OpenAI Limits Photorealism in ChatGPT's Political Images

Key Takeaways

1OpenAI restricts ChatGPT to illustrated images for sensitive electoral content

2AI verification methods remain limited, and Google does not detail its protections

3Anthropic relies on usage policies, and Claude refuses certain sensitive content

Why it matters : The technical and policy measures taken by major AI players aim to limit the impact of deepfakes on election integrity, but their effectiveness and sustainability remain uncertain.

For election-related content, ChatGPT no longer delivers photorealistic images but rather illustrated renderings. This restriction aims to complicate the creation of political deepfakes, as technical verifications remain imperfect and other platforms offer less detailed protections.

Uncertain Safeguards and Limitations of Verifications

Verification tools or labeling for AI-generated content exist, but they are neither widely adopted nor entirely reliable. Google, for Gemini, refers to an electoral policy that announces additional protections to limit the creation of misleading images related to elections, without specifying how these measures are implemented. There is no guarantee that OpenAI will maintain its restriction after the end of the electoral cycle. However, these initiatives show that OpenAI claims to take into account, at least in part, the protection of the integrity of U.S. elections.

At OpenAI, Sensitive Queries Switch to Cartoon Style

OpenAI has implemented a measure to prevent the creation of political deepfakes by requiring, for certain queries, the generation of images in an illustrated style. Since an update on October 20, requests for images depicting electoral interference or likely to discourage voting no longer produce photorealistic visuals, as these queries are ignored. Tests have generated cartoon-style scenes, such as Immigration and Customs Enforcement agents in front of a polling station, and it has not been possible to obtain a photojournalistic image of a voting machine recording a wrong choice, a frequent electoral myth. Josh Lawson, head of elections at OpenAI, explains that the problem with deepfakes lies in deception, and the goal is to reduce the evidentiary value of images while preserving the possibility for illustration and social commentary, as well as adding a layer of friction to limit malicious uses. The transformation of images into cartoon style thus aims to make the creation of realistic political deepfakes more difficult.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

More Reliable Image Models Complicate Visual Identification

Recent image models, such as ChatGPT Images 2.5, make fewer mistakes and hallucinations, making generated content harder to distinguish from authentic images. This development comes at a time when the electoral period demands reliable information, with the midterm elections scheduled for November 3. A photorealistic image showing false scenes at polling places could cause anxiety among some voters, fueling concerns about deception through deepfakes.

Anthropic Focuses on Usage Policies Rather Than Generation

Anthropic does not offer a traditional image or video generator. The company specifies that any attempt to use Claude to deceive voters or disrupt an election violates its policies and has published general guidance on elections earlier this month. Claude has also refused to create a photorealistic image containing potentially sensitive election-related content.