That is the net model of Eye on A.I., Fortune’s weekly e-newsletter overlaying synthetic intelligence and enterprise. To get it delivered weekly to your in-box, sign up here.

Think about the lifetime of an promoting government and a scene from Mad Males is prone to come to thoughts: Don Draper snake-charming a pair of Kodak advertising executives with a perfectly crafted pitch in regards to the emotional pull of nostalgia (“It’s delicate, however potent…”) so as to win the account for his or her new slide projector. “This machine isn’t a spaceship,” Draper tells the entranced Kodak males of their slide carousel in a single well-known pitch from the tv present. “It’s a time machine.”

Effectively, it seems, these days have largely gone the best way of three-martini lunches, skinny ties, smoking within the workplace, and widely-tolerated office sexual harassment. Within the digital period, as an alternative of a high-stakes, high-wire act centered on excessive ideas, promoting has largely been lowered to a quantity recreation. Advertising departments or inventive businesses must churn out dozens or lots of of variations of digital adverts for Facebook, Instagram, or internet banners, every with barely completely different imagery, show copy, and calls to motion, after which conduct a collection of A/B experiments to determine what works for a selected target market. It’s a slog.

Just a few weeks in the past, I wrote about one company making an attempt to make use of machine studying to take a little bit of the drudgery out of this work, serving to to automate the testing of various adverts. Immediately, I need to speak about one other: Pencil, a startup that’s truly utilizing A.I. to create the adverts themselves. Based mostly in Singapore, however with staff working remotely throughout the globe, Pencil mechanically generate dozens of six, ten or 15-second Fb video ads in minutes.

“The advert trade has been transferring from massive concepts to small concepts,” Will Hanschell, Pencil’s co-founder and chief government officer, tells me. “As a substitute of a Superbowl advert, a multi-million greenback blow out yearly, it’s more and more about very small, on-line adverts. And in that setting, you need to run 10 adverts and throw out the 9 that don’t work and begin once more with one other 10. That has made the job unfun for lots of inventive folks.”

Pencil hopes it will probably liberate these inventive of us to work on the massive image whereas A.I. does the remaining. “It cuts movies into scenes, generates copy, applies animations after which makes use of a predictive system that appears at selection and tries to find out what feels most on-brand and appears much like issues which have labored up to now for the model,” Hanschell says.

An organization offers Pencil’s software program the URL of its web site, and that software program mechanically grabs the logos, fonts, colours and different “model picture data” discovered there to make use of in a enterprise’s adverts. It may use pictures from the web site or a enterprise can select to offer the system further pictures or video. It makes use of refined pc imaginative and prescient to grasp what is going on in a picture or a video in order that it will probably match that to advert copy. To write down the copy itself, Pencil makes use of GPT-3, the ultra-large pure language processing A.I. constructed by OpenAI, the San Francisco A.I. analysis agency.

Hanschell says that when Pencil began out, utilizing GPT-3’s predecessor, GPT-2, the advert copy it generated was usable solely 60% of the time. Now, with GPT-3 and higher understanding of tips on how to use the present internet copy to immediate the system, Hanschell says the system generates usable copy 95% of the time. What’s extra, the system can truly generate novel concepts, he says. For example, for a corporation that sells protein powder, the system can give you concepts round vitality, however it will probably additionally give you concepts in regards to the morning ritual or health, he says.

I watched a demo of Pencil’s software program by which it created a collection of Fb adverts for an eyeglasses firm. It got here up with the tagline, “Your frames, your means,” in addition to, “Your wildest appears to be like, completely crafted,” every paired with acceptable nonetheless pictures. Not precisely Don Draper. However not unhealthy. And as Hanschell factors out, within the quantity recreation of right this moment’s digital promoting jungle, a lot ok to start out buying prospects.

What’s extra, the system can present a prediction for the way good a selected advert will do in comparison with what the corporate has run up to now. For example, it forecast that the “Your wildest appears to be like, completely crafted” advert would do 55% higher than earlier adverts the identical firm had run. That’s one thing most human advert executives can’t do.

Pencil is already being utilized by about 100 corporations, together with some massive multinationals comparable to Unilever. It’s a good instance of a brand new era of merchandise—and even entire companies—which are being made attainable by speedy advances in pure language processing, or NLP. (For extra on this, try the most recent episode of Fortune’s Brainstorm podcast. Additionally, final 12 months, my Fortune colleague David Z. Morris wrote about several other companies utilizing A.I. to mechanically craft or refine digital adverts. )

However on the similar, a rising variety of moral issues are being raised about these underlying NLP techniques. For example, GPT-3, regardless of all of its seeming energy, nonetheless fails simple tests of common-sense reasoning. It additionally has a problem with bias: As a result of it was educated on the whole lot of the Web, there’s a very good probability it could have picked up an inclination to write down sexist or racist prose.

One space the place OpenAI itself has already acknowledged an issue: The system can exhibit a transparent anti-Islamic bias, with an inclination to depict Muslims as violent. A latest paper by two researchers at Stanford discovered that in additional than 60% of circumstances, GPT-3 related Muslims with violence—and that the system was extra prone to write about Black folks in a unfavourable context.

This lead the tech journalist David Gershorn, who covers A.I. for tech website OneZero, to question why OpenAI would allow it to be used in a commercial setting and why OpenAI’s investor and accomplice, Microsoft, can be incorporating GPT-3’s capabilities into its personal merchandise. How damaged does an A.I. system must be, Gershorn requested, earlier than a tech firm decides to not launch it?

I requested Hanschell about the issue of potential bias. He famous that OpenAI had developed filters that screened out a number of the worst examples. And he stated that in Pencil’s case, no adverts are ever run with no human approving them first. “One of many principals of that is that we needed a human to be in management always,” he says.

So I assume possibly we are able to’t get again to these three-martini lunches fairly but. There’s nonetheless work for us to do.

With that, right here’s the remainder of this week’s A.I. information.

Jeremy Kahn
[email protected]