AI - Elixis Technology

Bad AIs Eat Their Own Data

Elizabeth Technology December 10, 2021

Poorly optimized algorithmic content is frustrating for users, in more ways than one.

Ads (For Items)

It’s a new, somewhat dystopian warning: look for gifts in Incognito Mode so the ads don’t give away what you were looking at. Unfortunately, in a world run by websites that want you to make an account for your purchase, Incognito Mode is less helpful than it used to be.

Websites take notice of what you look at and buy, and then they juggle that into a measure of intent – are you actually planning to buy X item? How many times did you check it, and how long did you look at the listing? Did you look at other listings like it? Did you message the shop owner, or ask a question? Did you ‘heart’ it? If you did, it’s going to recommend more proportionally to how much you interacted with said item. But what about gifts, you may ask? How does the algorithm know I’m not buying this nurse-themed cup and this teacher-themed lanyard for myself?

Turns out any website using Google tools to track engagement knows what data to leave out in the long-term – they’re gathering so much data that it’s not really a loss! Given enough time to read your patterns, they’ll be able to figure out you’re done looking and will squirrel that knowledge away for the Gift Finder stuff (or whatever Google does with all of the data it stores on you) later. That’s… creepy, but not necessarily worsening your experience.

But what About the Ones that Aren’t as Optimized?

What is worsening the user experience is a lack of understanding context by other, less developed and less conscientious algorithms. Google Ads was notorious for following you with an item you looked at once before their target-testing showed users didn’t like it, and it was prone to mistakes anyway; companies following Google as an example didn’t always move on when they figured that out, though. Target sending coupons out for baby carriers and bottles came across as gauche, even when it was right – you hope nothing bad ever happens, but the first trimester for a pregnant woman can be very scary, which is why it’s tradition to hold off until the second trimester to start sharing that info. Imagine a company butting in with a mailed coupon and effectively telling your household that you’re pregnant before you get to!

And where ‘haunting’ a user with an item they glanced at is still popular, it can make it tough for users to go back to casual browsing without that item appearing, making a website less appealing to casually visit. For example, Etsy – Etsy does not seem to be able to distinguish between items you’d buy once, like musical instruments or coffee tables, and items you’d buy over and over, like soap and other consumables. As a result, if you buy an instrument off Etsy, you don’t necessarily get ads for items related to that instrument – you just get ads for more Instruments. Take these screenshots of my Etsy front page:

This was immediately after I bought an instrument from the shop OrientalMusic, and if this was candles or snacks or something, showing me more stuff from the same vendor would be reasonable – as it is, I can’t window-shop for stuff Etsy thinks I might like because all it thinks I might like right now are more instruments.

“Shuffle” and Spotify

Spotify allows its users to make playlists of songs, but it also attempts to generate separate playlists for the user. “Discover Daily” and Discover Weekly” are designed to show the user new (or new to them) music that they might like. “Release Radar” aims to get you to new songs from other bands in your playlist. And then there’s the “On Repeat” playlist, which is meant to play you the songs that you’ve heard most often.

The obvious issue with that: if you’re a free listener, Spotify decides which songs you’ve heard most often. If you’re a mobile listener on the free plan, you don’t have the option to not shuffle on the playlists you make, so the algorithm determining what song you’re going to listen to next is also ultimately deciding the On Repeat playlist, not you. The other playlists also learn that you like those same songs more, and Spotify’s algorithms scramble to provide you recommendations based off of the songs you like the most… the songs it thinks you like the most, which aren’t songs in the playlist but are instead songs you listened to, which Spotify decided.

Effectively, Spotify is feeding itself its own data, not yours!

Even worse, the shuffle function isn’t truly random – it’s run on an algorithm too. True randomness would be a saving grace for “On Repeat” – if you have a song in multiple playlists that you listen to often, statistically, it’ll pop up in On Repeat before songs you only have in one. That is, if it were actually random – unfortunately, it’s also decided by an algorithm. If you’re getting the same three or four songs every time you start a playlist, and the same handful the majority of the time afterwards, even with plenty of other songs in the list, that’s not a coincidence.

OneZero says that Spotify divides its functions into exploit and explore, and when it’s trying to exploit, it’s easily tricked into a feedback loop of the same music you hear all the time. Explore is in the same boat, but it uses other people’s data to suggest songs that Listeners of X liked – leading to the same conclusion every time you open the Discover playlist. If you didn’t like those songs last time, it doesn’t care – it’s recommending them again to you now because Listeners of X liked it, and you listened because the algorithm put it first in line in shuffle, which leads to it thinking you like X a lot. Wired.com says that it can get itself so stuck on what it thinks you want that trying to break out and get new recommendations in your Discover playlists is better done on a fresh account. Yikes.

Youtube Recommended

Youtube’s recommended page is usually pretty good at picking up what you’d probably want to watch… as long as it has some history about you first, and also as long as you don’t stray too far from what you normally consume. Countless Youtubers have filmed themselves opening Youtube in an incognito window so they can show how few videos it takes to get into some crazy conspiracy theory videos – turns out the Flat Earth is never more than five or ten clicks away! A phenomenon that some noted was that new accounts who didn’t have any other data would get funneled into a rabbit hole once Youtube had the slightest smidge of data about them – and when conspiracy theory videos have high engagement (i.e lots of comments arguing) and enough run time for ad breaks, they’re considered above average content. Wonder why Youtube is putting those little Context bars below videos with sensitive topics now? That’s because it was forced to reckon with the algorithm’s tendency to feed misinformation to newcomers and people who ‘did their own research’ right into believing the Earth was flat and lizard people were real.

Sources:

https://onezero.medium.com/how-spotifys-algorithm-knows-exactly-what-you-want-to-listen-to-4b6991462c5c

https://www.wired.co.uk/article/spotify-feedback-loop-new-music

Deepfakes: Should You be Concerned?

Elizabeth Uncategorized October 22, 2021

You might have seen those videos of “Tom Cruise” on TikTok, or maybe you saw someone’s face superimposed onto Superman. Deepfakes are getting better by the day!

Deepfake Software

Deepfakes are a species of visual edits that use pictures and video, combined with AI, to create something new! The AI uses a pre-existing video and a library of photos to replace one person’s likeness with another. If you have the pictures for it, you could deepfake your face onto Chris Hemsworth’s body, and other such shenanigans. And deepfakes aren’t just for videos! They can also be used to create better still images as well. Where Photoshop relies on a human’s touch to make it believable, deepfake tech can create a realistic still mostly by itself given the tools.

That’s the catch, not all deepfake AI has all the tools – some deepfakes are noticeably worse than others, for a couple of reasons. The tech is still pretty new, so most programs are still ‘learning’ what is and isn’t possible for a human face. The second issue is the quality of the images fed to the deepfake – if the images don’t give the deepfake enough information to accurately recreate angles, it’s going to have to get creative. This is a bad thing when you’re trying to make a believable video.

Celebrities Vs. The Average Joe

Deepfakes rely on data, so if the software doesn’t have much data to work with, the resulting deepfake looks…uncanny. Even really, really good deepfakes right now, with a ton of data, look a little uncanny. Picture the last movie you saw a dead celebrity in – you probably realized something was wrong even if you didn’t know they were dead, like General Tarkin in Rogue One. He’d had his whole head scanned at high quality before he died, and he still looked a little strange on-screen. It was little things, like his neck not moving perfectly with his mouth. Young Carrie Fisher at the very end of Rogue One had a noticeable grain due to the source images, and that same young Carrie Fisher in The Rise of Skywalker looked strangely plastic even in low, indirect light.

The average person doesn’t have enough high-quality video or images from even one angle for deepfake AI to make something believable. It only takes a split-second of slightly misplaced nose or mouth for someone to get creeped out by whatever you’re making and identify it as fake. The uncanny valley is instinctual, but it’s reliable! It takes serious work to overcome that instinct. If Hollywood can’t manage it, is there anything to worry about for the average person? Well… yes. Because the average person has access to it, and the tech is always getting better.

Controlling it

How do you control it? Big stars have to deal with their image being stolen all the time. If anyone’s prepared, it’s the celebs, who have to fight magazines and movies alike to be represented like they want to be. But what about the average folks when it starts to bleed downwards? Minor politicians, or competition for the cheerleading squad? Or explicit images made specifically to harm someone’s image, made by an amateur with juuust enough knowledge to make something that, at first glance, looks believable.

How do you account for that?

Lets look at the Tik Tok Tom Cruise account. The creator has gone out of his way to make it clear that Tom Cruise’s likeness there is not real. Even so, the videos are jarringly realistic. He used a Tom Cruise impersonator as the ‘base’ for the deepfake, and the end result barely catches any uncanny valley at all. He just looks a little stiff. That guy’s videos are still up, because it’s obviously not really Tom Cruise no matter how realistic it is.

And then there’s an account that’s putting Charlie D’amelio’s face on their own body, in an attempt to impersonate her. Tik Tok is removing these because it’s not obvious that it’s not Charlie, even though the quality is worse. Someone who watches it more than once is going to recognize that it’s not Charlie, but it’s still getting pulled, because it’s not being clear enough. They are crossing a line.

There’s also a distinction between the two for intent. ‘Tom Cruise’ is showcasing his technical skill, the Charlie impersonator is trying to be Charlie.

Legally, copyright law does have some precedent from: if an the music and art world: if an impersonator is so close in performance to the original that an average person can’t distinguish it from reality, then they’re violating copyright. Singers use this when covers get a little too close to the original. See Drake songs, for instance: the only covers you’ll find on Youtube are by female singers or men who sound totally different, because he’s very strict on his copyright. When the audience can’t tell them apart, they’re pulled.

The problem is enforcement. The average person is not going to have to time or resources to hunt down impostors and report them all. Charlie is famous on Tik Tok, but if she wasn’t, Tik Tok mods likely wouldn’t actively hunt down these impersonator accounts for her. If someone really, really hated an obscure user, they’d be able to overpower their reporting efforts with fake content, and that fake content only has to be believable enough for someone to scroll past it and think “wow, I can’t believe they’d do that”.

The average person is not equipped to scrutinize every single little bit of media that comes their way, it’s exhausting and unrealistic to expect that of them. It’s how disinformation campaigns work. If the deepfake is believable enough, and the original’s not aware of it, that deepfake may as well be fact for everyone who sees it and doesn’t realize it’s fake.

Implications

If you’re online a lot, you might have heard of that new Mountain Dew ad featuring Bob Ross’s likeness. This was… weird, to a lot of people, and for good reason. Using a person’s likeness to sell something has been a matter of debate ever since money became mainstream – you’d probably sell more spices if you said the king bought from you back in BC times. But normally the person is able to call them out for it. Now, with deepfakes, you can make celebrities say anything post-mortem, and nobody but the estate will be able to challenge it.

And, even if the estate gives permission, how specific do you have to be about that image? Actors struggle with Paparazzi images even today – Daniel Radcliffe famously wore the same shirts and pants for weeks while filming a movie, so the paparazzi’s images of him were worthless. Imagine having the ability to put Daniel Radcliffe in any pose or outfit you wanted for the front of a magazine. The person wouldn’t make unflattering faces for your pictures before they died? Well. Now they will.

Presumably Bob Ross’s estate allowed the use of his image, but in the same way we don’t take organs from dead bodies without consent of the deceased, maybe we shouldn’t allow the selling dead loved ones’ images for advertising purposes, without their consent beforehand. Especially now, when it’s easy to deceive people with this tech!

Is There Good?

And then there’s the other side of the spectrum, where deepfakes can be used to bring people back to their glory days, or color black-and-white movies. They can be used to de-age actors, as seen in Captain Marvel, Star Wars, etc. Samuel L Jackson was 40 years younger thanks to deepfake tech, and Mark Hamill appeared as he was forty years ago for another Star Wars series.

Deepfakes, given the tools, do a better job of recreating someone’s face than human-controlled CGI ever could. They could have been used to make Henry Cavill’s Superman in Batman Vs. Superman mustache-less, instead of whatever they did instead that made his face look unsettling. He couldn’t shave his ‘stache because he was also filming Mission Impossible at the same time, so the only way out was either prosthetic facial hair, or CGI-ing over it. They picked the CGI. People noticed. Deepfake tech might have made his mouth’s movement a little less uncanny.

Deepfake tech could be used to disguise facial injuries, like Mark Hamill suffered during the original Star Wars trilogy, or create alien races without the heavy prosthetics traditionally used or sweatshop CGI-studio labor. They could make dubbed movies less visually jarring, and line up actors’ mouths with the words they’re supposed to be saying.

Deepfake technology is a very double-edged sword. All the good it could do isn’t outweighed by the bad. It’s dangerous technology, and in a world that’s increasingly using the internet to share information, disinformation is a powerful pollutant.

Sources:

https://www.theguardian.com/technology/2020/jan/13/what-are-deepfakes-and-how-can-you-spot-them

Captchas – How and Why

Elizabeth Uncategorized September 29, 2021

Captcha, which stands for Completely Automated Public Turing test to tell Computers and Humans Apart (what a mouthful) was first conceptualized in the early 2000s. Websites were already struggling with bots, and a website known as iDrive recognized their inability to ‘see’ the way people did. Paypal, also struggling with bot attacks, began using the same method to keep brute-force attacks from getting in. This is the true essence of Captcha – in 1997, the tech was first described as anything that could differentiate robots and humans, but it wasn’t known as ‘Captcha’ until Paypal got in on it. It’s more advanced form, the reCAPTCHA, was first coined in 2007 and then absorbed into Google in 2009.

Type These Letters, Hear These Sounds

The original style is becoming easier to get past as AI improves, but it’s still better than nothing. An AI would still leave clues that it was ‘reading’ the letters (or trying to) as it tried to decipher the captcha text from the other random lines and fuzz on screen – Cloudflare, a security website, notes that AI couldn’t do much better than keysmashing and hoping to get in that way when this was first implemented. Now that AI can ‘see’ much better than it used to thanks to endless training to recognize text out in the real world, it gets more and more accurate for every captcha box it sees. Captchas may be algorithmically generated – AIs designed to account for algorithmically generated content in front of them are now capable of deciphering the text, and Captchas are actually sometimes used as tests!

That doesn’t mean they’re obsolete or useless for protection. Just because some people can create AIs that can get past it doesn’t mean that everyone can. Many basic bot creators would much rather go to an easier, less-well-defended site than sit there and try to program an advanced, specific AI for such a simple task. It’s not perfect protection – no protection is.

However, there were problems: unimpaired users often complained that solving them was hard. For visually impaired or deaf users, the captcha might genuinely be unsolvable. Screen readers, a common tool for blind folks who use the internet, allow them to browse the web by reading the page out loud. Because a captcha is a picture, not a text box, the screen reader doesn’t know it’s there. Accessibility software is often simpler than cutting-edge bots (and incapable of reading images), and so they were left behind.

Audio versions are a better solution, but their nature still makes it difficult for screen readers to ‘see’ the play buttons. Besides, audio-to-text AI was already more advanced than picture-to-text because there’s a market for automated captions and auto-transcripted phone calls. Transcription software has been around for ages, and it only gets better at separating noise from information as time goes on – there is almost nothing a captcha could add to the sound to make it hard to interpret for a machine and not a person. As such, these captchas are less common than the fuzzy text and image ones still seen everywhere today.

“I am Not a Robot”

One of the simplest types of Captcha is the “I am Not a Robot” check box. It seems like it could be easy to trick it – and it sort of is, but it’s not a walk in the park. The box works by tracking cursor movement before the user hits the little check box. On a desktop, an AI might jump directly to the box it needs to click, with no hesitation, or it might scan the entire page to locate the box visually if it’s unable to detect the clickable element. That’s not human behavior – people don’t ordinarily need to select the entire page and then contemplate it before clicking the right area. People are also generally unable to jump directly to the clickable elements as soon as the page loads, even if they’re using the tab keys or a touchscreen device.

This was easily one of the most user-friendly kinds of Captcha out there. No reading. No listening. No selecting blurry images or trying to guess at misshapen letters. As such, it was quicker to use than a number of other types of captcha tests were, even though someone with a lot of time and determination could rig something up to bypass it.

Click These Pics

This is the previously impossible barrier that stopped AI dead in its tracks. Training an AI to see and recognize like humans do used to be impossible, but now… now it’s on the horizon. Self-driving cars will need it. Google uses it for reverse-image search. Facebook uses it to find you in friend’s photos. If humanity was going to truly master AI that behaved like people did, AI was going to have to learn how to see; that meant other, outsider AI would also be learning how to see.

The pictures are easy – you get an image separated into 9 or 16 tiles, and you select imagery that matches its request within those tiles. An AI might be able to measure ‘red’ in an image, but the sort of uncomplicated AI that most amateur hackers could crank out wouldn’t know a fire truck from a stop sign. Even if it got lucky that one time, other human users are picking all of the right squares every time – so if it misses even a sliver of the red in another tile, or over-selects, it doesn’t pass and has to go again.

Is that… being used for something?

Google is using Captchas to crowd-source training for their AI. However, doing this meant that Google had access to a metric ton of training time – Wikipedia claims that people around the globe spend 500 hours completing CAPTCHAs every week. Unlike those text and audio ones, pictures with the features they need can’t just be generated indefinitely. If you’ve noticed a decline in picture quality for these captchas, you’re not alone. The quality really is getting worse. The sharper pictures are already trained into the database, so now all that’s left is the blurry, fuzzy, poor-quality ones everywhere else, the ones that weren’t ideal for the initial training.

Now, millions of people every day are telling the computer what a red car or a street sign looks like, instead of just a large handful of researchers. Some of this research is for smart-car training, some of it’s for reverse-image searching, some is purely to advance the state of AI – once AI can recognize things in its environment visually, it can usually behave with less human intervention. And the more training it has, the less likely it is to become confused at a really inconvenient time. Tesla has famously struggled with AI mis-recognizing things, such as the moon, blinking streetlights, and partially graffiti-d signs, but the more training it gets, on worse and worse quality images, the better it will eventually perform.

Sources:

https://www.cloudflare.com/learning/bots/how-captchas-work/

https://support.google.com/a/answer/1217728?hl=en

https://googleblog.blogspot.com/2009/09/teaching-computers-to-read-google.html

https://elie.net/publication/text-based-captcha-strengths-and-weaknesses/

Car Screens – Is It a Good Idea, Really?

Elizabeth Uncategorized September 10, 2021

We all know how addictive screens are. And yet, after endless campaigns to get teenagers to stop staring at their screens while driving, we’re introducing cars that practically require it. Why?

The Good

Screens exist basically everywhere. They’re oftentimes a good substitute for analog buttons, as in the case of phone keyboards, and can provide more flexibility and wear-time if the screen is going to be in front of the public, as in self-serve checkout screens in grocery stores. They’re easier to clean and more difficult to break.

However, screens and analog buttons don’t have to be enemies. Some modern cars come with air conditioning that can be set to an exact degree to aim for, but the number of possible answers means a digital readout is required. It gets both a screen and a set of buttons.

Other things necessitate screens if the customer wants them as a feature. You don’t have to have a screen for the radio, the buttons could perform all of the functionalities just fine, but if the customer wants to know the temperature outside, that takes a screen. Plus, that screenless radio would be annoying to select presets for, so they almost universally come with some sort of screen or indicator for the channels. Bigger, more complex car radios with screens can show more information about the broadcast, too!

Some features that help with safety and ease-of-driving come with screens too – backup cameras need a screen to function. There’s no way for that feature to exist without a screen somewhere, so it may as well be in the dashboard of the car.

The Bad

That being said…

Some things are better suited for buttons and physical inputs. Someone can adjust their volume by simply grazing their hand along the surface of the interior dash until they get to the right knob. Trying to do that with analog buttons and a digital readout is also doable – they will be able to both feel and hear the difference (of switching stations) in hitting different buttons, so they’ll eventually be able to land on the right one, as long as their fingers are on the buttons. Doing it with no physical feedback requires taking eyes off the road, otherwise the user doesn’t know if their fingers are even on the buttons on-screen.

Extra features that are useful are also often distracting while on-screen, so it’s not totally the screen’s fault. GPS hooked into the car’s screen makes sense. It’s safer than looking down at the cupholder or blocking off a bit of vision for a suction-cup phone holder on the windshield. However, typing on one is usually a nightmare because the positioning is awkward, right in the middle of the dash, even if the screen is top-of-the-line responsive. Syncing to a phone to use the GPS there, and then BlueTooth it to the screen fixes that problem but creates new problems in it’s wake. Even worse if these things are in separate menus, which means spending time navigating said menu to get to the GPS, Bluetooth hookup, or other assorted features in the first place. All of that should take place before driving – but isn’t it annoying to have to fix all of that up before even leaving? Flipping through the radio was effortless before screens made it more difficult than it needed to be.

Deeply Unnecessary and Largely Unwanted

Bizarrely, automakers also offer options to connect to the internet for reasons beyond simple GPS or music. As The Turning Signal points out, the layers upon layers of menus and features offer a lot of distraction while in the car, and no hierarchy of features. Radio should probably be fewer steps to get to than GPS, for example, because you’re not going to use GPS on every trip you make, but radio is almost always on. Another obvious downside is that if anything goes wrong with the screen itself, you’re trapped with the settings you had when it broke, and that’s really annoying.

Part of this isn’t even due to the screens – it’s because the automaker is desperate to stuff as many features as possible into the car. The sheer number of things a car can do now means even if everything were analog, the user would still be glancing down pretty often just to find the right button for the task. Seat warmers, directional AC, GPS, motorized seats, built in chair massagers??, the heater, turbo heating or cooling, the radio, Bluetooth, etc. etc. would all need their own buttons – multiple buttons for each. If automakers were to make these all real, physical buttons, your dashboard would look like something from Star Wars. It’s too late to go back unless the automaker wants to ditch features that other cars (and their previous cars) still have.

Even Worse

Ford announced plans to beam billboard information directly onto the screen, via a complicated system of computers and AI. While it’s not literally beaming every sign it sees into the car, and it is theoretically possible to shut off, it’s still an awfully ugly statement. The dashboard has become advertising space for billboards that used to be ignorable. A big question will be how it interacts with other apps on-screen. Does it get priority over the radio, or the GPS? Even assuming that’s all sorted, and the customer willingly has the ads open, glancing down to peep at a flash on-screen is a little bit dangerous, is it not? Their reasoning is that the consumer may have missed information they could be interested in. If the information is interesting, that’s worse! That makes the distraction issue worse! The screens are already horribly distracting as they are, with all of the menus and buttons and stuff to dig around in, so having an ad, which is inherently trying to snatch your attention away from what you were doing before you saw it, beamed directly into the car while the driver is driving, is effectively putting revenue above safety. I thought Ford had learned from the Pinto. Apparently not.

Many people jumped on Ford for even suggesting the option. As they should – billboards themselves have gotten into trouble for being too distracting, how beaming directly into the car was supposed to avoid those same issues is anybody’s guess.

And then there’s things like games and social media apps built into the system. It’s weird anyway, because most people have phones, but whatever. Assuming it has the most basic of safety features built in, and won’t activate if the car is in drive – what’s to stop the driver from shifting into park at every red light to check up on their accounts?

Phones can at least be stuffed into pockets – this screen would have to be disabled.

Sources: https://www.motortrend.com/news/ford-billboard-ad-patent-system/

https://www.theturnsignalblog.com/blog/touch-screens/

https://www.motorbiscuit.com/why-are-automakers-replacing-buttons-with-touchscreens/

https://gizmodo.com/get-ready-for-in-car-ads-1846888390

https://newsroom.aaa.com/2017/10/new-vehicle-infotainment-systems-create-increased-distractions-behind-wheel/

Stop Hyping Autopilot

Elizabeth Uncategorized September 8, 2021

It’s not done yet!!

Tesla’s autopilot is really impressive. It’s just not done yet. Between failure to detect real objects and detecting ghost objects, the new Auto-pilot has a lot of really terrifying anecdotal cases.

A Word of Disclaimer

Tesla does tell users not to get in the back seat or otherwise take their eyes off of the road while autopilot is driving. They’re constantly updating their programs to include edge cases discovered on the road, and it’s really hard to do that if the car never gets to use the feature that’s causing bugs. However, I’m not sure it’s impossible to catch some of these user-reported issues in a testing environment. Elon Musk’s consistent belief that people will die for science is not comforting in this situation.

However, many of the issues in the following article are rare, fringe-case scenarios. It doesn’t represent the cars as a whole, it’s more of a warning – you really can’t trust the autopilot 100% yet, because users report multiple different issues stemming from the programming. Nothing most Tesla owners don’t already know.Drive without autopilot or drive while paying careful attention to the autopilot, and Tesla’s as good as any other car.

The irony of using cars out in the wild to ‘test’ is that a regular car’s cruise control is actually less stressful – the driver doesn’t have to pay active attention to the car’s surroundings on regular cruise control! The old-style cruise control couldn’t make the car suddenly brake or swerve into another car.

The Brakes, the Reads

Speaking of which, the brakes! A car capable of braking can brake itself into an accident in a split second on busy roads if it sees something it thinks is dangerous.

This is a cool feature, but it’s not done yet. Reddit’s Tesla subreddit has numerous accounts of the brakes engaging for little to no reason: phantom animals, suddenly ‘seeing’ a stop sign on the highway, misinterpreting special vehicles’ rear lights, and more. The biggest one is phantom overpasses, where it misunderstands the shadow as a reason to stop (users say that this was an older version of the software, and that newer ones don’t do it as much unless there are other, compounding factors, like tow trucks or construction lights. Still not ideal).

Nature released an article detailing how someone could hypothetically trick the car into seeing a speed limit sign instead of a stop sign, and get it to accelerate into an intersection. Specially painting trucks and cars so that the AI misinterprets what it’s seeing might turn into a great way to cause accidents. The AI seeing things is trying it’s best to look for issues, but as Nature describes it, AI is often ‘brittle’. The computer’s not totally sure what it’s looking at, so it makes its best guess. Unfortunately, it’s best guess is often pretty bad. A computer’s best guess as to what a food truck with a hot dog on top is might be that the truck’s actually an overpass, or maybe a deer, while even a baby human can tell it’s some sort of vehicle. Fringe cases like the hot-dog truck have to be manually added to the computer’s repertoire so it doesn’t freak out next time it sees it. However, it has to do this for each instance of a ‘hot dog truck’ it doesn’t recognize. Dale Gribble’s famous ant-van would confuse it too, for example, and it’s not hot dog-like enough for the AI to snap to that memory. It would be starting from scratch, every time.

It also occasionally fails to brake or move when there is something there. Commentors theorize that the computer is deliberately programmed to ignore things along its sides, so it doesn’t freak out about the railings and concrete barriers that run alongside highways.

The Lights and Cameras

Tesla’s auto-pilot is easily confused by wet road surfaces. One user reported that their Tesla couldn’t understand reflections from signs, or wet ground. It would see it’s own high-beams in the reflected light, and lower them automatically. And then it realizes it’s dark once it’s past the sign, so it flips them back on. It keeps doing this until it has a continuous level of darkness or brightness in-line with what it’s expecting from a dry road with few signs. Unfortunately, that means the car has to make it to an area with streetlights or other cars for it to figure out the low beams should be on, not the high beams. Or the user can flip it manually, which means turning off the autopilot, on some models. Speaking of light, it can’t always tell that lights are lights and not more white lines.

It also struggles with overpasses – it doesn’t understand bridges, and there are so many bridges, overpasses, and assorted vertical shadow-casters that distinguishing it from a regular stoplight pole is a Herculean challenge. As such, it often erred on the side of caution before reprogramming fixed its confusions.

The built-in monitor can also display what the camera thinks it’s seeing, which gives the user some valuable insight into how it works. When it pings something as a thing, that thing is there now. See this gif of someone driving behind a truck with stoplights on it:

This is a hilarious edge case, and I don’t blame the car for not understanding what’s happening, but the lights stick to the place in the road where the Tesla identified them. Once it’s there, it’s there – a box or bag in the road that’s incorrectly identified might not get re-identified correctly. Of course not! Because if the Tesla was told to constantly re-ping it, it might misidentify things it got right the first time, and the more opportunities the programmers give it to do that, the more likely it is to happen. Right now, what Tesla has going on is good for ideal conditions. The struggle is getting all of that to work in the real world.

The Hardware

The cameras are great. This issues with the autopilot are purely AI-driven. The flash memory used in older models was prone to failure and had to be treated like a warranty item to avoid a total recall, which sucked for users, but otherwise – the hardware directly tied to software functions is more or less working as advertised. It’s the other parts of being a car where Tesla falls down.

It’s unfortunate, but Tesla’s ‘Model S’ front axels are prone to deforming. It doesn’t happen quite often enough to warrant a recall, but enough for some disgruntled users to post about it online. Something as simple as driving onto the curb bends the front axle, and the user then starts to hear strange noises from around the wheel area when they turn. Many Tesla superfans attribute these complaints to one guy in Australia harping on it, but scattered posts (from various devices, locations, and dates) across the Tesla subreddit as well as Tesla forums suggest this is a bigger issue than those superfans (and Tesla) want to believe. Tesla revolutionized electric cars, but it also re-did a lot of design work itself, from scratch. Is it really that unbelievable that cars across nearly a decade could be suffering from a premature parts failure? It happens to non-electrics all the time!

Design

Also, from a design standpoint, I just… don’t think the cyber-truck looks that good. The previous four-door Teslas look great! They’re very slick, but they look a lot like some of the hottest cars in the market. A family car, or a commuter car. It blends in with the pack, and only stands out in traffic in good ways, like it’s lack of noise. The cyber truck looks nothing like the trucks it’s meant to compete with. The sides of the bed are raised so it meets the rest of the body on a nice, straight line. That sure looks cool, but for anything of actual weight, the driver can’t toss items in over the side. That’s one of those minor-but-annoying things that peeves owners off over time.

The glass is also armored, which is cool, but… what for? Who is driving this? Who’s afraid of getting hailed on or shot at, and doesn’t want a less conspicuous vehicle? Or, the inverse – bougie celebrities with a lot of money and a lot of enemies might want a really conspicuous car but with stronger glass. Does the cyber truck do that? Kinda… but so do many sports cars.

It’s a cool idea, but it’s just that – an idea. The truck of the future, not the truck of right now. An electric truck is a great idea! But it doesn’t look anything like other company’s versions of the same concept does, so people may be reluctant to jump to Tesla, instead of Ford. Differentiation in cars can either give you the VW Beetle, or the Pontiac Aztec. Only time will tell how the cyber truck fares.

Sources:

https://www.tesla.com/cybertruck

https://www.nature.com/articles/d41586-019-03013-5

https://forums.tesla.com/discussion/60330/model-s-axle-problems

https://www.nature.com/articles/d41586-019-03013-5

https://www.forbes.com/sites/bradtempleton/2020/10/23/teslas-full-self-driving-is-999-there-just-1000-times-further-to-go/?sh=7c7734c32ba6

AI: You Get Out What You Put In

Elizabeth Innovation, Technology June 7, 2021

AI needs training to understand what it’s meant to do. The quality of the training determines its outcomes.

Tay Fed Tweets

Microsoft’s Tay was exposed to the worst of the internet at incredible speed. Once Microsoft announced the project, Tay began forming her own tweets out of content she was sent, and it went about as well as you’d expect. Racist, sexist, anti-Semitic language ruled her feed, and she was shut down shortly after. This is an unfortunate experiment, because Tay might not have turned so quickly if she’d just been exposed to open Twitter anonymously. There are still racist tweets out on Twitter, but being targeted by the mob produces a disproportionate amount of ‘bad’ tweets towards the target. If they didn’t announce Tay’s existence, they wouldn’t have gotten as many messages, though. Knowing what the experiment is allows people to screw with it in the same way it allows them to participate.

The experiment was still considered a success: Tay took exactly what she was given and rebuilt it in new, interesting ways. Racist new ways, but still new ways. A machine was able to successfully learn how to make a targeted threat towards another Twitter user. That was huge for machine learning. Of course, Microsoft doesn’t exactly want that to be the face of their machine learning program, so Tay was reset and reinstated with filters in place, under different names.

This is a key experiment, not only because of how fast Tay keyed on to how she was ‘supposed’ to behave, but also because it highlights issues with reinforcing learning in the active environment. AI may end up learning things it’s not supposed to, to the detriment of the environment it’s supposed to be part of!

Google Deep Dream Fed Eyes

Google’s Deep Dream software was famous when it first reached the public. It was fed pictures of living things, and so living things were all it could see, anywhere. Everything it touched was coated in eyes and fur-texture. It was horrifying. Deep Dream pictured all sorts of eldritch horrors in everyday items – it was doing it’s best, but all it knew where pictures of dogs.

Google fed Deep Dream a set of images from a database assembled by a university, but it wasn’t given all of the images, since that would have been a huge amount of data for the then-small Deep Dream. Instead, Deep Dream consumed the library in smaller pieces, and one of those pieces was a very fine-grained sub-library of images of dogs. Deep Dream’s specific instructions were to take the picture – illustrate what it saw – repeat. Little aberrations slowly turned into eyes and whiskers.

Since then, Deep Dream has added filters that allow users to pick which database of images they want to use, each of which creates a new, wacky image out of their own uploaded images, but the dog filter still sits strong in people’s favorites. Sometimes things are so creepy they’re cute! The story of an AI chugging along and doing it’s best is one for the ages.

Art-Breeder

Art Breeder is an AI-powered character creation tool, and it’s already been covered by some of the largest YouTube channels on the website. Art Breeder breaks down human expression inputs into granular emotions, such as happiness or fear. Using sliders, the user can then alter a preexisting image, or create a face out of thin air! Art Breeder uses it’s database to put together what it thinks happiness or sadness does to the human face. It’s difficult to get a perfectly realistic human face – most still appear animated or cartoony – but it’s also frighteningly easy to accidentally create a blob-monster.

Art Breeder’s AI doesn’t actually know what it’s seeing, it’s just doing it’s best based on pictures it’s been fed. It doesn’t know that glasses are separate from a person’s face, for example, or that mouths don’t have four corners for smiling and frowning at the same time. It also doesn’t necessarily understand freckle patterns, or where blush belongs. Art Breeder’s fascinating. It can make faces, or it can make motorcycle accident victims, all with a single mis-click of a slider.

AI-Dungeon Fed Fanfiction

AI-Dungeon, a young and upcoming Steam game, made some waves when it announced that it was trying to fix issues within its script generation – especially the AI’s tendency to get explicit with named characters. Why, you ask? The source AI (which AI Dungeon uses) was partially trained on fanfiction alongside Wikipedia and assorted other text sources. Fanfiction, for those of you who don’t know, is fan-written fiction about popular media.

Fanfiction is great because it can go absolutely buck-wild with the content: characters are put into alternate universes where a certain character’s death doesn’t happen, or maybe the entire cast is working in a coffee shop. Maybe two characters end up in a relationship even though nothing in the canon of the work suggested that could happen. It’s a great place to start for aspiring writers, since the characters are all already written – all that’s left is to put them together in a different way.

Unfortunately, a lot of fanfiction is… explicit, so filtering is very necessary. Feed AI Dungeon explicit content, and it will attempt to recreate what it was trained on, which was an absurd amount of explicit content mixed in with all the general-audience and PG-13 rated content the developers wanted to use.

The worst part is not the explicit content, which is allowed – it’s that the machine didn’t know it was only supposed to apply to adult characters, which ended up creating some very awkward, uncomfortable content for people who discovered the flaw. As such, they’ve updated their reporting system to keep that from happening again, and the Dungeon is now allowed to auto-flag itself if it spots content it’s not supposed to be making.

Potential for Racism

Unfortunately, training facial recognition software with mostly white people means that the computer only understands white faces, and it doesn’t have sufficient training in other areas as a result. When the only bird you’ve seen is budgies, every budgie looks different – but all cockatiels look the same until you’ve gotten more experience handling them. AI isn’t being given the necessary experience, and as a result it’s flagging Black and Asian men disproportionately.

It’s happened before. It will continue to happen unless steps are taken to prevent it. All it takes is a mistake that the human handlers either don’t catch or deliberately ignore, and an innocent person is a suspect where they otherwise wouldn’t have been.

Interestingly enough, this phenomenon is also identified in more primitive AI, the kind that soap dispensers and automatic door openers use. “Racist soap dispensers” sounds like political fluff, but it is an issue: the dispenser is programmed to detect white hands. It doesn’t know if it’s supposed to respond to darker palms, so it just… doesn’t. Older styles that relied purely on movement were actually doing a better job than the kind that’s supposed to identify hands to dispense. Exclusion may be an accident, but its result is still the unfair treatment of different races.

The biases of the researcher are reflected in their research, and if they don’t notice it themselves, they may believe they haven’t had a hand in how the data was collected, or how the questions were chosen. That’s why it’s so critical to test, re-test, and re-re-test experiments. Biases are difficult to rule out, but not impossible. Don’t assume that a machine is perfectly logical and always right: it was made by humans, after all.

Sources:

https://www.fastcompany.com/3048941/why-googles-deep-dream-ai-hallucinates-in-dog-faces

https://www.artbreeder.com/

https://latitude.io/blog/update-to-our-community-ai-test-april-2021/

https://www.theverge.com/21346343/gpt-3-explainer-openai-examples-errors-agi-potential

https://www.nytimes.com/2020/06/24/technology/facial-recognition-arrest.html

https://www.nature.com/articles/d41586-020-03186-4

https://www.nature.com/articles/d41586-019-03013-5

Censoring Image Info: Do it Right

Elizabeth Cyber Security, Uncategorised February 3, 2021

Redaction.

Once upon a time, I stumbled into a forum thread about image censorship. The forum was made up of clipped images of funny Facebook posts, and at the time people were beginning to realize that you can’t just post names online willy-nilly. Censoring out the names attached to the posts was a requirement, and there were many ways to do it, but some of them could be undone.

What is Censoring, or Redacting?

That’s questionable. Merriam Webster gives a run-around definition, where censorship is the act of censoring, and censoring is the work of a censor, so I’m having to shave off the little bits of definition I got from each of those steps to make a cohesive definition here. Censorship is the act of keeping need-to-know info out of the hands of people who don’t need-to-know. This information could be moral (censoring swears out of public TV shows) private (people’s faces) or some other sort (no free branding). This isn’t a perfect definition, but it’s enough for this limited article.

The same goes for redaction, but with a little more intensity – need-to-know info has to be shared, but it could put people or property in danger. The easiest way to share that info without putting people in danger is to make them anonymous. By my own example here, redaction is the act of cutting out specifics (and anonymizing people) so the information can be shared.

People can guess – Tom Clancy is infamous for connecting dots to write what-if stories about redacted info – but the info is more or less anonymous to the general public.

Pixelate

Many choose pixelating over other methods of image redaction because it’s less harsh to the rest of the image, and destroys more than most kinds of “smooth” blurring. A lot of people can still make out what brand of soda a pixelated can is, and context will usually tell people that an obscene gesture is what’s behind the boxes on a TV show, but in general it works pretty well to get rid of the finer details that could identify somebody. More or less.

As machines get better and better at identifying patterns and finding the stop sign in Captchas, the human face is easier and easier to recreate. Gizmodo has an article on the subject here, and it’s a good demonstration of why – when the info is really important – it shouldn’t be used. Picture this: you have a 10,000-piece puzzle, most of it is one color, and you don’t have a box to look at. You do your best, but end up with a blob. This is early computers trying to un-pixelate an image.

It was great! It was very difficult to decipher who a protected witness was.

Then, further down the line, you get the box, and a set of glasses that lets you distinguish colors better – turns out that one color from before is actually like 30! So you get to work piecing it together. The box is blurry, so that’s a bummer, but other people with a completed puzzle can show you theirs. And someone posts to your database/puzzle forum an image very similar to the parts of the puzzle you’ve already completed. Suddenly you’re able to finish decoding the image for what it is: a human face.

That’s where we’re at right now. Pixelating the face of someone in the background of a TV show likely won’t lead to anybody going through all this effort to find them, but it could turn into a problem for folks being pixelated out of compromising images, court hearings, interviews, etc. where it’s very important that they aren’t found.

Text is even easier: picture the scenario above, but you know what letters are, there’s only two or three colors even with the glasses, and the puzzle’s only about 500 pieces. Don’t. Pixelate. Text. There’s a reason that governments go the permanent marker route. This article here does a great job of describing the undoing process.

Blur

Blur is very similar to pixelating, in a lot of ways. The pieces to the puzzle are much smaller, but you should begin to see a pattern with algorithmic censoring: once somebody knows how to do it, it can be undone. Fortunately, most people using it for important things know to go so hard on the blur factor that the image could have been a lot of things (or people), and poorly written AI can confuse matters further. Algorithms to undo blur aren’t perfect, so creating a face out of nothing doesn’t mean it’s the right face.

For example: a blurry picture of Barack Obama. It’s blurry and pixel-y, but still clearly former US president Barack Obama. The computer, instead, turns him into someone else, noticeably whiter. In a perfect world, databases would have perfect access to the entire population, but they don’t. They have access to what the researchers and engineers feed them. If your goal is to keep people from discovering someone’s identity, but you don’t want to slap a blackout square on their face, blurring is still a choice. Just make sure it’s too blurry for both people and machines to make out. Obama in this image has not been blurred nearlyenough to thwart human eyes, even though the machine can’t figure it out. As a side note, this is a great example of why facial recognition technology is too immature to use right now!

Black Out (And Sticker) Redaction

From Sci Fi shows to taxes, redacted documents pop up frequently. Completely covering text in a document with black ink or unremovable black squares should completely destroy data. It’s a government favorite for that reason! As long as it’s done right, the info is lost.

The problem is doing it right.

The American Bar Association has a blog post on the matter here. A failure to completely redact information digitally led to the case falling apart. Separately, the US government got into some hot water with the Italian government a while back over a document with information in it they were not supposed to see, including names of officials and checkpoint protocols relating to an Italian operative’s death in Baghdad.

The Stickers

In less serious stakes, digital stickers can be imperfect depending on the app used to place them on the document, but that’s more of a .png problem than a problem with sticker apps. Since these are mostly used to post funny exchanges online, rather than conceal government secrets, bulletproof security is normally not necessary. As such, you should treat them that way: security is not their main goal. Don’t use them for tax forms.

Additionally, printing the page, marker-ing over info to redact it, and then scanning it back in is an option if you truly don’t trust digital apps to completely destroy the data. It’s tedious, it’s annoying, and

it requires a scanner, but it’s an option. This is also not infallible, because even in real life things can look opaque when they aren’t. Kaspersky made this image with a digital marker, not an ink one, but it’s still a good demo. Use something marketed for redacting, not just some Crayola water-soluble marker.

Side Note: Government and Redaction Programs

Sometimes art programs store images in layers. Sometimes checking a PDF for redactions means making the redactions not permanent until publish. With these two problems in mind, mistakes like not merging layers, or using a program that doesn’t actually remove the text (as in, you can still copy it from behind the box) are somewhat understandable. That doesn’t mean it’s not a huge mistake. Redaction is there for a reason.

A major program flaw leaked government secrets. Users could simply copy the text behind the box, like it wasn’t even there. Why would you ever leave the text intact when that’s exactly the opposite of what it was advertised for? It wasn’t an isolated incident, either, as you can see mentioned above with the ABA and the Italian case. Other ways to unsuccessfully redact include putting a vector of a black box over the information in Word and cropping the image in an Office program. The entire picture’s still there, it’s just hidden, not destroyed. Don’t do that.

Swirl Redaction

Swirl is the worst of all of these options unless the others are executed very poorly. Besides being the ugliest option, it doesn’t do a good job at destroying information that other computers could use. Another algorithm doesn’t need to make assumptions like it would for pixelating. All of the information is still there, just stored in the shape of a crescent. That’s it. The algorithm stretches the image, and then warps it around a central axis, but everything is still there. See the side note below on the Swirl Man who assumed he’d done a good enough job of redacting his face. Now that this cure for swirling is out there, it’s basically obsolete.

Side Note: They Caught The Guy

A while back, police caught a child trafficker. He only hid his identity by swirling his face. Swirling, like any other computer effect, uses an algorithm. Algorithms follow rules.There’s a clear pattern in the swirling that can be undone to retrieve the original image. Simply knowing what tool he’d used was enough to reverse-engineer it and undo the face swirling. He was caught, thankfully, as a result of his own hubris. Here’s the Wikipedia article on his case and capture.

Sources: https://www.makeuseof.com/tag/easily-pixelate-blur-images-online/

https://stackoverflow.com/questions/4047031/help-with-the-theory-behind-a-pixelate-algorithm

https://en.wikipedia.org/wiki/Pixelation

https://gizmodo.com/researchers-have-created-a-tool-that-can-perfectly-depi-1844051752

http://news.bbc.co.uk/2/hi/europe/4504589.stm

https://vowe.net/archives/005838.html

https://www.kaspersky.com/blog/how-to-leak-image-info/34875/

http://www.cs.cornell.edu/~shmat/shmat_imgobfuscation.pdf

https://help.adobe.com/archive/en_US/acrobat/8/professional/acrobat_8_help.pdf

https://talkingpdf.org/redacting-with-acrobat-8-professional-vs-redax/