Mapping the misuse of generative AI

Responsibility and safety
New research analyzes the misuse of multimedia intelligence today, in order to help build safer and more responsibility technologies
Artificial intelligence models (AI) that can produce the image, text, sound, video and more new era of creativity and commercial opportunity. However, with the growth of these capabilities, the possibility of misuse, including manipulation, fraud, bullying, or harassment.
As part of our commitment to the development and use of artificial intelligence responsibly, we published a new paper, in partnership with Jigsaw and Google.org, to analyze how to misuse of childbirth intelligence today. The difference throughout Google uses this research and other research to develop better guarantees for our artificial intelligence technologies, among other safety initiatives.
Together, we gathered and analyzed nearly 200 media reports that capture the incidents of public misuse, which were published between January 2023 and March 2024. From these reports, we identified and classified the common tactics of misuse of artificial intelligence and found new patterns in how to use these technologies or endanger them.
By clarifying the current threats and tactics used across different types of artificial intelligence outputs, our work can help forming artificial intelligence governance and directing companies such as Google and other artificial intelligence technologies in developing more comprehensive assessments and mitigation strategies.
Highlighting the main categories of misuse
Although the tools of obstetric intelligence represent a unique and convincing way to enhance creativity, the ability to produce custom realistic content has realistic content the ability to use them in inappropriate ways by harmful actors.
By analyzing media reports, we have identified two main categories of abuse of artificial intelligence: exploiting the capabilities of obstetric intelligence and the compatibility of artificial intelligence systems. Examples of techniques that are used to create realistic images of human effects included the impersonation of a public figure; While the hacked techniques included “fracture of protection” to remove model guarantees and use hostile inputs to cause breakdowns.
The relative frequency tactics of abuse of artificial intelligence in our data set. Any specific condition of the report reported in the media can include one or more tactics.
The exploitation cases-which include malicious actors that exploit artificial intelligence tools that can be easily accessible at the level of consumer, often in ways that do not require advanced technical skills-are the most prevalent in our data set. For example, we reviewed a high -level case from February 2024, when an international company was said to have lost $ 200 million Hong Kong (about 26 million US dollars) after a employee deceived a financial transfer during an online meeting. In this case, every “person” was at the meeting, including the company’s financial manager, in fact a convincing and newborn.
Some of the most prominent tactics that we have noticed, such as plagiarism, fraud, and artificial figures, before the date of the invention of artificial intelligence before influencing a long time to influence the ecosystem of information and deal with others. But the broader access to the tools of the artificial intelligence may change the costs and incentives behind the manipulation of information, which gives these old tactics effective and new capabilities, especially for those who previously lacking the artistic development to integrate these tactics.
Determine strategies and misuse groups
Falsifying evidence and manipulating human suspicions lies behind the most prevalent tactics in cases of misuse in the real world. In the time period that we analyzed, most cases of misuse of artificial intelligence were published in efforts to influence public opinion, enable fraud or fraudulent activities, or achieve profits.
By monitoring the extent to which poor actors collect between their artificial abuse tactics in seeking to achieve their various goals, we have identified specific groups of misuse and describing these groups as strategies.
A planning drawing of how to map the goals of bad actors (left) on the strategies of misuse (right).
The emerging forms of misuse of artificial intelligence, which are not perfectly malignant, are still ethical fears. For example, new forms of political communication weaken the lines between originality and deception, such as officials, suddenly a variety of languages friendly to voters without transparent disclosure that they use obstetric artificial intelligence, and activists who use sounds created from artificial intelligence for victims available to repair weapons.
While the study provides new visions on the emerging forms of misuse, it should be noted that this data set is a limited sample of media reports. Media reports may give priority to exciting accidents, which in turn may distort the data group towards certain types of misuse. The discovery or reporting of misuse cases may be more challenging for those involved because artificial obstetric intelligence systems are very new. The data collection also does not directly compare the misuse of artificial intelligence systems, traditional content creation tactics and tactics of manipulation, such as editing photos or preparing “content farms” to create large quantities of text, videos, photos, photos, and more. So far, anecdotal evidence indicates that traditional content tactics remain more prevalent.
Staying ahead of a possible misuse
Our paper highlights the opportunities to design initiatives that protect the public, such as progressing in widespread literacy campaigns, developing better interventions to protect the public from bad actors, or warning people and causing them to discover and refute the manipulation strategies used to misuse artificial intelligence.
This research helps our teams to protect our products better by informing us of developing safety initiatives. On YouTube, we are now asking the creators to participate when their work is changed useful or industrially created, and it appears realistic. Likewise, we updated our electoral advertising policies to ask advertisers to reveal when their ads in the elections include materials that have been changed or generously generated.
As we continue to expand our understanding of the malicious uses of the IQ and create more technical developments, we know that it is more important than ever to ensure that our work does not happen in a silo. We have recently joined the content of the coalition and originality source (C2PA) as a member of the Steering Committee to help develop the technical standard and pay the approval of content accreditation data, which is a futile -resistant graphic data that shows how to make content and edit it over time.
In parallel, we are also conducting research that progresses with current red progress, including improving best practices to test the safety of large language models (LLMS), and developing pioneering tools to make the content created by artificial intelligence easier to identify it, such as Synthid, which is integrated into a growing range of products.
In recent years, Jigsaw has conducted research with wrong information creators to understand the tools and tactics they use, developing pre -videos to give up people who have attempts to manipulate them, and showed that pre -campaigns can improve the flexibility of misinformation on a large scale. This work is part of Jigsaw’s broader information interventions to help people protect themselves online.
By treating potential potential use processes in a proactive way, we can enhance the responsible and moral use of the Obstetrician Organization, while reducing its risks. We hope these ideas about the most common abuse methods and strategies for researchers, policymakers, trust and safety teams in the industry will help build safer and more responsibility technologies and develop better measures to combat misuse.
Thanks and appreciation
This research was a collective effort by Nahama Marchal, Rachel Xu, Rasmi Elasmar, iason Gabriel, Beth Goldberg, William Isaac, with reactions and advisory contributions from Mikel Rodrigue, Vijay Bolina, Alexios Mantzarlis and Seleem Elied, Mevan Babakar, Matt Akbut and Matt. Sébastien Krier, Ziad Reslan, Boxi Wu, Frankie Garcia and Jennie Brennan.
2024-08-02 10:50:00