Generative Artificial Intelligence (AI): Other Tools

Various tools have been designed to distinguish between content written by humans and AI-generated content.

• One analysis of ten such tools found that Winston AI (84% accuracy) and Sapling (68%) were the best paid and free AI detectors, respectively.¹
• There is some controversy as to the real-world effectiveness of AI detectors.²
• As LLMs improve in their ability to imitate human language, AI detectors may become less effective over time.
• We must be wary of false positives. Since presenting an AI's output as your own work may be considered tantamount to plagiarism, it is a serious accusation.

The most prominent generative AI systems are concerned with plain text inputs and outputs, but other applications have been successful.

While many text-focused LLMs are free or have serviceable free versions, audiovisual specialty AI systems tend to have very limited free trials before charging for all outputs.

Most LLMs capable of text processing can handle code in common languages (Python, JavaScript, etc.) as well as human languages.

• LLM chatbots can write code snippets, functions, or even whole programs.
• Chatbots can also help you debug and rewrite your code.
  • You can copy your code and/or error messages into the chat interface and explain the problem in plain English.
• Code writing assistant bots, such as GitHub Copilot, are available for integration with common integrated development environments (IDEs), such as Visual Studio Code.
  • These assistant bots can write code in response to plain text prompts or suggest "completions" based on code you have already written like an advanced autocorrect or intellisense system.

There are serious limitations to be aware of before using LLMs to help you write code. LLMs make mistakes, and since small changes in syntax can radically alter the behavior of a program, it is imperative that you scrutinize AI-generated code before running it.

• LLMs sometimes confuse languages that have similar syntax or share libraries, such as Python and R.
• Less commonly used languages, such as Perl or Lua, may not be as well supported as Python or JavaScript. Some code assistants have a list of supported languages in their documentation.
• LLMs that have a knowledge cutoff, such as ChatGPT, cannot assist with code that depends on updates or modules released after that cutoff.
• Do not give LLMs access to code containing confidential data, such as API keys.

A number of tools have been specifically designed to facilitate literature review. The developers seem to share two main objectives:

1. Find research relevant to supplied plain text (other research papers or plain text research questions) rather than a search string.
2. Synthesize or summarize key findings from search results.

The degree to which these objectives are achieved is as yet unknown.