How to tell when a robot has written you a letter

The next Turing Test? Handwriting.


A few weeks ago I got duped by a robot. In the mail.

I was sifting through my dead-tree postal mail and tossing junk in the recycling bin. Nearly everything that arrives in my mailbox is junk, so I was tossing, tossing, tossing … until suddenly, whoops: A hand-addressed letter. This looked legit, so I ripped it open — only to find it was an oily invitation to take out a second mortgage on my home. I’d been fooled.

Normally, I can tell the fake, printed-font “handwriting” of a marketing missive from ten paces away. That stuff looks barely more human than Comic Sans, and it’s printed in laser toner instead of ink from a pen.

http://www.youtube.com/watch?v=uyXlChb7-HI

But it turns out that marketers are working diligently to develop forms of mass-generated mail that appear to have been patiently and lovingly hand-written by actual humans. They’re using handwriting robots that wield real pens on paper. These machines cost up to five figures, but produce letters that seem far more “human”. (You can see one of the robots in action in the video adjacent.) This type of robot is likely what penned the address on the junk-mail envelope that fooled me. I saw ink on paper, subconsciously intuited that it had come from a human (because hey, no laser-printing!) and opened it.

Handwriting, it seems, is the next Turing Test.

This shouldn’t be surprising. The quest to get people to open spam and malware online has, in the last decade, produced some of the most innovative and hilarious attempts to make robots seem human. (My favorite was the mid-00s wave of spambots that quoted classic literature in an attempt to thwart Bayesian filters designed to recognize penis-enlargement invocations.)

So now robots are trying to write like us. But they’re not perfect yet! It turns out there are some intriguing quirks of human psychology and letter-formation that the machines can’t yet mimic. Learn those tricks, and you can spot the robots.


Brian Curliss and Daniel Jurek

I first heard of these human-machine handwriting differences in a conversation last week with Brian Curliss and Daniel Jurek, the cofounders of the startup Maillift. If you need to send out 200 personalized letters to sales leads but haven’t got the time to handwrite them yourself — or if your handwriting is, like mine, grotesque — then Maillift will generate them for you, using teams of genuinely carbon-based people. (What sort of person enjoys handwriting letters for others? “Teachers,” Curliss replies. Apparently teachers have spectacular handwriting, take enormous pride in the craft, and want to make some extra coin in their evenings and weekends.)

Curliss and Jurek also own a handwriting robot, so they’ve studied thousands of human-written letters and compared them to ones produced by machines. They’ve identified three crucial distinctions. They are:

The Clear Center

This was written by a human — specifically Curliss, using a Pilot G2 pen on my notepad.

Look closely at the strokes. When humans write by hand, we use irregular force, often pushing down too hard on the pen. This produces a “clear center”: If you consider each stroke as a road of ink, the middle of the road — the center of the strip — is occasionally somewhat transparent and devoid of ink. The pen has smooshed the ink out to the left and right sides of the pathway. In this example, the clear center is particularly visible in the “t” of “Maillift.”

Robots don’t do this. “With a machine-written letter it’s very consistent and very even,” Jurek says. “The pen is down or the pen is up, but it’s never really in between.” Here’s an example of robot handwriting:

See? No clear center.

Robot-making companies are currently busily working to match this human irregularity, but for now, it’s a giveaway that you’ve got robot mail.

(My apologies for the crappy quality of these pictures, by the way. When I met Curliss and Jurek, all I had with me was my Nexus 4 phone, which has a pretty wretched camera.)

The next difference?

The Dots On the “i”s

If you suspect a robot has sent you a love note, check the placement of dots on the letters “i”. This letter above was penned by a robot, and as Jurek notes, “the ‘i’s are always dotted in the same way.” They’re too regular. The dot appears in precisely the same place over the stalk of the “i” each time.

Humans are never this accurate. Everything from emotions to fatigue and shifting hand-position affect where the dots land. “People dot their ‘i’s and put punctuation in at the end of their sentences depending on how they feel as they’re writing,” Jurek adds.

Once again, we can expect the robots to get better at mimicking our all-too-human irregularity. The best handwriting machines already vary some letters to avoid seeming too repetitive. Check out “offer” in the third line of that letter above: The “f”s are different from each other. The bot has been programmed to pen that letter a few different ways, the better to fool readers.

But the final difference between human and bot penmanship is the most subtle yet, and to my mind the most intriguing. It is:

The Rounded Right Margin

When humans write by hand, our margins are — you guessed it — irregular. On the left-hand side of a paragraph, we try to keep a nice straight margin, beginning each line in the same location. But we usually can’t keep it very precise. We drift inwards, producing a left margin that slowly slopes towards the center of the page. Here’s an example of a human-written letter Maillift composed to send to customers:

See how the left-hand side drifts inwards? The lines beginning “choose”, “color” and “logo” all move to the right.

But the activity on the right-hand margin is even weirder.

Curliss and Jurek have noticed that their human handwriters often produce a “rounded” right margin: It bulges outwards and then tucks back inwards. In the example above, “showcase” sticks out further than the line above, “option”. But then the lines begin rounding inwards, with “size” and “own” moving further and further into the center of the page.

Why would humans do this? Probably because we aren’t terribly good at judging how many words fit on each line. If we accidentally overpack a line — trying to cram too many words into it — we immediately overcompensate by becoming too cautious, and putting too few words on the next couple of lines. You can see that dynamic at work in the example above. The composer wrote a nicely-kerned first line, but in the next line crammed in too many words, such that “showcase” comes uncomfortably close to the edge of the page. So he or she pulled back sharply, and the next three lines shrink in length.

A robot, in contrast, knows precisely how much space each word takes up, and doesn’t make these mistakes. You won’t see that rounded margin on the right.

Until, of course, robots are programmed to mimic that, too.


Half the time, I’m cynical/alarmed/wearied that so many people are working so hard to make machines fool humans.

But the other half the time I’m kind of cracked up by the fact that the most avid prosecutors of Alan Turing’s sly and audacious 1950 thought-experiment have been not philosophers or computer scientists or advanced A.I. labs but … marketers. The former folks have foundered for years on the difficulties of understanding the fractal contours of human consciousness. The latter just want you to open up their damn mail. Comprehending the mysteries of human thought and behavior is hard. Emulating it? Not so much! It’s partly why Turing’s test is so unsettling: Man, are we really that easy to copy?

Mind you, this particular Blade Runner dimension of modern life could quickly diminish in relevance, because frankly, postal mail is itself declining rapidly. The amount of upright, breathing humans who regularly write letters by hand has been shrinking steadily for years. So maybe it’s not long before handwriting flips its its existential polarity. A handwritten envelope will become not a litmus test of humanity but sure-fire proof that we were sent a form letter by an impersonal database. We’ll sort through our paper mail with the inverse logic of today, tossing aside immediately all the letters addressed with pen-script (robot, robot, god, another one sent by a robot) but then pausing at the sudden, startling appearance of an envelope addressed by a dot-matrix printer.

Hmmm, we’ll say to ourselves: Now this might be real.