Separating the Hype from the Hope for AI in Print

Sign up for The Key Point of View, our weekly newsletter of blogs and podcasts!

It’s difficult to avoid the topic of artificial intelligence (AI) in the news this year, and the hype surrounding the subject is spectacular. At any moment, you’re likely to hear industry experts predicting that AI will drive our cars, steal our jobs, fight wars for us, or fight wars against us. We even have senior executives pushing for a moratorium on AI research (as if that will help). While AI is undoubtedly an important topic, let’s take a moment to strip away some of the hype and consider possible ways it could impact the printing industry.

Can AIs Think?

First, AI is software. Yes, while it is highly complex code, it is still just software. Arthur C. Clarke’s third law states, “Any sufficiently advanced technology is indistinguishable from magic,” and AI systems certainly seem magical. We use terms like learning, thinking, problem-solving, and understanding while describing AI operations, which leads to anthropomorphism. But, no, AIs are not alive, nor are they sentient. However, thinking and sentience are challenging to define and measure.

Back in 1950, mathematician Alan Turing contemplated whether digital computers would ever be able to think. Turing punted on the question because he couldn’t find a meaningful way to define or measure thinking. Instead, he proposed a test he termed “the imitation game.” In it, he asked, “Are there imaginable digital computers that would do well in the imitation game?” Turing’s imitation game involves someone having a text conversation with both a person and a computer. The computer passes the test if the tester cannot determine which conversationalist is human and which is the computer. Programs like ChatGPT can now either pass Turing’s test or soon will, but this means these programs can mimic human behavior, not that they can think. We’re going to need a new test.

That doesn’t mean that AIs aren’t valuable tools. Nor does it mean that they can’t get out of control and wreak havoc. First, our world is largely controlled by software. Second, AIs build their models and generate new code based on what they’ve read, much like a high schooler whose research involves scouring the Internet. Third, people are using AIs to generate and post new content that sounds legitimate and which will, in turn, be used by AIs to “learn.”

Can AIs Evolve?

One of the most challenging aspects of AI is that it “grows” and “adapts” beyond its original programming. I apologize for using those terms, but the result is that AI systems “evolve” beyond their original programming in ways that their developers don’t understand because they entail so much data and computations. Sam Bowman, Associate Professor of Linguistics, Data Science, and Computer Science at New York University, wrote a fascinating paper titled “Eight Things to Know about Large Language Models.” In it, he made the following observations about large language models (LLMs) used by AI software:

• LLMs become more capable with more computing power and data, even without innovation. This unexpected result is unlike any other programming style where programmers must code new features and capabilities. AIs can do this simply with more data and computing power without new coding.
• LLM behaviors emerge unpredictably as they become more capable. Unlike all other programming styles based on “IF this THEN do that,” LLMs can quickly and unpredictably alter their responses and capabilities with no additional coding.
• There are no reliable techniques for steering LLM behavior. LLMs are not controllable. Programmers try to put “guide rails” in place to steer AIs, but human term-programmers and AIs always manage to get around them.
• The inner workings of LLMs and unknown. As previously stated, these self-adapting systems become so large and complex that no one knows how they work after a while.
• LLMs can express different values than their creators or training text. Just as LLM behaviors are unpredictable, so are their interpretations of their training content. Again, no one can know how these complex systems work after they have evolved.
• LLMs can exceed human performance. This is one of the most significant benefits and scariest fears about AIs. They can perform analyses that humans or traditionally programmed computers could never do. This is especially true for applications with large data sets and variables. AIs are responsible for many recent breakthroughs in medicine and science.
• Brief interactions with LLMs are often misleading.
• LLMs often appear to learn and use representations of the outside world.

AI Versus Rules-Based Programming in Print

It is crucial to consider the differences, benefits, and limitations of rules-based and AI programming to determine the appropriate approach to use in the precision-oriented printing business.

Rules-based programming is precise and gives the same result for the same input conditions (unless you code in a random number generator). This is ideal for financial calculations and many forms of prepress automation. Xerox FreeFlow Core, for example, provides efficient prepress automation for specific job types and is all driven by rules-based programming. However, rules-based programming becomes increasingly complex and challenging as the number of variables and options increases. Coders need to account for every possible condition. Thus, while different workflows can be automated, determining which one to use for each incoming job is still a human-operator decision.

With all its faults, AI programming dramatically simplifies handling large data sets and variables.

• Translating a free-text email in any language into a job ticket or even determining whether the email is a job request would be nearly impossible for rules-based programming. There are just too many possibilities. But off-the-shelf AI systems can quite easily do a pretty good job.
• Programming job routing and scheduling can become quite complex for handling an ever-changing array of jobs, deadlines, media inventories, press availabilities, and other factors. AI-driven approaches are well-suited for doing this sort of work. The AI software evolves and adapts, so the initial programmers don’t need to consider every eventuality.
• Business and workflow analysis of months of customer, shop, and press data can involve millions or billions of records. Many traditional systems can provide dashboards to help people visualize their shops, but identifying potential issues and likely causes as well as making recommendations is beyond what traditional coding can provide. AI software, however, would be ideal for this type of work.
• Design, layout, and development of marketing materials and campaigns typically involve fixed libraries of pre-built templates, workflows, and special coding. AI software could develop these using branding guides and best practices examples.

Keypoint Intelligence Opinion

Will AI destroy humanity? While that scenario is possible, it isn’t worth losing sleep worrying about it more than any other world-destroying possibility. Will we have fully autonomous, AI-driven print shops in our lifetimes? Probably not, but we will be using AI to be more efficient in the future. It is our opinion that AI software will increasingly help printers and in-plants address labor issues like aging workforces and unavailable skilled incoming workers by assisting in print-oriented tasks. We will see AI used in customer service, job onboarding, prepress automation, estimating, design, scheduling, and business analysis. However, AI solutions need constant verification and management because they can become unpredictable. Almost every workflow automation software vendor we meet is investigating how to utilize AI in their offerings. The following 24 months will be world-changing in print automation.

Log in to the InfoCenter to view additional research on AI and print automation solutions in our Production Workflow Advisory Service. If you’re not a subscriber, contact us for more info by clicking here.

Keep Reading

INFOGRAPHIC: The Basics of Large Language Models

Managed Service Providers Are Integrating Generative AI

ChatGPT Produces Yet Another Blog