Big tech using AI as a lie detector.

[Spart0n]

Summary

 Google Amazon and the Israeli government are working together on project nimbus that among other enhanced AI surveillance will be able to tell if someone is lying. We don't yet know how accurate it is though, and some employees don't like it. 

 

Quotes

Quote

Nimbus training documents emphasize “the ‘faces, facial landmarks, emotions’-detection capabilities of Google’s Cloud Vision API,” and in one Nimbus training webinar, a Google engineer confirmed for an Israeli customer that it would be possible to “process data through Nimbus in order to determine if someone is lying”. 

 

My thoughts

This is definitely a cool use for AI but definitely will be used for evil from everyone to companies to nations, but it could be nice to see game shows or Internet shows use this kind of technology for fun not just evil. 

 

 

Sources

 https://www.androidcentral.com/phones/google/google-project-nimbus-ai-surveillance-evil

 

https://techcrunch.com/2022/09/01/google-workers-protest-1-2b-project-nimbus-contract-with-israeli-military/

 

 

[PeachGr]

Google can't have a good face unlock, but it can detect your lying face!

I feel like we have much way to go before we have something accurate 

[strajk-]

26 minutes ago, Spart0n said:

definitely will be used for evil

Define evil.

Pretty much all tech advances are used at some point at the detriment or exploitation of the user or society as a whole, as long the pros highly outweighs the cons that'll be fine.

[williamcll]

Sounds like it's made to compliment advanced interrogation techniques.

[Spart0n]

2 minutes ago, strajk- said:

Define evil.

Pretty much all tech advances are used at some point at the detriment or exploitation of the user or society as a whole, as long the pros highly outweighs the cons that'll be fine.

Well like like detectors used to be considered infallible but now it's common knowledge that they were never all that accurate, evil use of this AI would be just to outright believe it at all times and let governments use it to determine guilt or innocence, or letting companies use this on employees or even something as simple as a landlord using his office security camera to look at potential renters interview answers and decide if they can have housing there based on AI telling him if they lied. All that would definitely be evil

[Kisai]

1 hour ago, strajk- said:

Define evil.

Pretty much all tech advances are used at some point at the detriment or exploitation of the user or society as a whole, as long the pros highly outweighs the cons that'll be fine.

 

Evil is any self-serving selfish motivation. Someone or some company that is "Evil" is one that only looks to take advantage of situations for personal gain. This is why many media organizations get villainized, because they have the duty to report the news, but they are under the thumb of corporate evilness to only report things that bring eyeballs for ads, not truth.

 

AI use for evil purposes would be about eliminating the human from the process. So never mind lie detecting, the AI will learn how to perfectly deepfake someone to gaslight them into believing they are lying by creating videos and audio that they didn't do and ask them to "is this not you?" An AI as interrogator would be absolutely ruthless.

 

Now let's flip the script, what if we used the AI for detecting propaganda? When media outlets are outright lying, because the people at those news shows do not even believe the lie they are selling.

 

[Doobeedoo]

Yet it won't be able to detect a psychopath since they pass basic lie detector crap, that is not even precise in general and don't change expression as you'd expect from normal people in xyz situations.

Maybe AI model will have extra algorhitm layer to accomodate for that and infrence situations with that in mind who knows. 

But yeah funny basic face detect ulock thing.

[05032-Mendicant-Bias]

Technically google can make such tool. Even I can make a tool that from a video feed that finds faces and guess if it's a liar or a truth sayer. Sometimes my tool would find lying plant pots. And truth saying baloons with smiley faces.

 

ML models tend to be pretty opaque, you never know if the model decided that a person was lying because it picked up subtle facial cues that on average are associated with lying or because there was an excess of red pixels from a plant pot in the corner.

e.g. some researcher tried to find which pixels were most important for their dog breed model to classify dog breeds. In this case, huskies were detected by focusing on the snow in the image, because training husky pictures all had snow while most other dog had no snow. That's was the biggest differentiatior for the model.

 

It's going to be hard to know what the error rate of the tool is, what the blindspots and biases are and if the people using this tool are trained in knowing the limitations of the model. Already such tools used by banks to detect loan qualification in applicants are not exactly fair.

[mr moose]

Great, another shit tool that isn't effective enough to be used unless it gets you the results you want.   Can't wait to see it being abused by idealists/politicians all over the world trying to prove something that can't be proven.

[Rauten]

Just now, mr moose said:

Great, another shit tool that isn't effective enough to be used unless it gets you the results you want.   Can't wait to see it being abused by idealists/politicians all over the world trying to prove something that can't be proven.

Or to make sure that the mandatory smiles upon entrance to the factory aren't a lie.

Joyous.

[mr moose]

1 minute ago, Rauten said:

Or to make sure that the mandatory smiles upon entrance to the factory aren't a lie.

Joyous.

 

Wife holds up phone in front of husband and asks "does my bum look big in this"?

 

10 seconds later google has caused another divorce.

[Heliian]

6 hours ago, Spart0n said:

Well like like detectors used to be considered infallible

They never were,  it was only perpetuated by the Americans. 

[BuckGup]

This sounds very anti productive. They already have lie detectors Vsauce showed years ago that you cannot fool as it watches for an involuntary subconscious reaction in your brain to stimuli you have already seen before.

[Vishera]

Looking at the sources quoted here and around the internet those articles are a clear anti-Israeli propaganda.

The AI Lie detector thingy was just something that a Google employee said is possible - Not something that is planned.

 

In fact project Nimbus is:

Quote

Project Nimbus (Hebrew: פרויקט נימבוס) is an Israeli government initiative under which cloud computing services will be provided to its agencies. The Israeli Finance Ministry announced April 2021, that the contract is to provide "the government, the defense establishment, and others with an all-encompassing cloud solution."^[1] Under the contract, the companies will establish local cloud sites that will "keep information within Israel's borders under strict security guidelines."^[2]

And they will have access to the compute and AI/Machine learning capabilities of Amazon and Google:

Quote

Under a $1.2 billion contract, technology companies Google (Google Cloud Platform) and Amazon (Amazon Web Services) were selected to provide Israeli government agencies with cloud computing services, including artificial intelligence and machine learning.^[4][1]

Source: https://en.wikipedia.org/wiki/Project_Nimbus

[Vishera]

12 hours ago, Spart0n said:

Nimbus training documents emphasize “the ‘faces, facial landmarks, emotions’-detection capabilities of Google’s Cloud Vision API,” and in one Nimbus training webinar, a Google engineer confirmed for an Israeli customer that it would be possible to “process data through Nimbus in order to determine if someone is lying”. 

That's basically Google explaining to the Israelis about the capabilities of their AI/Machine Learning and give them an example to what is possible to do with them.

[wanderingfool2]

12 hours ago, PeachGr said:

Google can't have a good face unlock, but it can detect your lying face!

I feel like we have much way to go before we have something accurate 

There's a differences between id'ing a face where the goal is to do it quickly on underpowered hardware, also not store too many id'ing things about the owner vs using motion and stronger computing devices to detect facial features.  An example would be looking at the face filters (and how they work decently well).

[SansVarnic]

-= Topic Locked =-

As interesting as this would be, this thread has gone the way of conspiratorial/political and some derailment. I don't believe cleaning this up and re-opening would accomplish much.

Edited September 7, 2022 by SansVarnic