In Do We Expect Too Much of AI? Too Little? I wrote about how ChatGPT posed a serious problem for teachers because their students were very likely going to use the technology to write their essays.  Would teachers be able to detect that they’ve done this?  This is a serious challenge.

And then, as luck would have it, a student named Edward Tian released a tool called GPTZero on January 2, 2023. The goal of GPTZero is to identify whether something was written by an AI, doing so by looking at two factors that Tian calls perplexity and burstiness. Perplexity is how likely each word is suggested by an AI and burstiness measures the spikes in how perplex each sentence is. An AI will likely have a consistent level of perplexity from sentence to sentence whereas people tend to write with more spikes.

We’re In An AI Arms Race

This is an example of applying AI to counteract the effects of someone else applying AI. In short, we’re in an AI arms race – one group starts to use AI to their advantage and another group develops AI to reduce the impact of the work of the first group. In most cases I expect this to go back and forth. For example, in the case of GPTZero I would expect that the people behind the chat platforms, such as ChatGPT, will improve the perplexity and burstiness of their sentences. This in turn will motivate improvements to GPTZero to detect the improved writing. And so on, likely until the quality of the writing gets so good it can no longer be discernible from human-written text.

Here are a few more examples of this AI arms race:

1. Resume scanners. In Applying for Jobs But Getting Rejected? You May Need to Get Your Resume Past the AI I wrote about how people are using AI platforms to provide feedback about their resumes to help increase the chance that their resume will make it through AI-based resume screening software.
2. Disinformation detection. Social media platforms are applying AI to detect disinformation being posted by new bots developed using AI technologies.
3. Cybersecurity. The “bad guys” are using AI to mount attacks on your systems and organizations are combatting that with AI detection technology. For more information, read this AT&T AI cybersecurity article.

Implications

There are several important take-aways:

1. It isn’t surprising that this is happening. As with any tool, AI technology can be used for many purposes, including for competing purposes.
2. Some of the benefits that you gain from AI will be fleeting at best. You always run the risk either of a competing AI-based product being released that outdoes yours, and realistically you should likely assume that it’s only a matter of time until that happens.
3. This is a Darwinistic evolutionary improvement process. Improvements in one AI product will motivate improvements in another to counteract them.

Related: