While conducting our review of generative artificial intelligence (AI) options at IDS, we tested and scrutinized various tools and adopted a conservative approach to AI, keeping a primary focus on human writing.
Amid our testing process, two things became apparent:
Businesses can maintain the safety and integrity of their online platforms by using a well-rounded approach that utilizes content detectors in tandem with AI-generated content.
Table of Contents
These tools use advanced algorithms to sift through large amounts of information, identify language patterns, and flag content that meets programmed parameters indicating the use of AI. Although the intention behind these systems is good, they’re not perfect and can create more problems than solutions, often leading to negative consequences.
While the latest iterations of AI might seem like cutting-edge, modern-day marvels, generative AI research actually began in the 1960s. It wasn’t until 2006, however, that generative AI began to develop into the predecessor of its current form, and up until 2014, there were few innovations in the field.
In the following years, developments allowed generative AI to advance to where it is today. However, most users weren’t aware of generative AI until the past couple of years, and once the technology was released to the general public, awareness skyrocketed.
Consumers were quick to test generative AI technology like text-to-image model services, image-to-image translations, and large language model-based chatbots capable of delivering conversational dialog responses.
As more users utilize generative AI tools, there’s been a rising prevalence of AI-generated content. Here are some interesting industry stats:
As generative AI technology continues to advance and is increasingly infused into every program, more industries are expected to adopt these tools into their daily operations. Innovations are already driving advancements in the journalism, marketing, entertainment, design, automotive, healthcare, retail, finance, and manufacturing industries.
As AI use increases, so does the amount of machine-generated information in circulation. Most users seem open to the idea of more content from generative AI systems. Still, some individuals and organizations have taken a firm stance against publishing AI-generated content without making it clear that the content was generated via machine.
Users may take it upon themselves to employ AI detector tools to verify that their content is, in fact, generated by a human. This is becoming especially prevalent in the following fields:
And it’s not without its merits - with more humans adopting and accepting AI, it’s important to ensure the benefits of automation are balanced with ethical and societal considerations, transparency, and responsible practices.
Many concerns, limitations, risks, and issues still surround AI and have yet to be firmly established. Building trust in AI requires a multidisciplinary approach involving technology, ethics, regulation, and education.
There are many detection tools available that use different methods to determine if content is produced by a human or a machine. Just as generative AI tools are powered by data and pattern recognition, so are AI content detectors.
Using machine learning algorithms and natural language processing techniques, content detectors analyze a variety of features and patterns for known markers of AI generation. Two of the overarching categories for detection methods are linguistic analysis and comparison with known AI-generated text.
Linguistic analysis involves using classifiers and embeddings to sort data. Classifiers analyze the language patterns to recognize text generated by a specific AI model, and embeddings are representations of identified clustered data points.
With classifiers and embeddings, content detectors compare the text against characteristics commonly associated with AI writing and identify patterns. This includes how fluent the text is, how frequently words are repeated, how unique words are, and other markers of AI generation, as well as if the text is similar to previously identified AI-generated content. By identifying known AI patterns in text, content detection tools can thus determine if the sample was generated by a machine or a human.
Now that you have a brief overview of how AI detectors work, it’s time for the million-dollar question: Are they accurate?
. . . . No!
OpenAI, the company that developed ChatGPT, is very transparent regarding the reliability (or, rather, lack thereof) of its AI detection tools. When its AI Classifier detection tool was introduced in January 2023, it was quietly decommissioned six months later because it was not reliable.
Given their experience with complex large language models, OpenAI has found that AI detectors DO NOT WORK. Here are three main points they provide to illustrate their findings:
The company also noted that even if these detectors could accurately identify AI-generated content, there are two considerations:
Content detectors seem promising in a world that’s evolving to adopt more AI-generated media. However, this incipient technology has also introduced a set of potential pitfalls that are bound to increase as its use expands. Becoming aware of these issues can help people make a more informed decision about using detection tools.
Here’s a roundup of the problems with AI-generated content detectors we’ve already touched on:
So, what are some other problems with AI-generated content detectors? Let’s take a look.
The need for accurate and effective AI-generated content detectors grows as algorithms become more sophisticated and detection avoidance strategies evolve. In the meantime, it’s important to develop and employ strategies that can help you combat misinformation when using AI in digital marketing.
With generative AI tools increasing in efficacy, it’s getting harder to distinguish between human-written and machine-generated text. Earlier, we talked about how content detectors look for AI through two main categories - linguistic analysis and comparison - and the two main detection methods these categories use.
Linguistic analysis uses classifiers and embeddings to determine if the text has AI characteristics. In addition to the linguistic analysis classifiers and embeddings, there are other metrics factored into content detectors that are worth considering.
Perplexity refers to the predictability of text. This can be measured by whether the text is diverse and harder to predict, as human-written text has a higher complexity due to nuance and intent.
Burstiness is a measurement of variation within sentences. As generative AI tools only contain the information set they’re programmed with, this can lead to an overuse of certain words and a lack of variation.
There are many benefits of AI in digital marketing, but content detectors have a ways to go before they are truly useful, effective, and reliable. These tools should help users manage and filter vast amounts of information in addition to providing other benefits like increased efficiency, accuracy, and scalability.
Developers are teaching AI detectors techniques like classifiers and embeddings to more effectively examine content, and they’re good practices to consider when reviewing content for authenticity yourself.
The continuing evolution of AI and its detectors is an ongoing opportunity to reshape how we communicate, express ideas, automate processes, and consume information in the digital world.
To learn more about how IDS can help your business with a specialized digital marketing campaign, get in touch with a member of our team today!