An AI Lie Detector Is Going to Start Questioning Travelers in the EU
gizmodo.com
gizmodo.com
And it's a similar promise to a lot of other over-promising AI technologies that don't work, like "fake news detectors".
And how do you know that?
I'm not for it, but it's not a polygraph as described.
Quote: "Looking for cues to deception merely from ephemeral facial micro-expressions is questionable and likely fruitless. Micro gestures may be indicative of internal emotional turmoil that is being suppressed, but that is it. The distinguished Paul Ekman, who in fact coined the term micro - expression has stated in his book Telling Lies that micro expressions are rare and they "don't occur that often" (Ekman 1985, 131, 165). Plus as others have said, there is no single behavior indicative of deception (Matsumoto et. al., 2011, 1-4). I am concerned that machines that focus solely on the face will no doubt miss other information from the body (sweating, jittery hand, etc.) or generate lots of false positives because negative emotions abound especially where such machines are intended such as airports (stress of travel, stress of being subjected to searches, or inconvenient interviews, etc) or in a police setting."
[1] https://www.psychologytoday.com/ca/blog/spycatcher/201203/th...
All in all, just another worrying piece of news from Hungary.
Please take a step back and look at all the replies to you. They all essentially say the same: We don't know of any reliably objective measurement for measuring lying. Humans can sometimes do that intuitively, yes, but to my knowledge there hasn't been a single successful attempt of a generalization of detection. That's why state of the art means using trained humans, like the airport security in Israel.
Without a mechanism to detect lying without using humans, no amount of statistics (which ML boils down to) can help you. ML can't change the fact that you don't know if the data point is a signal or just noise. Without knowing what the signals look like, ML can't calculate the noise away.
Even simpler: For ML you need a dataset with people lying and people saying the thruth. That's what you train the model on. But we have no way of building that dataset, because, to my knowledge, we don't have a way to reliably build up this dataset: We can't reliably detect if someone is lying.
[0] https://en.wikipedia.org/wiki/Burden_of_proof_(law)
[1] https://www.wired.com/story/inside-polygraph-job-screening-b...
And nobody is being accused of anything, there is no burden of proof involved yet. If you mean that we can't subject someone to a lie detector while asking questions, i.e. we cannot subject anyone to anything without a reasonable suspicion, then we couldn't have airport security at all unless someone tweeted "today's plane is gonna be da bomb!"
Which would be a lot better than the current state of things in the USA, but that's not how this works.
I think it's unlikely that they'll be perfect, but they may very well be good enough to filter people for further, more intensive screening.
You'd be surprised. AI is just a catch-all marketing term, see for example IBM's Watson umbrella brand, that's sold like advanced AI but has had tons of fiascos, and is shown to be different sets of often simplistic implementations.
E.g: https://www.massdevice.com/report-ibm-watson-delivered-unsaf...
https://spectrum.ieee.org/the-human-os/robotics/artificial-i...
https://www.theverge.com/2018/7/26/17619382/ibms-watson-canc...
In this case, there's excellent reason to be sceptical about whether the service has a remotely adequate dataset, because even if lie detection were a simple problem, how on earth are you going to get calibration data involving hundreds of data points of real airline passengers lying about having weapons in their luggage or intending to overstay their visa? Never mind calibration data by gender, ethnicity and nationality...
Usually when people say 'AI', IBM included, what they mean is fairly trivial machine learning. Machine learning requires a dataset to learn on.
> In this case, there's excellent reason to be sceptical about whether the service has a remotely adequate dataset, because even if lie detection were a simple problem, how on earth are you going to get calibration data involving hundreds of data points of real airline passengers lying about having weapons in their luggage or intending to overstay their visa?
You could do it by watching film of interactions with border agents when it is still unknown whether or not the person was lying, and then use whether or not they were caught smuggling (or whatever) ex-post as your ground truth.
You're going to need a lot of film. Good luck picking the controls for that that ensure your ML system picks up on subtle differences in humans that are lying and not glaringly obvious differences in appearance, accent and even background noise between the footage of the convicts and the control group...
I assume you would isolate the people in pre-processing, and do feature extraction of things you might think are interested in. Yes, it'd still require a lot of data, but that data may exist somewhere. A lot of these things are filmed.
From CIA, and TSA, to your local police department, everybody would have jumped all over it.
This is some vendor managing to sell some BS to bureaucrats (probably with the required bribes).
If CIA had it, you wouldn't hear about it. I don't see any reason why these countries couldn't pilot it for their border security before TSA, though.
FWIW, polygraphs are still used for security clearance screening. So, they are at least accurate enough for that use. They get a bad rep because they aren't admissible in court, but it's not like they have zero correlation to truth telling. They are still quite accurate.
a) it'll work,
b) it's actually any sort of AI,
and that, for example, even reputable companies like IBM can peddle all kind of shit as AI.
So, IBM was used as an example of a common practice, not as some logical axiom that "because they did it, nothing AI will ever work". Example != logical necessity.
It does, however cast doubt, especially for such a thorny problem as "lie detection", that nobody has, thus far, solved.
Even if all this "AI" does is match the quality of a polygraph, it is probably extremely useful in screening people rapidly at the border.
You obviously don't work for IBM :)
A skilled IBM Watson sales consultant as capable of selling simple linear regressions as Big Data AI!
While I don't disagree that "AI" as a term is being abused, the distinction you're making seems very much a philosophical one.
Where did they get that dataset from? Did they ask a bunch of people a bunch of questions and then hired private investigators to verify the answers (and then hire extra investigators to cross-verify answers given by the previous bunch)? Because beyond something like this, there's no known way to create such a dataset.
Iraq and other countries purchases millions of dollars worth of phony bomb detectors. Something anyone with a bit of science education can debunk.
Look at all the things on Kickstarter that are obvious pipedreams or scams yet people invest.
https://www.vanityfair.com/news/2015/06/fake-bomb-detectors-...
This is a hilarious read.
Former inmates, drug addicts, surgeons, lawyers, mechanics, welders, astrophysicist, particle physicists, pilots. Need I go on?
Not once have I read anyone here who’s admitted to being a high level politician.
[0] https://slate.com/technology/2013/04/dowsing-for-bombs-maker...
That anonymity disappears much more quickly if said Member of Congress has also posted in the past about where they're from and which political issues they're passionate about...
Our society was built on large amounts of discretion being exercised by law enforcement and the ability to convincingly lie to law enforcement.
Society is going to get flipped on its head if we are able to detect total truth, and enforce all laws with 100% effectiveness. Even worse when these tools will be mostly in the hands of the government, and not the people.
This I get.
> and the ability to convincingly lie to law enforcement.
This less so. Can you elaborate?
If everything were some day 100% honest, the polite niceties of "oh sorry didn't see the speed limit change" or "Forgot I left my pocket knife in carryon, sorry" would go away. Of course these are only two of countless examples.
"Do you have any intention of overthrowing the government?" - "No" - beep, and now you're in a forced labor camp
Even if these tools worked, they wouldn't be able to detect truth, only belief.
Personally, I don't like the concept of deploying algorithms for which we don't understand how they work, in scenarios of high health or social impact. It's cute to deploy DNNs to classify cat pictures, as when - not if - it spectacularly fails, nothing of value is lost. But if you can't point at the verified model the algorithm follows, or at least trace the algorithm after the fact, understanding and assessing every step it took, then that algorithm has no place in making decisions about people's lives.
I'm just wondering if it's too much to hope that they'll base the decision whether to implement it beyond the trial on data and not on marketing.
And there is of course the argument that it should be possible, albeit with a lot of effort and/or a reasonable punishment, to break the law. (If this was impossible, we'd never get out of regimes we don't agree with, and with the advance of technology I'm afraid we're heading somewhere where a small group is actually able to oppress a huge number of people.)
Border agents are trained to look for all sorts of things, none of which make you guilty, but a few of which may make you suitable for more questions or a quick search.
So I'd imagine it'll get used in the same context.
Their system is supposed to detect lying from facial expressions. The only kind of "science" purporting to back this possibility is the work of the psychologist, Paul Ekman, which is based on flimsy evidence at best. A gigantic hint that this "research" is a bunch of hooey is his claim to have identified 29 "wizzards of deception detection" [1]- in very, very literal terms those are people with the magickal power to tell when someone is lying just by looking at their face (and magickally perceiving revealing facial expressions, subconsciously).
The iBorderCtrl system might indeed be replaced by a wizzard, or, why not be more inclusive, a witch, with a magick wand pointed to the traveller, while questions are being asked of them. If the traveller is telling the truth, the wand will jump up, if they're lying it will dive down. The witch is not moving the wand! She's only channeling the MAGICK!
The same magick being channelled by this revolting misuse of technology for the most odious purpose imaginable. It is not a coincidence that this "prototype" is being deployed in Hungary, the country in the EU that has embraced populist, xenophobic tendencies as no other, having elected a master of the craft, Viktor Orbán, as a prime minister and head of government.
This is such utter, utter bullshit. I cannot believe that the Commission agreed to all this. What the fucking fuck.
___________
[1] http://www.communicationcache.com/uploads/1/0/8/8/10887248/o...
I can. I mean, the driving forces in the EU got extremely xenophobic. Italy is ruled by the fascist Salvini, the PiS in Poland are not far behind Orban, the UK government has been xenophobic for decades, France suffers from xenophobia too (after the terror attacks, though) and our own Horst Seehofer risked collapsing the German government over (literally) 0 migrants (per https://www.focus.de/politik/deutschland/seehofers-deal-zahl...)...
What are human rights, what is democracy worth, when there is no one left to enforce them? The EU won't do shit against Hungary, Poland or Italy (as there is a 100% consensus required!), the US under Trump are going isolationist and trampling on human rights wherever they can, and the UN Security Council is powerless against the stuff that Russia, China, Saudi-Arabia etc. do because there's always a veto power that bails out their "friends".
The unique approach to ‘deception detection’ analyses the micro-expressions of travellers to figure out if the interviewee is lying.
http://ec.europa.eu/research/infocentre/article_en.cfm?artid...
To OP: maybe today with a huge GPU, lots of data and more complex models you could get something going. Do you think a more advanced model could really pull this through? Maybe using other features -- perhaps coupling with a thermal camera can help detect changes in temperature that could signal the person is lying.
[0] quotes because I've heard this complaint a few times already, and it's usually from people that have bought the buzzword but usually don't understand much about machine learning in general.
In the research I was working on, there were two types of deception/lies: high stakes (aka suicide bomber or shoe bomber) and low states (you have a pound of cocaine strapped to your leg). If you want to build a lie detector on facial data, you need consider things like
1. It is important to have a robust data set that contains lots of example of both for training and cross-validation
2. False positives can be very expensive (because you start pulling people aside, lines build up, people get pissed, etc)
The project I worked on was trying to solve a specific problem: replace Behavior detection officers [1]: These are people that sit inconspicuously at the TSA checkpoints and decide who gets to get pulled aside for further screening and who goes through (I do know anything else about what they do and where they are, numbers etc.) Their training is advanced, laborious, and expensive for TSA. They are, of course, trying to catch terrorists (aka the high stakes lies) but in reality they are dealing with the low stakes lies. The only terrorist I can think of since 9/11 that got close to taking down a plane was Richard Reid, and they missed him so they seem to be 0/1 in high states lies).
Anyways, the problem with building their replacement is we do not have a lot of good examples of high stakes lies in video data. There are some (Scott Peterson, etc) but not enough. Also, the models they were built off of were "micro-expression" [2] detectors, and micro-expressions show up differently across cultures (maybe the consensus has changed since I was modeling this stuff) . So, in short, the project I was on failed because the data sucked and even if we did build a model, it was overfit to the data we trained it on. I am not sure how deep learning and more advanced GPUs would solve that. Maybe they found better data, but I would be skeptical. I honestly would be surprised if they actually did much better than random (ROC > .5)
Sorry for the rambling - end of the day for me on the east coast.
[1] https://www.tsa.gov/news/testimony/2013/11/14/tsa-behavior-d...
The article also mentions microexpressions as the basis for this new system, and that they worked on a slow universe of 30 people, with ~76% accuracy. What remains to be seen is if the new, larger deployment will work better. Also, they may be working more focused on the low stakes lies rather than the high stakes.
What I was thinking that "more compute power" could bring to the table would be deeper and more complex models, faster training algorithms[0] and data augmentation. But then again, all that relies on having good data to begin with.
[0] I worked with neural nets in the past (~10 years, I pretty much left the field right before the deep learning big bang), and coming back to the field I'm pretty much amazed at how training a 7 layer deep MLP used to take a few minutes on CPU back in 2010, and now can be done in seconds (still on CPU). And while there's a part of microarchitectural improvement to account for, that could at best be a 5x speedup -- the rest is owed to the newer training algorithms.
Once again, thanks for your answer!
It's crass and juvenile and totally inappropriate but after more than ten years it's also second nature and still represents the gist of what the synergizer's plans are likely to accomplish.
Magical 85 percent accuracy. so it is basically toy.
With 85% accuracy, 85 of liars are flagged as lying (correctly), and 15% x 9900 = 1485 of non-liars are flagged as lying (incorrectly).
Thus, a bit more than 5% of people flagged as lying are actually lying, while nearly 95% of people flagged are innocent. This is not even taking into account the possibility that hardened criminals might be less nervous than somewhat anxious normal people.
Enjoy your border crossings, everyone.
EDIT: fix italics
EDIT to add: And that's after they get the accuracy up to 85%. And unless accuracy is defined somewhat differently.
This is not what accuracy means.
85% accuracy just means that 85% of all the decisions the system makes are correct. A system in such a setting, where a single false negative matters a lot more than a single false positive (which would simply be handed over to a human for further investigation) would necessarily be tuned for extremely high recall at the cost of precision. In other words, it would often flag innocent people for further investigation (as you've said), but it would almost never clear people that should've been flagged.
https://en.wikipedia.org/wiki/Accuracy_and_precision#In_bina...
Unless I'm mistaken (and that's possible, I've changed my opinion twice now), my example outlined above is
- conceivable, and
- has 85% accuracy (85 people correctly identified as liars, 85% x 9900 = 8415 correctly identified as non-liars, thus a total of 85+8415=8500 of 10k total "accurately" identified), and
- still only 5% or 6% of flagged liars are actual liars.
EDIT to add:
And if the system is tweaked as you suggest, to very rarely fail to flag a liar:
- suppose it correctly flags all 100 liars as liars
- suppose accuracy is still 85%, thus 8500 people in total classified correctly
- thus 8400 non-liars flagged correctly, and the remaining 1500 non-liars flagged incorrectly
Now still only 6.25% (100 of 1600) of people flagged as liars are actually liars. Thus, even with the tuning you suggest, this remains.
(Note to self: 1. think 2. write)
You really have to compare precision and recall values to know if the accuracy statement holds true. You could have have 100% precision and low recall and still have 85% accuracy (meaning you could never flag someone as lying and be wrong while missing a bunch of liars and still have 85% accuracy).
but if everything is totally evenly distributed, then 85% accuracy means 85% accuracy and your first statement is correct.
The real issue is that accuracy is only one piece of the puzzle.
Let's make the spherical cow approximation that "a lie" is a fully defined concept, then we have 4 conditional (bayesian) probabilities:
P( "sincere" | sincere) The probability a sincere person is reported as "sincere".
P( "lying" | sincere) The probability a sincere person is reported as "lying".
P( "sincere" | lying) The probability a lying person is reported as "sincere".
P( "lying" | lying) The probability a lying person is reported as "lying".
The first 2 probabilities should sum to 1, and the latter 2 possibilities too, so we have 4-2 = 2 degrees of freedom. A reported "accuracy" tells us nothing without knowing the distribution of liars and sincere people in the test group..
This is going to be awful - imagine all the people who are anxious anyway, perhaps dont speak the langauge very well, get confused by the questions etc. There are going to be a lot of people who will get "enhanced screening" and generally treated like a criminal for no other reason than "the computer said you are a liar".
Awful.
So it maybe that the bigger problem is the false negatives.
Even if they bring it up to 85% accuracy with 50% base rate, by the time you are dealing with base rates that are more realistic, you're going to run into way more problems than just 1 out of every 7 people.
As always your contact will be via your MEP, who appoints the commission president (based on the fact the EPP was the largest party in 2014 and Junker was the presidential candidate of that party)
The countries in which the pilot takes place were mainly EPP (odd correlation...?)
By the way, direct links to newpress release and project page, which lists involved countries :
- http://ec.europa.eu/research/infocentre/article_en.cfm?artid...
A "driving while X", a terry stop or other fishing expedition type stop is adversarial.
Occasionally we have comment threads on HN that remind me how far removed a lot people here are from these issues and it does not make me confident that there will be improvement in my lifetime. This comment thread is one of them. I don't support this AI thing at all but as far as problems with law enforcement go border agent's demeanor should be really far down the priority list. I'm all for things that put the issue on people's radar and I don't wanna come off as saying "but X is worse so we shouldn't care about Y" but this seems like a first world problem.
Thinking about being stopped by the police for speeding as being something routine is insane. You should't be routinely doing something dangerous and illegal (speeding is illegal per se in most states.)
>You should't be routinely doing something dangerous and illegal (speeding is illegal per se in most states.)
You should read up on what the academics and engineers have to say about speed limits instead of pearl clutching. Unrealistically low speed limits are worse than no speed limits, often times there's politics and road metric gaming involved in setting a speed limit so you often wind up with speed limits that are unreasonably low. You can't just pick a limit and expect people to follow it if it's unreasonably low.
I don't see why not - I manage it fine, have never had a speeding ticket, and don't see why other people can't do the same.
Without being told the speed limit, how fast do you go?
Now that you're going that fast, you enter a section of highway that has been annexed by a nearby incorporated town, its speed limit in town lowered to 35 mph, and its only cop sitting out on the highway, clocking cars and looking for speeders to cite for "70/35".
Without the threat of cops, most drivers generally drive at the subjective and situationally variable speed described by Adams in HHGttG as "R1", which is the maximum reasonable and prudent speed for the ambient conditions. When lanes narrow, R1 drops. When fog blankets the road, R1 drops. When traffic gets congested, R1 drops. Everyone already has ample incentives to drive at a safe speed to protect their affordable insurance premiums, their expensive vehicles, and their priceless lives.
In general, the locales that recognize this set their posted speed limits to the 85th percentile of driver speeds in normal conditions. They only ticket people that are clearly beyond the community norms for R1. The locales that want ticket revenue set the speed limits to R0.9 or less, so that everyone is occasionally subject to citations if they aren't extremely vigilant about avoiding the enforcement.
The latter is why not. When there is profit in catching law-breakers, the law will be adjusted to criminalize otherwise normal behaviors.
If none is posted, then whatever the national or state speed limit is.
> The locales that want ticket revenue set the speed limits to R0.9 or less, so that everyone is occasionally subject to citations if they aren't extremely vigilant about avoiding the enforcement.
So the only people who aren't caught are those that are carefully observing the speed limit? Sounds good to me. These supposedly revenue-seeking police departments never bother me.
There are plenty of people willing to actually drive R1, which would be 60 mph, if there weren't any cars poking along at just under the speed limit, trying to avoid giving the cops their pretext for a highway robbery. Setting the limit too low makes the road less safe, because it increases the range of speeds at which the cars are moving.
It is small consolation to see literally half a dozen other cars pulled over on a 5 mile stretch of road-that-should-be-highway, because the town certainly could save money by firing at least three of their cops, and possibly the whole department. There's no excuse for that kind of official depredation. And yet, I don't live there, so I can't vote for them.
It is no sin to break the law when the law is turned against the people, but I am reluctant to pick fights that other people would have to join in order to win. I'd rather enact a boycott on that particular town's businesses--and any on the other side of them--until they clean up their own mess.
...why not?
Without naming countries, you are handled like a criminal very often.
Your definition of 'not adversarial' is intriguing.
> as far as problems with law enforcement go border agent's demeanor should be really far down the priority list
Oh really? When border agents can seriously screw up one's life (for instance, by placing a decade ban, or detaining them, and worse) with no oversight, I'd say it's a pretty big problem.
EDIT: > They're just routine transactions.
People have been killed in 'routine enforcement stops'. Agent demeanor is a big deal.
Which makes perfect sense since it's extremely hard for customs agents to actually catch smugglers. (Compared to immigration screening which is largely based on hard criteria which human agents evaluate with much less leeway.)
The bigger question is whether it's effective - but anything is likely to be more effective than customs agents selecting people to search based on gut feeling.
Well, I didn't know my taxes were used to build such systems. It would have been really more acceptable if they sold the system as a simple chat bot that records what you says, in case you are involved in a case later. But the "lie detector" part is scary and this is the typical instance that Elon Musk warned against.
I don't understand how academics accepted to work on that.
I think the idea was that, with any kind of profiling, it would be easy for drug gangs to figure out who to use as mules. And with any judgement, they can pressurise the guy making the call. But a lottery machine everyone can see is harder to defeat.
Furthermore, I believe the government has been given broad exceptions from GDPR for anything related to doing government stuff. This is form of legitimate interest "processing is necessary for the performance of a task carried out in the public interest or in the exercise of official authority vested in the controller;" It is in the public interest the border control laws be enforced and a manner that is both effective and minimized cost to the public, also this involves the exercise of official border control authority, therefore this is acceptable.
(Standard reminder that if you are relying on the advice of a stranger for legal issues that matter you should get an attorney.)
https://ico.org.uk/for-organisations/guide-to-the-general-da...
It is interesting that there is a comment about explicit consent - I imagine this will be like the millimeter wave scanners: you can either go in and get your genitals imaged by some remote voyeur, or if you dont like that then you can get a thorough strip search by someone in blue gloves. Have a nice day :-)
That's the point. If you refuse then they do a thorough search.
If you answer suspiciously, or refuse to answer, then you get the full search.
This is similar to how you can refuse body scans at the airport, but if you do then you get the full pat down search.
If you act very weirdly, like you are suggesting, then they will just make the decision to definitely search your bags.
Targeting people who act weirdly is kinda the entire point, though.
Also, the use case of border control is a great application. There's a lot of misunderstanding in this thread. The AI screening is just a first screening, and if someone fails that, then they go to a human. So moderate false positive rates are acceptable.
The commentators on this thread generally make a few mistakes: assuming lie detectors won't work because polygraphs don't work, assuming the border control use case needs to be perfect or not have false positives (it doesn't), assuming Paul Ekman's microexpressions are all that iBorderCtrl is basing their research on (I agree Ekman's research is questionable, and I don't know what exactly iBorderCtrl is doing, but it seems highly likely they're doing more than just looking for microexpressions), assuming racist intent or that it'll just flag non-Europeans as liars.
[citation needed]
I would be very surprised if the "research suggesting AI-based lie detection to be possible" that you mention is from anyone who has any sort of reputation to protect. Machine learning scientists, the vast majority of, would not touch such obvious pseudo-scientific claptrap with a ten-foot pole. It's the kind of thing that tarnishes one's reputation and never washes off. And rightly so.
I would not welcome, but grudingly accept, your references to the contrary.
Also, if iBorderCtrl are not using Eckman's work, then what kind of theoretical framework are they basing their work on? Why is it "highlly likely they're doing more than just looking for microexpressions"? Where is all the science of detecting lies from looking at peoples' faces?
And if they're not basing their work on someone's research, then what are they basing it on?
The IBORDERCTRL system has been set up so that travellers will use an online application to upload pictures of their passport, visa and proof of funds, then use a webcam to answer questions from a computer-animated border guard, personalised to the traveller’s gender, ethnicity and language. The unique approach to ‘deception detection’ analyses the micro-expressions of travellers to figure out if the interviewee is lying.
From the Commision's website, posted here by another user:
http://ec.europa.eu/research/infocentre/article_en.cfm?artid...
Domestically, many airlines already have automated checkins (including luggage dropoff in some places), automated boarding, and in some places automated passport checks at immigration. The last human hurdle is security screening, which is mostly theater at this point.
So, automated screening of travelers at the beginning of their journey makes a lot of sense. Mostly it's just confirming the obvious: are they who they claim they are (i.e. does the person showing up match the automated profile available already, are there any red flags warranting extra attention) and are they carrying anything they should not be carrying. Like AI already outperforming physicians in the job of scanning medical images for anomalies, surely state of the art luggage scanners are also outperforming humans (probably by magnitudes). Add to that some screening for clear markers that somebody's behavior is a bit off and you have basically automated away security staff likely to miss those signals because they are human beings that get tired, stressed, bored, distracted, biased, etc.
So the combined checks of identity, profiling, automated luggage scans and escalation to humans in case of any doubt should be vastly more efficient. The default case should be zero humans involved with the whole process. When it escalates, you still get the humans in the loop that are then a lot more effective because they already know there were some red flags.
This stuff will initially perform quite poorly probably. But it will still be worth it in identifying the "definitely not lying, don't waste your time on this" category of travelers.
That is absolutely not how any of this works, at all.
That all just seems to be fancy way of saying it's discriminatory and racist.
There's more than a whiff of Checkpoint Charlie about the place (portakabins installed directly on what was the motorway surface, concrete traffic calming measures, 5 km/h speed limit, armed police, more police sitting in chase car should anyone decide not to stop, floodlit at night)
I'm sure it's worth it, what with Austria and Germany sharing a fairly long land border which is more or less completely unsecured. Side roads - of which there are plenty - don't get checked much either. <rolls eyes>
https://www.reuters.com/article/us-amazon-com-jobs-automatio...