Generative Artificial Intelligence

General Principles

The bad news is that there is no one-button solution for identifying if a piece of text or media is fake or incorrect. The good news is that some of our oldest methods of information verification still hold true today.  Always verify the source of an image, video, or soundbite.

For current events or claims about historical events, trace the information back to sources that would be in a position to know about the event and to be truthful in their telling of it. For a source referenced by the generative AI, that would mean going back and checking the source. For a quotation from a politician, that would mean locating a news source that recorded the statement, rather than a quotation or clip shared on social media claiming to come from that source. For a quotation from a historical figure or a picture of that person, a more reliable source might be an original or scanned copy document itself shared by an archive or quotations provided by a historian who has access to the original document. 

Misrepresented Authorship and Undisclosed AI Use

Sometimes the issue isn't a claim about what someone else said but a false claim of authorship. While some generative AI use is blatant, often the the differences between AI-generated text and human generated-text are subtle.  It would be nice if AI detectors worked to catch it, but they are only somewhat better than human readers.

Aside from racing against improvements in generative AI sounding more natural and against workarounds people create, both computerized detectors and human readers have false positives. False positives are when a human-generated text gets incorrectly flagged as being AI-generated. In other words, both human readers and computerized systems have an imperfect ability to detect AI generation.

Both human and AI detectors can be correct much better than chance, but they can't be certain. Other methods, such as directly interacting with the person claiming authorship to understand their knowledge of the content of the text or comparing with writing samples from similar tasks performed without internet access, can supplement the imperfect human or computerized AI detector to increase confidence on whether an item is the author's own work. 

Text: References and Quotations

If the generative AI gives a citation to a particular source, check that source to verify it exists or to examine it directly for the quotation. 

For a quotation claimed to have come from a particular person, a searching for a "phrase in quotation marks" can sometimes lead directly to news sources that report on it. If multiple sources that you trust contain the phrase, it's more likely to be true than if the statement is only being repeated by sources that you aren't familiar with or know to be unreliable.

For historic figures, online quotation lists often repeat the same mis-attributions to things someone never said, but they can sometimes be helpful by also providing information about where the original quotation came from. Those sources can then be consulted to verify where they came from. Online archives of books are often helpful for tracking down those quotations. 

Images

For still images, Google reverse image search can be an extremely helpful starting point. This will identify if the image came from a personal social media page (unreliable) or from a more reliable publication (credible news story, peer-reviewed article, etc.). It also can help track back the date the image originally appeared, to verify if the description of the image's context is correct. 

Video

For video, verify the source of the video. Is it from a credible news source or posted on social media? Are either multiple angles of the event or multiple streams across different news stations? Is this a video that has been verified by experts? 

Audio

For audio, search part of the soundbite using quotation marks "quote from the soundbite" to see if a full transcript can be found. Then verify the quality of the sources of that transcript and listen to the quote in its full context. This soundbite should not be trusted until you can verify it within the context of its full original recording.