problem in commercial AI systems?

✓ Correct — ✓ Correct. Gender Shades quantified intersectional demographic bias in facial analysis accuracy — a landmark study that prompted IBM, Microsoft, and Amazon to audit their products.

✗ Incorrect. Gender Shades specifically measured facial analysis error rates across demographic groups, finding up to 34.7% errors for darker-skinned women vs. 0.8% for lighter-skinned men.

8. What happened to Google Photos' solution to its 2015 gorilla-labelling bias problem, as reported by Wired in 2023?

✓ Correct — ✓ Correct. Rather than fixing the training data, Google suppressed the relevant search terms — a technical patch that avoided the underlying representational problem for years.

✗ Incorrect. Google's reported solution was to block searches for primate-related terms entirely — a content filter rather than a fix — persisting for at least eight years.

9. China's Deep Synthesis Regulations (effective January 2023) include which of the following requirements?

✓ Correct — ✓ Correct. China's Deep Synthesis Regulations are among the most detailed AI content regulations globally, covering consent, labelling, and data retention for AI-generated media.

✗ Incorrect. China's Deep Synthesis Regulations require consent for likeness use, conspicuous AI content labelling, and 15-day log retention — among the most comprehensive AI content mandates globally.

10. What was the central finding of the Bloomberg 2023 study testing five AI image generators with occupational prompts?

✓ Correct — ✓ Correct. Bloomberg found consistent bias across all five tested systems — the bias was not limited to any single tool but was a structural feature of AI image generation trained on Western internet data.

✗ Incorrect. All five systems showed the same pattern: white men overrepresented in high-status roles, darker-skinned individuals overrepresented in low-status and criminal contexts, exceeding actual workforce demographics.

11. What distinguishes Adobe Firefly's training data approach from that of Midjourney or the original Stable Diffusion?

✓ Correct — ✓ Correct. Adobe's consent-based, compensated model is cited as the ethical standard for AI training data — though Firefly also uses some openly licensed and public domain content.

✗ Incorrect. Adobe Firefly's distinguishing feature is its consent-based, compensated training data model using Adobe Stock images — in contrast to scraped, uncompensated datasets used by other major generators.

12. What specific action did Amazon take regarding its Rekognition facial recognition tool in 2020, and why?

✓ Correct — ✓ Correct. Amazon's 2020 halt on Rekognition sales to law enforcement was a direct response to documented racial bias research showing the real-world harm potential of deploying biased facial recognition in criminal justice contexts.

✗ Incorrect. Amazon halted law enforcement sales of Rekognition in 2020 following research demonstrating that bias in the system could lead to wrongful identification of Black individuals in criminal investigations.

13. The EU AI Act's maximum financial penalty for non-compliant generative AI providers is:

✓ Correct — ✓ Correct. The EU AI Act sets penalties for generative AI providers of up to €15 million or 3% of global annual turnover — significant deterrents for large AI companies operating in the EU market.

✗ Incorrect. The EU AI Act sets penalties for non-compliant generative AI providers at up to €15 million or 3% of global annual turnover, whichever is higher.

14. The Leica M11-P, launched in 2023, was notable in the context of AI ethics for which reason?

✓ Correct — ✓ Correct. The M11-P's hardware-level C2PA integration means every image leaves the camera with a cryptographic provenance record, establishing authenticity before any post-processing occurs.

✗ Incorrect. The Leica M11-P was the first camera to embed C2PA Content Credentials at the hardware level — attaching a cryptographic authenticity record to images at the moment of capture.

15. The Photography Society of America's 2023 rule update created a separate exhibition division for AI-generated work. What does this represent conceptually for the photographic medium?

✓ Correct — ✓ Correct. The separation of AI categories in competitions reflects a deeper ontological shift: "photography" is ceasing to be a single medium and becoming a cluster of distinct practices with different truth claims, disclosure obligations, and ethical frameworks.

✗ Incorrect. The category separation reflects photography's fracture into disclosure-dependent practices — each with different truth claims, ethical obligations, and legal frameworks — rather than a temporary or purely technical distinction.