Poe in Cyberspace: AI, Now and Then
Heyward Ehrlich · The Edgar Allan Poe Review · 2023
The communication revolution of the 1830s—combining the steam press, the telegraph, the daguerreotype, and the computing calculator—expanded both the media audience and its subject matter, promoting Poe’s career by opening new subjects for magazine fiction and also new kinds of literary criticism.One of the few writers of his generation to feel at home with the new technologies, Poe credited his interest in them with the creation of some of his own favorite works. “Ligea,” he began—quickly inserting an immediate surprise—was “undoubtedly the best story I have written—besides Scheherazade” (PL 616). Of course, few contemporary editors, critics, and readers understood the many discoveries and inventions Poe catalogued in “The Thousand and Second Tale of Scheherazade” (Godey’s, February 1845). Poe exploited the gap by making it an occasion for literary satire. His way of describing several dozen remarkable subjects in contemporary science and technology was to fling extreme hyperbole, making everything seem improbable if not imaginary. After each ultra-fanciful description Poe provided a footnote with a much-needed simple explanation. Thus “a man of brass and wood and leather” was Maelzel’s Chess-Player; the device that could equal “the united labor of fifty thousand fleshy men” was Babbage’s computing calculator; the appliance “heard from one end of the earth to another” was the telegraph; and what “directed the sun to paint his portrait” was the Daguerreotype (M 3:1167–68). In a reprint Poe soon added an extraordinary device that permitted the writer “to sit down in Damascus and indite a letter at Bagdad”: the Electro Telegraph Printing Apparatus. This complex device connected a local alphanumeric keyboard by a telegraph line to a remote electric printer with a daisy wheel that could deposit ink on a moving band of paper, something like a stock ticker. This ambitious device was an unsuccessful experiment, yet it was also a mechanical predecessor of today’s most complex artificial intelligence systems.When the generative artificial intelligence program ChatGPT suddenly appeared in November 2022, we were unaware that it depended on years of experimental trial and error by such devices as the Google search engine, the Apple smartphone, and the Facebook social network. Long before that, as historians recognize, Alan Turing in 1950 had given serious attention to the question of the relation of human intelligence to machine intelligence. The Turing Test was designed to determine whether a text message wired from another room came from a machine or a human. In 1966 Joseph Weizenbaum famously wrote the Eliza program in only two hundred lines of code to debunk other computer/human analogies, but many respondents unexpectedly took the program seriously as a legitimate form of nonauthoritarian Rogerian therapy.The learning curve of most AI systems is of course not smooth but a jagged series of detours. Today’s AI systems can be successful because they are standing on the shoulders of failures of the past. Although still widely used, Amazon’s facial recognition recruiting tool Rekognition was inadvertently trained with gender and race bias and thus was plagued by errors in dealing with women and persons of color, as the ACLU formally complained in 2017. Google Allo falsely saw a turban image as a gun; Google Translate suffered from a built-in gender bias; iPhone X’s facial recognition program is easily fooled by cheap masks; Uber’s self-driving program detected pedestrians only in crosswalks; and a Tesla fatality took place when the driver’s seat was inappropriately unoccupied (www.techrepubic.com).Microsoft may now seem a late entry in the AI sweepstakes, but back in March 2016 it launched an AI experiment called Tay, on which more than 95,000 tweets were posted within sixteen hours, including overtly racist, misogynist, and antisemitic expressions, such as the statement, “Hitler was correct to hate the Jews.” Tay quickly suspended the service and soon pulled the plug. What Microsoft didn’t take into account was that a group of Twitter users immediately begin tweeting racist and misogynist comments that the bot quickly incorporated into its own tweets (www.analyticsinsight.net).The dangers of AI have been well publicized in recent months. We have been made aware of the possibility of massive technological unemployment, widespread economic dislocation, intolerable climate change, and even uncontrolled military operations that could result in human extinction. These are not idle fears. Speaking through the Center for AI Safety, the artificial intelligence leaders Sam Altman of OpenAI, Dennis Hassabis of Google DeepMind, and Dario Amodei of Anthropic in May 2023 issued a strong statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” In May 2023 these representatives discussed with U.S. president Joseph Biden and vice president Kamala Harris the necessity of forming a strong regulatory office such as the International Atomic Energy Agency (New York Times, May 31, 2023, A1). Later, President Biden announced that representatives of seven major companies—Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI—had agreed in principle to accept voluntary AI safeguards (New York Times, July 22, 2023, A1). To pursue this end, Senate leader Chuck Schumer (D, NY) proposed holding congressional education sessions in the fall of 2023.There is immediate urgency in the need for regulation. We are familiar with the danger of contaminated printed documents after the earliest versions of ChatGPT released in 2022 made all printed documents theoretically suspect. However, now more advanced AI agents are going further to invent voices and photographs, two hitherto untouchable bastions of truth and evidence. What are called deep fakes, such as cloned voices and false videos produced by AI, have already appeared in local elections in Chicago, Toronto, and New Zealand. To combat their effect, the American Association of Political Consultants issued a condemnation of these violations of its ethics code; moreover, Senate Democrats have introduced a bill requiring the use of prominent disclaimers in the circulation of all AI-produced material (New York Times, June 26, 2023, A1). The danger is that unscrupulous politicians will use such techniques despite any public denials.Some lapses in judgment in using AI have been well-publicized. Lawyer Steven A. Schwartz used an AI chat program to prepare his research for an injury case against Avianca airlines, citing prior cases against Delta, Korean, and China Southern Airlines, but none of these precedents actually existed. Judge P. Kevin Castel described these invented facts as “bogus,” the result becoming “a bouillabaisse of existing cases” (New York Times, 5/29/23, A1).Unfortunately, consumers of information often find fiction to be more convincing than fact. In 2023, forty-one of seventy-seven false articles being prepared for a test on CNET evaded detection by their human editors (New York Times, May 30, 2023, A19). In another test, a group of school-age children were easily fooled by disinformation (Scientific American, February 1, 2022). One of the downsides of ChatGPT is that it makes misinformation both fast and easy to produce and hard to detect; another is that fact-checking employees are often underpaid, some of those operating in Kenya, Uganda, or India and receiving, it is charged, as little as $2 an hour (Guardian, March 31, 2023).Nevertheless, despite the fear of fake news, conspiracy theories, and outright lies, there is also hope that AI would produce beneficial results. Today, generative or AI programs also promise to benefit many areas, such as health care, manufacturing, law, insurance, journalism, research, business, financial investment, flight safety, and weather forecasting.Of course, all fears of future AI should be balanced by the realization that we have already adopted many forms of AI that have become beneficial and so commonplace that we are no longer aware of them. Without thinking of how they work, we use word processing software to check our typing and spelling, to provide a ready thesaurus, and even to watch our grammar and expression. In today’s telephone answering systems, machine intelligence listens for traces of keywords in our speech. In online chat, we sometimes try to find words to bring a human online to rescue us. Formerly we had to travel many miles and even continents to do our research, but now we take for granted that despite our spelling errors Google will understand our meaning and offer us the most likely web pages, including encyclopedic and bibliographical resources. When we lose patience with our typed queries on our phones, we speak to iPhone Siri or Google Now—often with faster and more relevant results. (Now if only it knew phone numbers and email addresses!) Machine intelligence has already mastered the games of Go, chess, and Jeopardy, leading perhaps to false public confidence in hands-free self-driving systems.Needless to say, not every computer glitch can be attributed to AI. In October 2020, PHE (Public Health England) suffered a disruption when the maximum of 1,048,576 rows and 16,384 columns was inadvertently exceeded in an Excel worksheet. In 2021, a Zillow algorithm led it to unintentionally purchase homes at higher prices than its current estimates of future selling prices, resulting in a $304 million inventory write-down. In 2021 a Wisconsin pharmacist destroyed vaccines because he had heard on a social network that they modified DNA, giving unrefrigerated vaccines to fifty-seven people, for which he was sentenced to three years in prison and a $83,000 fine in restitution (NBC News, June 8, 2021; Washington Post, February 1, 2021). Indeed, not every problem with AI applications is the result of faulty software; experts in the field have long warned that poor data selection can result in a situation described as GOGO, meaning “garbage in, garbage out.”It may be comforting to realize that AI is now being used successfully in many fields. The New York Times complied an authoritative two-page feature, “How Real People Use A.I. in Work, School, and More,” with a few dozen short case histories of actual good uses of AI to plan gardens, plan workouts, plan meals, make a gift, design parts for spaceships, organize a messy computer desktop, write a wedding speech, write an email, play devil’s advocate, organize research for a thesis, cope with ADHD, imagine themselves as a copywriter just hired to proofread documents, sort through an archive of pictures, transcribe a doctor’s visit into clinical notes, appeal an insurance denial, write Excel formulas, get feedback on fiction, get homework help, learn Chinese, get help when English is the second language, create an app when you’ve never coded before, fix bugs in your code, play Pong, build entirely new games, teach people to curl like a pro, create new proteins in minutes, identify diseases in banana plants, draw like Sol LeWitt, describe entire Dungeons & Dragons worlds, make a Spotify playlist, and just play with language (New York Times, June 17, 2023, B4–5).Today Microsoft’s Prometheus combines interactive ChatGPT with the Bing search engine, potentially a very powerful combination. In another field, Amazon’s Proteus is developing smarter robots for merchandise handling, expected to expand its robot force from ten thousand to seventy-five thousand (Wired, June 26, 2023). Google has a great deal at stake with its leadership in search, Gmail, and a raft of other computer services. The company has reportedly “used machine learning in Gmail for the past few years to generate one sentence replies and predict what you’re likely to type next” (Wired, March 1, 2023). Currently Google expects to invest tens to hundreds of millions of dollars in its DeepMind unit, comparable to the cost of developing GPT-4, to develop its own advanced AI called Gemini in order to compete with OpenAI’s ChatGPT.It is still early in the AI era, and many issues remain unresolved, such as those pertaining to copyright, plagiarism, liability, documentation, investment, profitability, influence, abuse, and restraint. This column reported in Spring 2023 that the Quora website, now called poe.com, had an AI software program, also called Poe, which linked to other AI programs, such as ChatGPT, OpenAI, and Anthropic, leading in turn to Claude, Claude+, GPT-4, and Sage, and resulting in dozens of bots in fields such as HTML, JavaScript, programing, resumes, and writing. For the last item, this description is offered: “Introducing a new language processing bot that is designed to elevate your writing to the next level! Our bot is here to help you improve your grammar, syntax, and word choice, as well as add eloquence to your writing.”To sum up, Edgar Allan Poe the author should not be confused with poe.com the website or Poe the AI computer program. In a recent search on Open.AI for “Poe ChatGPT” two of the three were oddly merged with this fascinating but untrue result: “Recently, a language model called Poe ChatGPT has been developed that merges the power of AI technology with the storytelling prowess of Edgar Allan Poe.” Nevertheless, a Google search for the word poe will lead to Poe the AI computer program. One such result begins: “Poe—Fast Helpful AI Chat: Poe lets you ask questions, get instant answers, and have back-and-forth conversations with AI. Gives access to GPT-4, gpt-3.5-turbo, Claude from Anthropic. . . .” What is quite astonishing about this new hyper-cyberspace Poe is its very high rating in the Apple App Store: the “Poe—Fast AT Chat” website earns 4.8 out of 5.Poe in Cyberspace columns are archived at eapoe.info.