Ad Fontes Guidelines for Generative AI, Errors and Retractions, and Plagiarism
Generative AI content
Ad Fontes Media considers content pieces rated 36 and above to be reliable. We believe the news landscape is best and most accurately served by human writing, not generative AI. Therefore, only human-generated content can be rated 36 or above. Disclosed use of generative AI can score 36 or above if it is used for things like topic selection, initial outlines, suggesting alternate language for final drafts, generating translations, text to audio, with a human doing the actual writing and final editing before publication. Disclosed generative AI summaries can also be included in content rated 36 and above.
Disclosed generative AI content that is fully fact checked and edited before publication can be rated between 24 and 35 (the higher up the disclosure is on the page, the better). AI-generated summaries (disclosed or undisclosed) that are confusing, irrelevant, or awkwardly worded get rated in the 24-35 range.
Artificial images need to be disclosed if they alter a real event, person, or place, or present a realistic scene that didn’t actually occur. Purely decorative AI images do not need disclosure.
If we find some evidence of undisclosed AI generation, but it’s inconclusive, then we can bring the reliability score down to the 24-25 range, but not below the 24 line.
Any content with an AI-generated mistake will automatically be rated below 24.
If a source discloses generative AI content, but the disclosure is not on the article page, not even as a hyperlink, then it will score a 23 for reliability. If we find a preponderance of evidence for undisclosed generative AI, the content will score in our 16-23 “incomplete” category.
Disclosed generative AI content that was likely not fact checked or edited before publication or contains minor (grammar), AI-generated mistakes, including AI images, will score in our 16-23 incomplete category.
If we find clear and convincing evidence of undisclosed generative AI that is beyond a reasonable doubt, then the content will score in our 8-15 “misleading” category.
Disclosed generative AI content that was clearly not fact checked or edited before publication, contains plagiarized content that AI has reworded, or contains misleading AI-generated mistakes, including AI images, will score in our 8-15 misleading category.
If we find the same evidence level of undisclosed generative AI as listed in the 8-15 range, but the source asserts human authorship or contains plagiarized content that AI has reworded and is under another person’s byline, then the content will score in our 0-7 “inaccurate” category.
AI-generated content that is inaccurate, including AI images, will score 0-4 for reliability.
Accountability Process for Corrections
Ad Fontes Media monitors mistakes, errors, and changes in the news landscape. For prominent articles with corrections, we have an accountability process that takes into account factors like the severity of the error, whether it was a generative AI error, the wording of the correction, and the length of time before the correction. Scores of the corrected article are averaged with the original article score that took the error into account. The final score is not necessarily a strict average, as our analysts have discretion in weighing mitigating and aggravating factors to apply to the final score.
Ad Fontes Media considers these the best practices for corrections:
- The correction is at the top, before the body of the article begins.
- If the error is present in the article url, it redirects to a new url with the corrected article.
Corrections should always:
- State exactly what the error was.
- Take responsibility for the mistake.
- Be labeled a correction instead of an editor’s note, if an error was corrected.
If those conditions are met, we consider it a full correction and will give full credit when rating the corrected article. If those conditions are not met, we consider it a half-correction and will take away four reliability points when rating.
If the premise of the story was bad and cannot be salvaged, the text of the story should be removed and replaced with an editor’s note. Such retracted articles are given 8 additional reliability points. Articles originally rated under 24 that are removed with no note are given some credit, but never above 24. At best, they will receive a score in the incomplete range of 16-23.
Publishing a new article with correct information but leaving the original article up and unchanged will not raise the low score of the original, uncorrected article.
Any updates to an article should always explain exactly what was updated.
Plagiarism
Before rating any content as being plagiarized, we must first conclude beyond a reasonable doubt that the source does not pay for syndication, since some syndicates do not require attribution.
If we find some evidence of plagiarism, but it’s inconclusive, then we can bring the reliability score down to the 24-25 range, but not below the 24 line, which marks the upper boundary for unreliable content.
If we find a preponderance of evidence of plagiarism, we rate the content in the 16-23 “incomplete” category. If we find clear and convincing evidence of plagiarism, we rate the content in the 8-15 “misleading” category. If we find clear and convincing evidence of plagiarism with an incorrect person listed for someone else’s work, we rate the content between 0-7, in our “inaccurate” category.
