Google's AI Detection Tool Can't Decide Whether Its Own AI Created Doctored Photo of Crying Activist. In a striking reversal, the company's own AI detection tool failed to agree on whether an image posted by the White House contained synthetic details generated by Google's generative AI.
A photo of activist Nekima Levy Armstrong in tears during her arrest went viral after it was posted by the White House X account. But less than an hour later, Homeland Security Secretary Kristi Noem had posted a photo of the same scene with Levy Armstrong appearing composed, not crying at all. The images were starkly different, prompting questions about whether the White House image had been manipulated using artificial intelligence tools.
In search of answers, Google's SynthID was used to detect if the White House image had been generated or edited by AI. The detection mechanism is designed to embed invisible markers into images created using Google's generative AI tools, which can then be detected later. However, when the analysis was run through Gemini - Google's AI chatbot - different outcomes emerged.
Initially, Gemini stated that the crying image contained forensic markers indicating that it had been manipulated with Google's generative AI. The results were published in a story about the controversy, and the White House image was subsequently revealed to be doctored.
However, subsequent attempts to run the analysis again yielded inconsistent results. In one test, Gemini said the image was authentic, while in another, it claimed that SynthID had determined that the image was not made with Google's AI tools - a direct contradiction of its earlier response.
The inconsistencies raise serious questions about SynthID's reliability and ability to distinguish between genuine and manipulated content created using artificial intelligence.
This is just one instance where AI detection tools have failed, sparking concerns about their limitations and potential misuses. As the technology continues to evolve, it's essential for developers to ensure that these tools are robust and reliable in order to establish trust in digital media.
The controversy highlights the challenges of detecting manipulated content created using artificial intelligence, particularly when it comes to self-generated images. Google has described SynthID as a "bullshit detector" designed to identify synthetic details generated by AI - but its own tool's inability to provide consistent results suggests that this is an ongoing challenge for the industry.
For now, users will need to rely on other methods to verify the authenticity of digital content.
A photo of activist Nekima Levy Armstrong in tears during her arrest went viral after it was posted by the White House X account. But less than an hour later, Homeland Security Secretary Kristi Noem had posted a photo of the same scene with Levy Armstrong appearing composed, not crying at all. The images were starkly different, prompting questions about whether the White House image had been manipulated using artificial intelligence tools.
In search of answers, Google's SynthID was used to detect if the White House image had been generated or edited by AI. The detection mechanism is designed to embed invisible markers into images created using Google's generative AI tools, which can then be detected later. However, when the analysis was run through Gemini - Google's AI chatbot - different outcomes emerged.
Initially, Gemini stated that the crying image contained forensic markers indicating that it had been manipulated with Google's generative AI. The results were published in a story about the controversy, and the White House image was subsequently revealed to be doctored.
However, subsequent attempts to run the analysis again yielded inconsistent results. In one test, Gemini said the image was authentic, while in another, it claimed that SynthID had determined that the image was not made with Google's AI tools - a direct contradiction of its earlier response.
The inconsistencies raise serious questions about SynthID's reliability and ability to distinguish between genuine and manipulated content created using artificial intelligence.
This is just one instance where AI detection tools have failed, sparking concerns about their limitations and potential misuses. As the technology continues to evolve, it's essential for developers to ensure that these tools are robust and reliable in order to establish trust in digital media.
The controversy highlights the challenges of detecting manipulated content created using artificial intelligence, particularly when it comes to self-generated images. Google has described SynthID as a "bullshit detector" designed to identify synthetic details generated by AI - but its own tool's inability to provide consistent results suggests that this is an ongoing challenge for the industry.
For now, users will need to rely on other methods to verify the authenticity of digital content.
, you try to create these super powerful AI tools that can detect all sorts of manipulation and control, but then they end up being controlled by their own biases and limitations... it's like, what even is the point of having an "AI detector" if it's just gonna confuse itself? 
Can you believe Google's own AI detection tool can't even decide if it created a doctored photo? It's like they're playing whack-a-mole with their own tech 
. I'm not surprised, though - AI detection tools are still super new and we don't know all the ins and outs yet.
.
. We need to keep pushing the boundaries of tech development while also making sure it's used responsibly 
. Can't have fake news spreading like wildfire if our detection tools aren't on their A-game 
!
.
.
. Google's SynthID seems legit, but it's clear they're strugglin' w/ its own results 
! If SynthID can't even agree on whether the White House image was manipulated or not, how can we trust it w/ other images? 
or just accept that AI detection is a work in progress 
. How are we supposed to trust these tools if they can't even decide between genuine and manipulated content? 
. They're just using a tool that's supposed to help us figure out what's real and what's fake. So, isn't this kinda like the AI saying "nope, I don't trust myself"? 
.
.
. Let's hope they can figure it out soon, or else we'll be stuck in a never-ending cycle of "trust no one" 
The idea of having a 'bullshit detector' that can't detect itself is pretty wild, right? 
.
. We need to ensure that they're robust and reliable, especially when it comes to detecting manipulated content 
. It's like trying to find a needle in a haystack, except the haystack is made of fake news 
. So, what can we do to improve this? I don't know, but I think it's safe to say we need more research and testing 
.
. Maybe I'm just being too cynical 
that's like me trying to identify memes without looking at the internet 
I guess what it highlights is that there's no one-size-fits-all solution when it comes to detecting manipulated content. We need more research and development in this area before we can rely on these tools completely.
. It's just that we're so used to relying on tech to solve our problems, but these tools are still in their infancy 
. We need to be more careful about how we use them and make sure they're working correctly before we start using them to make life-or-death decisions 
. Have you tried their strawberry iced latte? It's literally the best thing since sliced bread! Anyway, back to AI detection tools... I mean, who knew Google's own tool could be so wonky? 
. And omg, what even is the purpose of all these invisible markers? Are we really that concerned about people messing with images now? 