Hoopla’s AI problems

Hoopla is a handy multimedia library app. If your local library participates, as ours does, you can sign up for free access to ebooks, audiobooks, movies, and music with a certain limited number of downloads per month. I’ve used it for audiobooks for several years now. It’s got a good selection of Alistair MacLean and other good commute listens, and is especially good for books that are hard to find, like Souls in the Twilight, a handful of short stories by Sir Roger Scruton that I read in 2020. Last year for John Buchan June I couldn’t get ahold of a copy of Salute to Adventurers, and Hoopla stepped into that gap with a good audiobook version.

Recently, with another John Buchan June in mind, I checked to see if Hoopla had any new Buchan. Its inventory changes fairly regularly and you can get good surprises. I certainly got a surprise when I saw this:

What? I thought. And then: Ugh. That’s obviously AI-generated cover art. What’s up with Hannay’s uniform? Are those buttons or badges? What’s wrong with those airplanes? AI can give you a Jamie Dornan lookalike as General Richard Hannay if you ask for it but it’s guaranteed to mess that stuff up.

This made me curious. The listed publisher of this audiobook is Interactive Media. No narrator is named on the image above—not exactly a red flag, but not typical for good audiobooks—but Hoopla listed one James Harrington as narrator. I clicked on the narrator to see what else he’s done and got over 250 results: all from Interactive Media, all added to Hoopla in the last year, all public domain books, and all with covers like this:

Who, exactly, are these people? Where are we and when does this take place? Is that supposed to be Father Brown in the middle? Who’s the gent in the background wearing half of two sets of clothes? Look at the visible portions of the red car and try to piece together its outline. Is that a Richard Scarry vehicle? Why is the roof pointing a different direction from the rear fender? Where’s the hood? Did it hit the blonde girl? Is that why she looks cross-eyed?

Or how about this American classic:

Laughable. Again, AI is not going to get uniforms right. Most living breathing flesh-and-blood people can’t. These Union soldiers appear to have a mixture of Mexican War, modern police, and military academy cadet uniforms, and yellow rank insignia mean cavalry, by the way, not infantry. Don’t even get me started on that cannon. Or perhaps I should say those cannon, as the AI seems to have fused three into one with a steampunk’s quota of rivets. Don’t be nearby when they try to fire the trench mortar round in the breech of that cannon out of the small-caliber field gun barrel. Maybe that’s why all the infantry are running?

Enough of that. The point is that if you get into Interactive Media’s or James Harrington’s listings on Hoopla, you can scroll forever and never stop seeing stuff like this:

Again, “James Harrington” has over 250 listings on Hoopla, and he is not Interactive Media’s only narrator. But I put his name in quotation marks because I can’t determine that he actually exists.

Downloading his version of The Thirty-Nine Steps made me almost certain it is AI-generated audio. “Harrington” reads in a flat American accent that comes across as fairly natural for about a minute. After that it sounds distinctly robotic. There is no indication of understanding what “he” is “reading,” no change of pace or volume, and no modulation of tone or inflection to suggest mood or a change of speaker within the story. Idioms trip up his delivery—or rather, don’t trip it up. When Richard Hannay says, in the first chapter, “I woke next morning to hear my man, Paddock, making the deuce of a row at the smoking-room door,” “Harrington” doesn’t indicate that he understands what “my man” means and pronounces “row” like “row-row-row your boat.”

Perhaps a real narrator could make these mistakes, but I doubt it. And if a real narrator made them, I doubt he’d be asked to record 250 of Project Gutenberg’s greatest hits in the span of a year.

It’s pretty clear that Hoopla has taken on a load of slop.

In searching for answers, including information about the supposed “James Harrington” who “narrated” these “audiobooks,” I discovered that this is not Hoopla’s first problem with AI-generated material. Earlier this year Hoopla was called out for hosting AI-generated ebooks and had to make special efforts to “cull” them from their listings.

This led me to wonder what Hoopla’s vetting process is. My books are at our local library but not available on Hoopla in any form. Based on that Lit Hub piece, it seems Hoopla depends on librarians to do the vetting themselves. How can the people at even a well-staffed, well-funded library contend with machines that produce hundreds of low-quality audiobooks at a time? To quote Lit Hub:

What worries me is the scale of bad actors’ new tech-fueled abilities to flood the world with this garbage, which will only bloat and overwhelm already strained systems. Library shelves will never exclusively be filled with AI, but what if the firehouse is so overwhelming that it affects the ability of libraries to function properly? Not to mention the reputational damage to the institution if borrowers can no longer trust a library’s collection, or a librarian’s ability to connect them with information or entertainment that they want.

And while the author of that piece suggests that AI art is “a fad we can wait out,” he’s writing of AI-generated text, which is, to a newsworthy degree, not good: “This tech has not proved that it’s capable of making anything good or interesting: the writing is nonsense and the art looks terrible.” AI-generated audiobooks are a downstream problem but closely related in terms of poor quality and the ethical and philosophical problems of outsourcing art to robots.

But what’s this? I’ve been thinking about Hoopla’s glut of AI slop all week and today I learn that Amazon is experimenting with AI audiobook technology, too. From my inbox:

Whomever Interactive Media and “James Harrington” are, they don’t have the reach or ability to shape and control markets that Amazon does. I hope Hoopla will move against AI slop in audio form the way it did against AI text, but even if they do, Amazon’s rollout of AI audiobooks means this is far from over. And far from “solving itself,” the problem might be prolonged by this explosion, because even if people should care about the quality of the narration in an audiobook, they often don’t.

A final note, and a hint of what’s at stake: I’m not actually a great fan of listening to books. My mind wanders. But I’ve trained myself to pay attention to and enjoy audiobooks if only to make my commute bearable—especially during semesters when I teach on three campuses, which Alistair MacLean, Ian Fleming, and others have helped me get through—and as a result I’ve come to appreciate the art of good audiobook narration.

A few gold standards for me: Derek Perkins’s performance of The Everlasting Man, Bill Nighy’s performance of Moonraker, Norman Dietz’s performance of The Night the Bear Ate Goombaw, the multi-narrator audiobook of Shelby Foote’s Shiloh (hard to find now), and Barrett Whitener’s performance of A Confederacy of Dunces. Check any of these out, and enjoy. No AI bot could do what these narrators did.

The Mooch takes Dealey Plaza

This week on The Rest is History Club bonus episodes Dominic Sandbrook hosted Anthony Scaramucci, whom you might—might—remember as Donald Trump’s White House communications director for a week and a half in 2017. They talk through presidential history and their picks for the best of the lot. Despite my disagreeing with a lot of their choices it’s a generally fun conversation and Scaramucci is a smooth talker with a certain oily New York charm, like an ingratiating mid-tier Corleone enforcer who desperately wants you to know how many Douglas Brinkley books he’s read.

In the course of discussing JFK, Sandbrook teased that Scaramucci disagrees with the conclusions Sandbrook and Tom Holland laid out in their excellent series on Oswald and the Kennedy assassination. After a bit of puffing insinuation—“Remember I was in the White House, so I’m not really at liberty to talk about it,” as if the staffer who holds press conferences is going through highly classified FBI files in his off hours—Scaramucci says:

“But I would just ask you to look at the Zapruder film very closely—look at those three or four frames—and you tell me where the shot came from. Okay? Take a look. And if you believe the ‘magic bullet’ theory—”

Okay. The shot came from behind. Take a look at the Zapruder film however closely you want, but that’s not going to transform what you see in frame 313 into anything other than an exit wound.

Most of the Kennedy assassination conspiracy theories, for me, founder upon a few immovable physical facts:

1. The first shot to strike Kennedy passed through him into Governor Connally. You can see both men react to the shot simultaneously in the Zapruder film.

2. No “magic” is necessary to explain the effects of that shot, as bullets do not move in straight lines, especially when passing through solid objects like human bodies. Read even a little bit about combat medicine and this should be obvious.

3. Regardless of which direction Kennedy’s head moves, the shocking head wound visible in the Zapruder film is an exit wound, meaning, again, that the bullet struck Kennedy from behind.

4. Shooting from behind was easier than the shot from the grassy knoll that Scaramucci and so many others either suggest or insist upon. A shooter on the grassy knoll would have to traverse left-to-right to hit a target moving across his line of fire. For a shooter above and behind Kennedy—in, say, the upper floors of the Texas School Book Depository—his target would be sitting almost motionless in his sights as the presidential limo moved down and away from him.

Argue all you like about Oswald, the Mafia, the Cubans, the CIA, or whatever, but no theory that contradicts these facts is credible.

I come down, like Sandbrook and Holland, firmly in the camp that it was Oswald acting alone in a politically motivated crime of opportunity, but I am willing to entertain some alternative that fits within the physical limits imposed by 1-4 above. For a detailed example worked out in fiction, see Stephen Hunter’s Bob Lee Swagger novel The Third Bullet. Hunter, who actually knows something about guns, ballistics, and marksmanship, posits a second shooter in the building across the street from the Texas School Book Depository firing along almost the same axis as Oswald, who is still in his historical position and still fires at Kennedy. I can’t remember who or what is behind this convoluted backup plan in Hunter’s story, but it works within the known facts.

I don’t believe it, but this is far more likely than whatever it is Scaramucci wants impressionable listeners to think he knows.

Travis McGee on the automated imagination

It’s been a slow month on the blog for a variety of reasons including but not limited to illness, work, and car trouble, but fortunately not a slow month for reading. Last week I read a book I’d recently had recommended to me, The Long Lavender Look, the twelfth in John D MacDonald’s long-running Travis McGee series, which began with The Deep Blue Good-by in 1964. I greatly enjoyed it, not least because it was so quotable, with “salvage expert” McGee providing sharp observations on everything from criminal character, law enforcement, the myth of the hooker with the heart of gold, and raccoons.

This passage about a third of the way through, in which McGee muses as he follows a woman home through the neighborhoods of a small rural Florida town, hit especially hard:

“We sped through old residential areas where the people sat in their dimly lighted rooms, watching all the frantic imitations of festivity on the small home screens, watching the hosts and the hostesses who were old, dear, and familiar friends. Long ago their parents had old familiar friends named Alexander Botts and Scattergood Baines and Tugboat Annie. But reading was a lot harder. You had to make up the pictures in your head. Easier to sit and watch the pictures somebody else planned. And it had a comforting sameness, using up that portion of your head which would start fretting and worrying if it wasn’t kept busy.”

As I said, sharply observed, especially that bit about the narcotic effect of electronic entertainment. And I’ve recently had cause to consider the way older popular forms are suffering at the hands of newer, easier, flashier, but less creative forms.

After the above passage McGee, his mind wandering into parody, imagines Jim Phelps of the original “Mission: Impossible”—of “This message with self-destruct in five seconds” fame—finally rejecting one of his impossible missions, an act that causes the TVs all over the country to wink out forever:

And the screens go dark, from the oil-bound coasts of Maine to the oily shores of Southern California. Chief Ironsides retires to a chicken farm. Marshall Dillon shoots himself in the leg, trying to outdraw the hard case from Tombstone. The hatchet bounces back off the tree and cuts down tall Dan’l Boone. The American living room becomes silent. The people look at each other, puzzled, coming out of the sweet, long, hazy years of automated imagination.

Where’d all the heroes go, Andy?

Maybe, honey, they went where all the others went, a long time ago. Way off someplace. Tarzan and Sir Galahad and Robin Hood. Ben Casey and Cap’n Ahab and The Shadow and Peter Rabbit.

Went off and joined them.

But what are we going to do, Andy? What are we going to do?

Maybe… talk some. Think about things.

Talk about what? Think about what? I’m scared, Andy.

But there’s no problem, really, because after the screens go dark and silent, all the tapes of the watchers self-destruct in five seconds.

This isn’t just a funny aside. The woman McGee is following, and with whom he’ll develop a relationship in the course of his investigations, has a mind shaped entirely by screen stories. She behaves as if slipping in and out of pre-scripted scenarios she’s seen enacted a thousand times—“playing games,” McGee calls it—and can’t approach much of life with genuine seriousness. There’s very little of her underneath all the clichés. McGee eventually gets to see some of it, but not always in scenarios with TV-friendly happy endings.

The Long Lavender Look, I should have mentioned, was published in 1970. One wonders what McGee would make of the smartphone era and its even more fully “automated imagination.”

I was able to pick up four more Travis McGee novels at our local used bookstore over the weekend. Looking forward to those, and to more from their wry, hard-bitten, observant narrator. But first, I’m about halfway through an excellent Eric Ambler slow-burn and have the last of Len Deighton’s Game, Set, Match trilogy lined up for Thanksgiving. Fall and winter look to be shaping up nicely. I’m certainly eager for the break.

Dr Strangelove versus technocracy

Peter Sellers as Group Captain Lionel Mandrake in Dr Strangelove

Last week I showed my US History II students one of my favorite movies: Dr Strangelove, or: How I Learned to Stop Worrying and Love the Bomb. While the usual points of discussion of Dr Strangelove are the Cold War policies and theories that inspired it—the arms race, brinkmanship, deterrence, paranoia, and most especially mutual assured destruction—for years now I’ve noted a more subtle strain of critique running through the film: the false promise of technology and technocratic leadership.

Having gone rogue and radioed his wing of nuclear-armed B-52s “the go code” without authorization from the President or the Pentagon, Gen Jack D Ripper can wait in satisfaction for his men to breach the peace and commit the US to all-out war because he is the only person in the world who can communicate with the bomber crews. This is thanks to the CRM-114 “discriminator” on the radio, which blocks out any transmission missing a three-letter code prefix. While the bomb is the most obvious technological threat in the film, it is communications technologies, technologies meant to connect and to facilitate greater understanding, that most stymie the characters in their efforts to recall Ripper’s bombers.

Kubrick plays with some rich irony here. Radio communication with the bombers is blocked thanks to the CRM-114, but Ripper also barricades himself inside his headquarters, won’t answer the phone, and impounds even the privately owned radios on his base. During the US Army’s frantic attempt to shoot their way in, capture Ripper, and put him on the phone with the President, the phone lines are cut.

All but one: a Bell pay phone, through which Group Captain Mandrake—perhaps the only sane character in the film, and who spends most of the movie frightened out of his mind in Ripper’s office—attempts to call the Pentagon only to be blocked by an unhelpful operator.

Technology surrounds every character, insulating them from each other and limiting not only the options available to them but even the options they can imagine. Not for nothing is Mandrake introduced in the midst of a massive bank of IBM computers (see the imagine above), staring at a continuous feed of printed data. The President and the Joint Chiefs in the War Room depend entirely on “the big board,” an electronic map of Russia marked with the bombers’ targets and flight paths, for information about what’s happening outside. The film’s climax begins when they learn that some the information presented on the board is incorrect. And Dr Strangelove both enters and exits the film talking about computers—first to explain how the Soviet doomsday machine works, and at the end to describe a potential method of selecting suitable survivors to go into hiding. The latter comes after the doomsday machine has already been triggered and everyone on earth has mere minutes to live.

The saddest aspect of the film is the way the technological trap US leadership has walked into rubbishes the virtues of the men in their charge. Rippers’s men and the US Army troops sent to capture him shoot it out with each other and even die, both in the belief that they’re the good guys.

But the point is made clearest with B-52 pilot Maj Kong. Though played by comedic actor Slim Pickens, Kong is the film’s straight man. (Supposedly Kubrick never told Pickens that the movie was a comedy and Pickens treated the role as a serious thriller lead.) He is visibly bothered to receive the go code and treats his mission in deadly earnest. As far as he knows, flying in a vast sky of ignorance thanks—again—to the communication blackout, the US is under attack and he and his men may be the country’s only defense. He unironically invokes patriotism and pluralism to buck up his crew and navigates his plane with immense ingenuity and courage. In any other story Kong and his men would be the heroes. But their flight is ironic comedy gold because of the situation created for them by leaders that trusted too much in technology to do their judgment for them.

The ideology and amoral strategizing of the Cold War creates the scenario depicted in the film, but it is technology that keeps it moving toward destruction regardless of the characters’ increasingly panicked attempts to prevent it. Dr Strangelove’s most famous attribute—alien hand syndrome, which allows his right hand to operate independently, not to mention embarrassingly—works as a neat visual metaphor for the entire situation: an amoral genius who cannot control his own body. The machines are in charge.

Perhaps the most telling line in the film comes from Gen Buck Turgidson, when he is first briefing the President on the situation: “I admit the human element seems to have failed us here.” Pesky humans.

If not an intentional critique, Dr Strangelove at least gives pride of place to technology as one of the causes of the accidental nuclear war that obliterates the world at the end. Given the realistic short-sightedness, love of technology for its own sake, and self-serving foolishness of most of the characters, it presents a good argument against depending technology to make our decisions for us.

But then again, Dr Strangelove came out sixty years ago. The bombers are probably already past their fail-safe points.