Feature
An AI-generated image of Pope Francis wearing a puffer jacket went viral. A stunning surfer photo won a photography competition, only to be revealed leater as an AI-generated image. These, and many other examples, were all preceded by an AI-generated digital painting winning a digital arts competition at the Colorado State Fair. All were created with early versions of the AI image generator Midjourney, which has since improved. Similar examples can be found across image, text, video and audio content created by generative AI.
What can be done? If art critics and internet users can’t distinguish between machine-generated content and human-generated content, is there hope for finding the truth online? Additionally, how much does it matter if something is AI generated versus created by human hands? While these questions veer into philosophical territory, which we’ve covered before, hope is not lost. There are specific actions individuals can take from a critical thinking perspective. Industry development and experimentation in analyzing and watermarking AI content are accelerating. The market is evolving quickly.
Here’s the latest on detecting AI-generated content.
Paying Attention is the Best Way to Manually Identify AI-Generated Content
The most straightforward way to identify AI-generated content is to pay close attention. A cursory glance at an image or text might miss subtle clues that a more careful examination might reveal as outlined below. To help with this critical thinking, use the “SIFT” method. Developed by Mike Caulfield, a research scientist at the University of Washington, “SIFT” stands for:
- Stop: consider the source of the image and whether it matters whether it is machine-generated or altered
- Investigate the source: understand the incentives and expertise of the source
- Find better coverage: if applicable, see if content is verified on other sites
- Trace: identify the original context and source of the information.
The consideration of “whether it matters” is often contextual and can help reduce anxiety and investigative rabbit holes into the truth of everything — otherwise, it’s easy to trust nothing.
Tactical Manual Actions
While it is becoming increasingly difficult to definitively identify machine-generated content manually, especially when produced by a skilled AI operator, there are several signs and key categories to consider for both images and text.
Manual AI Image Detection: Look at the ‘Irrelevant’ Parts of the Image
After a review of a variety of sources and forums, manual detection of AI-generated images generally falls into three key categories. These categories can best be summarized as look at the background and minutiae of an image that are often not the focal point:
- Visual errors: count fingers and toes, look at strands of hair, ears and eyes
- Unnatural situations: look for unnaturally blurry and warped objects, especially those in the background
- Overall style: AI images often have a vibe or a style. Play around with AI tools to develop this skill
Importantly, many of these issues don’t arise in every AI photo — and if the point of the AI image is to fool, then the creator might likely avoid these common issues.
Manual AI Text Detection: Use Contextual Cues and Word and Phrase Combinations
Automated, text-generating bots precede modern AI tools — consider the dead internet theory for a provocative idea. However, modern machine-generated text is much more malleable and responsive given the prompt-response nature and the large underlying data sets. Furthermore, normal people can sound robotic in certain contexts, such as professional email, so the various tone capabilities of a large language model (LLM), like ChatGPT, make it difficult to identify its creations consistently.
There are certain words that are increasingly associated with AI tools. For example, ChatGPT often uses the words tapestry and delve. The use of delve, in turn, has increased substantially in medical journals. On one hand, this might mean that users are directly using AI-generated outputs, while on the other hand, the users of common AI words might simply be mirroring words they have read from AI-generated content. Tools and culture influence each other, so unless words are universally known as only used by AI, judging manually is prone to biases and false-positives
Software Detection is Inconsistent at Best
While many websites and software analysis tools claim to detect AI, there are no clear, definitive solutions. Companies are economically disincentivized to make their outputs more detectable as AI generated than legally necessary, especially if it impacts quality. Models and creative use of tools make it a constant game to identify AI-generated content but change is likely. Here’s the state of software detection:
Text-Generation Analysis Tools are Prone to False-Positives
Those investigating text generation may use a tool like GPTZero. GPTZero uses a variety of proprietary methods to identify AI text and has been the focus of numerous studies evaluating success. It is often accurate, but it still has false-positives or human-generated text identified as AI-generated text. Notably, OpenAI retired its own “AI text detection classifier” after a few months last year, given a “low rate of accuracy.” The incidence of false-positives for tools like GPTZero may make their use akin to lie detectors — sometimes completely accurate and sometimes completely wrong.
No matter the contextual focus of a text analysis tool, a textual software analysis should be a component within a greater investigation.
Analysis Tools are Rapidly Being Released, Including by the Companies Themselves
While OpenAI retired one detection tool, OpenAI and other companies have released alternatives. OpenAI recently announced a new set of tools for researchers to identify whether OpenAI technology produced a particular piece of content. Hugging Face has a collection of tools for detecting a variety of content formats, and Meta labels images created on its platform as “Imagined with AI.” Some of these developments likely result from President Joe Biden’s executive order on AI, requiring the ability of AI companies to “… ensure that users know when content is AI generated, such as a watermarking system.”
SynthID, Google DeepMind’s watermarking system, started beta software last year with Gemini-produced images and has been expanded to include text and video. Using a subtle alteration to the next-token prediction process universal to all large language models, Google states that its process maintains the quality of generated output while making analysis easy. Reviewing the above blog post by Google is highly recommended to understand how watermarking works.
A watermark may be removed by alteration software or by re-writing the machine-generated passage. However, some early research suggests that human editors can still fail to remove watermarks — especially for text less than 1,000 words. Extensive editing and transformation of any image or alternative content may equally be able to remove a watermark, however, so no approach is complete.
It is also unclear how watermarking will work with open-source models, and President Biden’s executive order only applies to companies based in the U.S. Watermarking still requires an additional level of analysis, often with tools that are currently in beta and/or only available to certain researchers.
Digital Literacy is Increasingly Important, Especially for Divisive Content
Knowing the authentic from the fantasy and the real from the generated is an increasing challenge online. Digital literacy skills, actively engaging and considering the truth of content have always been useful, but the accessibility of machine-generated content makes it more important than ever to think critically about what we consume online.
The topics of conversation that are most important and highest stakes are likely to be those with the most artificially generated content. Unfortunately, 49% percent of the world votes in 2024, and while Google, Anthropic and OpenAI have pledged to reduce election misinformation, it’s likely many global citizens will be influenced — without their knowing — by AI-generated content. Alternatively, claiming something factual is AI generated may become more commonplace. Paying attention and not jumping to conclusions will be essential habits for the foreseeable future.
