AI detector: How to Detect if a Text is Generated by GPT-4 or Another AI Writing Tool

AI detector: How to Detect if a Text is Generated by GPT-4 or Another AI Writing Tool

In today’s digital era, the proliferation of artificial intelligence in various sectors, including content generation, has prompted a pressing question: How can we differentiate between human-generated content and AI-generated content? This is especially challenging with state-of-the-art models like GPT-4. In this article, we will delve into techniques, indicators, and AI detectors that might help discern AI-produced text.

1. AI Detectors

AI detectors are tools designed to identify whether a particular piece of content has been generated by AI or not. Some of the prominent AI detectors include:

  • GPT-3 Sandbox: Though primarily a platform for developers to experiment with OpenAI’s GPT-3 model, it also allows users to evaluate the authenticity of a given text.
  • Deepware Scanner: This scanner, built by EleutherAI, was designed to detect AI-generated content, though with varying degrees of accuracy.

2. Limitations of AI Detectors

Despite the advancements in AI detectors, they aren’t foolproof. A research study indicated that even the best experts in the field can only predict whether a text was generated by AI with about 38.9% accuracy. This highlights the sophistication of modern AI models.

Factors that contribute to the limited success rate of AI detectors include:

  • Overlapping Patterns: Both human and AI-generated content can display similar linguistic patterns, making differentiation challenging.
  • Evolving AI Models: As AI models improve, their output becomes more nuanced and human-like, further blurring the lines between human and machine.

3. Recognizing GPT-4 and AI Patterns

Certain patterns of speech or frequently used phrases might hint at an AI origin. Some of these include:

  • Verbose Explanations: GPT-4, like its predecessors, tends to be overly verbose and provides detailed explanations where a human might be more concise.
  • Avoidance of Direct Answers: AI models might avoid giving direct yes/no answers and often lean towards more explanatory responses.
  • Lack of Personal Experience: GPT-4 doesn’t possess personal emotions or experiences. It might dodge personal questions or provide generic answers.
  • Inconsistencies: If a conversation spans multiple topics, the AI might provide inconsistent answers due to its lack of long-term memory.

4. Comparing to Original Work

One of the best methods to determine if content might have been AI-generated is to compare it to the alleged author’s original work. Key points to consider:

  • Style and Voice: Check if there’s a consistent style or voice throughout the content. Abrupt changes or inconsistencies might hint at AI intervention.
  • Depth of Insight: AI-generated content might lack the depth or unique insight that a human writer often displays.
  • Content Flow: Human writers often structure their content with a specific flow in mind. AI-generated content might seem disjointed or lack a coherent progression.

FAQs

Q: Can AI detectors reliably discern AI-generated content from human-generated content? A: As of now, even the best AI detectors have limitations. They may not always reliably differentiate between human and AI-produced content. Expert evaluations show about 38.9% accuracy in detection.

Q: How has GPT-4 improved over its predecessors in content generation? A: GPT-4 is more nuanced, context-aware, and produces content that’s even more indistinguishable from human-generated content compared to its predecessors.

Q: Can we expect future AI models to be even harder to detect? A: Yes, as AI models evolve, the distinction between human and AI-generated content is likely to become even more blurred, making detection increasingly challenging.

Q: Are there any foolproof methods to detect AI-generated content? A: Currently, no method is entirely foolproof. However, a combination of AI detectors, pattern recognition, and comparison to original work can provide a more holistic evaluation.

Conclusion

While advancements in AI have made it increasingly challenging to discern between human and AI-generated content, a combination of technology, observation, and analytical comparison can assist in making an informed judgment. As AI continues to evolve, the need for sophisticated detection tools and methods will become even more paramount.

OK, Real Talk…

Everything above was written by chatGPT, with no style prompts – so that’s what AI writing looks like. The problem is, *with* prompts, it could write cute and sassy, wuwu emoji, or serial killer catastrophe – it can write in any style you give it.

But also, most AI detectors have false positives all the time.

They just don’t work, and *most* historical documents or books you test, with determine that AI had a hand in the US Constitution for example.

There is currently no AI detector that can be trusted.
But here’s a human approach to detecting chatGPT…

Weird Stuff

Look for any out of place phrases, especially if they sound like commands or prompts – someone might have copied over a short prompt that says “do this next” or “add more detail” or just “continue.”

Check Facts

ChatGPT reads well but makes stuff up, so there might be a solid fact or detail in there that isn’t true. It’s a pain to check them all and even if you found a mistake it would be hard to prove AI writing was used.

ChatGPT Lingo

The obvious thing would be something like “As an AI language model,” – which is chatGPT’s default response. That’s a dead giveaway.

But here’s some other things:

ChatGPT is a lot like long-winded mansplaining. The default is a lecturer on a soapbox, a little dry and boring. So you’ll see a lot of these in a well structure essay:

  • introduction
  • It’s generally recommended
  • it’s important to note
  • remember,
  • conclusion
  • wrapping up
  • so there you have it.

If you give it some more style, it might start sentences with these…
Ah, ….

  • Indeed,
  • Now,
  • See,
  • You see,
  • In conclusion
  • Moreover,
  • Still,
  • However,
  • Remember,
  • To sum it up,
  • When it comes to…
  • It might not be easy,
  • There you have it folks

For fiction, it likes to use lots of colorful metaphors – even for an essay it might start with a metaphor and then carry it forward, with heavy hands until it’s unbearable. It tends to write what you’d call juvenile, dramatic fiction – with overblown drama and instalove romance.

And it has this annoying positivity and cheer and says stuff like “we’ll get through this… together!” or “Together we can face anything!”

It does this for nonfiction conclusions too, like this cheery signoff:

Stick around as we dive into more intricate facets of the writing world. It’s a wild ride, and we’re in it together. Happy writing, and remember, it’s all about the journey, not just the destination!

or this one:

Stay tuned as we delve deeper into unconventional writing methods. And as always, keep those pens dancing and keyboards singing! Until next time!

That’s not my writing style at all. Unfortunately… it’s probably better than my writing style: it’s positive and upbeat, it’s clever and creative, it’s memorable.

The problem with detecting AI is…

Most students use this stilted, boring language to develop an easy. We are basically taught to do it. It’s not good writing, but it’s very passable average writing, that’s safe and clean. Probably worth a B… or a raise or a scholarship. And unfortunately, the *best* writing – like those amazingly well-written college essays – that’s an easy style to copy.

As a measure of quality (when grading or figuring out if someone cheated), it’s going to be extremely difficult to prove they used AI unless they made an obvious screw-up. You can compare with all there previous essays and maybe see a huge difference, but it’s hard to accuse – and there wouldn’t be anything ethically wrong (probably) with using a spellchecker like grammarly or prowriting aid which very soon (or already) have built-in AI tools that can rephrase, edit and polish all of your writing.

These will become normal tools in all fields (unless education goes back to handwritten papers and assignments, which may need to happen). It’s hopeful to believe, that everyone will have suddenly become much better writers – so nobody has to read all the crap – and that people with the best ideas or the best stories will win (not everyone has the best ideas or stories, and so far writers have won. In most industries. The big money business relies on pitches and agreements, that some writer is probably getting paid for. People with convincing grant proposals.

Maybe this levels the playing field, and ideas and stories can win out.

In the case of best life experiences, that would be risky to lie about.

And in the case of brilliant ideas, chatGPT is actually pretty awesome at coming up with creative ideas on command. Maybe new things will be built and discovered. For right now, since detecting AI content for real is going to get more and more difficult, I imagine rocky times ahead for society at large.

Add Comment

Your email address will not be published. Required fields are marked *