Photography in the age of AI

Learning goal

You will be able to explain in your own words how AI-generated images are changing photography and which value of photography remains. Through the concept of witness, you will discuss the difference between an image generated from a prompt and a photograph formed by light falling on a sensor or film at a moment, in a real place. You will tell apart the jobs in which generated images can take the place of photography from those in which nothing can replace being there in a real place, at a real moment and with real people. You will apply source verification steps to understand whether an image is real, and recognize how generative models can produce skin tones and people in stereotyped ways, that is, model bias. In this lesson's assignment you will complete the module's project as a draft release form and a list of ethical principles for your own field of work.

The question

One morning, on a news site, your friend sees a widely shared frame showing a family waiting on the roof of their house in a flood. A few days later it turns out the frame was produced with an AI tool and that no such family ever existed. At the club a member shrugs: soon no one will need a camera; a few words are written and any image wanted is produced.

Your friend thinks about the frame they shot on the night of the fire, which the mother allowed to be published. What is the difference between that frame and the family on the roof? Will AI replace photographers?

Context

While examining the new tools in editing programs, you saw that a tool filling an empty area can invent details that were never there, and that the viewer must be told about this (Lesson 21.12). In the stock photography lesson you saw how generated images are changing the stock market and the different rules of agencies (Lesson 28.10). In the manipulation lesson, too, we discussed that what a photo makes the viewer think it is lies at the heart of the ethical problem (Lesson 29.4).

The last question of this series built on law and ethics looks to the future. Debates about AI change fast; many details written today may be outdated tomorrow. What is taken up here isn't the names or capabilities of particular tools, but a question that won't change in the long run: what makes a photograph a photograph?

Simple explanation

What is the difference between a generated image and a photograph?

A generated image is an image that an AI model computes, starting from a text prompt or other images, by combining the patterns in the large number of images it was trained on. It may look like a photograph; but no light, at no moment, in no place, has fallen on a sensor. The people in the image may not exist, and the place may never have existed.

Witness is the value a photograph gets from having recorded, at a particular moment, in a particular place, the light of something real in front of it. A photograph can be interpreted, chosen with a framing, shaped with editing; but at its heart is the knowledge that the photographer was there and that something really happened. The value of the fire night frame comes from here too: the neighbors really did bring blankets, the mother really did give permission. However much the image of the family on the roof shakes a person, it carries no such witness. What's more, a fake image circulating in the middle of a real disaster gets mixed up among the real victims' calls for help, and can shake people's trust even in real frames. The value of witness is the value not only of one photograph's credibility, but of all photographs' credibility.

What can AI replace, and what can't it?

In some jobs generated images have truly begun to get ahead of photographers: stock visuals illustrating abstract concepts, sketches of places that don't exist, mood boards for advertising, backgrounds. For photographers working in these fields the market may shrink; this is a realistic situation, and a photographer ignoring it can put their work at risk.

Being there, on the other hand, is a photographer physically witnessing a real event, a real place and real people, and it is the core of the jobs AI can't do. A family's wedding, a child's first steps, a bakery master's morning dough, the real moment of a news story, a client's own venue; these can be shot only by being there. In the same way, asking a person for permission, earning their trust, helping them relax in front of the camera and showing them their photo is also a photographer's relationship labor, and can't be generated with a prompt. Many photographers, while using AI as a tool in some steps of their workflow, choose to base the value of their work on exactly this witness and this relationship.

How do you tell whether an image is real?

Source verification is trying to understand whether an image is real by researching where it came from, who took or generated it, when and where. A few questions help: who shared the image first, is there a photographer's or a news organization's name, are there frames of the same event taken from other angles, does a reverse image search point to another source, are there signs in the image such as inconsistent light, meaningless lettering or strange hands and faces? These signs may decrease over time; that is why the most reliable way is to look at the image's source and context. Some tools and organizations have started using content credentials, which carry inside the file how the image was created.

Model bias is an AI model carrying the imbalances and patterns in the images it was trained on into the images it produces. According to many studies and reports, generative models can pair certain professions always with certain skin tones and certain regions always with poverty, or produce dark skin with less detail. Using generated "diverse" faces in a campaign in place of real people can also erase the visibility and labor of real people. A photographer can notice and question this bias both in their own work and in the images they see.

Visual evidence

Your friend opens the image of the family on the roof again and asks the source verification questions one by one. The account that first shared the image has no name, no such house appears in frames of the same flood taken from other angles, and a reverse image search leads to no news source. Looking closely, one of the hands of the child on the roof has an extra finger.

Beside their own fire frame they place the information showing they were there that night: the raw file, the shooting time, the mother's written permission and what she said. When sharing on the club's page, they add some of this information under the frame.

Finally they write a list of ethical principles for their own field of work, neighborhood businesses and people: getting permission directly from the person, preserving each skin tone's own tone, not doing content manipulation in documentary work, stating clearly when they use AI, helping first with people in suffering. They keep the list next to their portfolio, together with the draft release form they prepared earlier.

Common mistakes

Summary

A generated image is an image a model computes from the patterns it was trained on, which is the record of no light at any moment in any place; a photograph's distinguishing value, on the other hand, is that it bears witness to something real at a particular moment, in a particular place. AI can be a strong competitor in fields such as abstract stock visuals and sketches, but it can't produce being there and the relationship of trust built with people; source verification helps understand whether an image is real, and how model bias can produce skin tones and people in stereotyped ways is questioned.

Mini assignment

In this assignment, as the module's project, you will prepare a draft release form and a list of ethical principles for your own field of work.

1. Bring the draft release form you started in the model and property releases lesson to its final form for your own field of work; add the parental consent and property release sections.
2. Write a list of at least eight ethical principles for your own field of work; permission, preserving skin tones, content manipulation, publishing on the street, people in suffering and disclosure of AI use must be on the list.
3. Apply the source verification questions to a widely shared image you saw in the last week and note what you find.
4. Write two lists of the steps in your own work in which you could use AI as a tool and the steps in which you will never use it; describe in a paragraph the value being there adds to your work.

Check yourself

• Can you explain in your own words the difference between the fire frame and the family on the roof?
• Do you know which questions you will ask to verify an image's source?
• Does your list of ethical principles clearly cover permission, skin tone and AI disclosure?

Glossary

Mini quiz

1. What basically distinguishes a generated image from a photograph?

• A) Its colors being more vivid
• B) It not being the record of real light at a particular moment in a particular place
• C) A smaller file size
• D) It being black and white

2. In which job does AI have the hardest time replacing a photographer?

• A) A background illustrating an abstract concept
• B) A sketch of a place that doesn't exist
• C) A family's real wedding day
• D) A mood board for an advertisement

3. Which can be done to verify the source of a widely shared image?

• A) Doing a reverse image search and looking for frames of the same event from other angles
• B) Sharing the image more
• C) Looking only at the number of likes
• D) Changing the image's color

4. What does model bias describe?

• A) The camera's lens choice
• B) The photographer's style
• C) Errors in light metering
• D) A model carrying the patterns in its training data, for example stereotypes about skin tones, into the images it produces

Answers

1. B — A photograph's value comes from the witness of a moment and a place; a generated image is no such record.
2. C — A real event can only be shot by being there and building a relationship with people.
3. A — Looking at the image's source and other witnesses is the most reliable way to understand its reality.
4. D — Models can multiply the imbalances in the images they were trained on; this can lead them to produce skin tones and people in stereotyped ways.

Summary card

• A photograph is witness; a generated image isn't.
• AI can take some jobs, it can't take being there.
• Look at the source of a widely shared image.
• Question model bias; photograph real people with their permission.
• When you use AI, state it clearly.

Bridge to the next lesson

The list of ethical principles and the release form are ready; your friend realizes they are nearing the end of the program. The older member suggests one last piece of work: a capstone project lasting weeks that will bring together everything they have learned in a single project. Your friend doesn't know what to shoot; on their mind are the neighborhood, the bakery, morning light and flowers, but they don't know which of these truly makes them curious. Which story do you want to tell? The next lesson takes up choosing the capstone project's subject, starting from a question that truly makes you curious.