Will AI Replace Inventors?: Computing vs Seeing

Does generative artificial intelligence (AI), including Large Language Models (LLMs), signal the end of the road for inventors such as me? Or will generative AI accelerate and enhance the quality of my work? I hold 204 issued and 54 pending U.S. patents, putting me in the upper 0.1 percent of inventorsⁱ with at least one U.S. patent and at the top of inventors in the agricultural equipment industry, in which I spent most of my career. Besides concerns for my future, how do I mentor novice inventors so they can catch the wave of AI-assisted invention? Or will that wave permanently submerge them?

What is an Invention?

An invention starts with seeing a need and then addressing that need. For this article, I add the U.S. patent law requirements for an invention: (1) patentable subject matter, including a device, system, composition of matter or method; (2) novelty or not being an exact copy of something already in existence; and (3) non-obviousness in the eyes of a hypothetical person who has “ordinary skill in the art” and access to all public information in the world. In Europe, non-obviousness requires an “inventive step” beyond common problem-solving.ⁱⁱ

It should also be noted that with the exception of South Africa,ii global patent law requires inventors to be “natural persons.”ⁱⁱⁱ AI software as an inventor is a topic of much current debate including court cases^iv and introduced legislation.^v On the one hand, companies using AI software to develop patentable inventions want legal protection for those inventions. On the other hand, there is interest in protecting the legal rights of humans participating in invention through AI software.^vi

Inventive Process

As an example of my human inventive process, consider thoughts during my drive home from a medical appointment.

What if we had digital twins, high-definition digital models of our bodies, to mitigate unpleasant and risky medical procedures?

For example, if a patient had a digital twin of their colon generated from the first physical colonoscopy, a medical procedure with a quite unpleasant preparation process. These procedures are first given when people turn 50 years of age and then are repeated every five to 10 years. Perhaps the digital twin could be used to “finish” the sequence of colonoscopies started by the human? In moving this concept to commercialization, there would be a many patentable details left to work out. At this point, there is early research on creating digital twins of human colons. There are questions that need to be answered such as what data should be collected in the initial colonoscopy and then what data needs to be provided to update the twin. While later colonoscopies might be eliminated, it is more likely that the common fixed five or 10 year retest intervals would be modified to longer, variable intervals, still resulting in cost and comfort benefits.

In a later coffee-shop conversation, an idea emerged regarding continuous blood-glucose monitoring, which would eliminate the need for a blood draw or wearing a patch, as an analogous application for a digital twin that could “complete” a series of invasive medical data collections. Currently, glucose monitoring is evolving towards wearable devices, such as smart watches, that will measure blood glucose using infrared or terahertz electromagnetic signals. The digital twin could then be used to predict the individualized effect of eating certain foods on blood glucose for managing diabetes or weight.

AI’s Computational Limits

Solutions, whether from computers or humans, need to be delivered in a timely fashion. Tomorrow’s weather forecast is of no value if it is delivered the day after tomorrow. Similarly, inventions by AI need to be delivered while still relevant to individual, business or societal needs.

In the early- to mid-20th century, work by mathematicians and early computer scientists explored the bounds of what could be computed or solved. The research identified a number of generic problem categories and, for each category, whether a solution could be calculated and the program end or halt. If the program does halt, how much computing time is required as the size of the problem increases? The fastest computing time is limited by the problem size, the ability of software algorithms to handle growth of problem size beyond a handful of inputs, and the physical limits of the computer These limits will now be considered in more detail.

Halting Problem

One such category is the Halting Problem, which seeks a general algorithm that, given a program and inputs to the program, determines if the program will eventually halt—that is, complete its execution or go into an infinite loop. Alan Turing,^vii a mid-20th century mathematician known for his work in breaking German codes during WWII, proved that there is no solution for the general case even though specific cases may be decidable.

For example, the following one-line program always halts after sending the phrase—“This program halts.”— to a printer or display device:

PRINT (“This program halts.”)

And this two-line program (below) never halts because the WHILE statement is an infinite loop that never exits. The PRINT statement is never reached in the execution of the program. Infinite loops are often experienced on consumer devices by a progress indicator that continuously spins or scans on a display, never advancing to the next step of program execution:

WHILE (TRUE) ENDWHILE

PRINT (“This line never gets printed.”)

Programs that can’t be decided typically involve calculations of numbers in sets for which the existence of a particular member or the largest member are not known. As an example, it is not known if there are any odd “perfect numbers,” or if the number of perfect numbers is finite.^viii A perfect number is an integer that is equal to the sum of its positive proper divisors, for example, 6 has proper divisors of 1, 2, and 3 and 6=1+2+3; 28=1+2+4+7+14; etc.) Consequently, it is not known if a program to find an odd perfect number, such as the one above left, will halt. ‘N’ is set to the odd number 3. ‘N’ is tested and if found to be a perfect number, it is printed and the program halts. Otherwise, the next odd number is generated by adding 2 to N, and then program execution returns to the testing step and, as necessary, trying the next odd number.

While the Halting Problem is typically considered in the context of finding solutions to mathematical problems, by analogy it can be applied to general inventive problems. Would an AI problem-solver be able to find a way to travel faster than the speed of light? Using all current physics knowledge, the answer is “no.” Could it discover new enabling physics? It is unknown if such physics exists and then if AI could discover it and halt. Thus AI problem-solvers could be turned loose on some problems without knowing at the start if they will ever return a solution; they could run indefinitely.

Growth Function

In addition to identifying if an algorithm will come to a solution and halt, another consideration is the growth function of algorithm-execution time as the number of inputs increases. This is typically expressed as a mathematical function using “big-O” notation.^ix One place computational time growth is observed is software for, say, an online store-product database. When a vendor demonstration is conducted with several dozen products in the database, the speed seems instantaneous. When the system is purchased and installed with data for tens of thousands of products, it runs painfully slow. Potential customers get tired of waiting and leave the online store before search results are delivered. People can also get tired of waiting for inventive problem solutions or find them of no value if they are delivered too late. The table below shows some examples of common growth functions for a baseline set of data, the 1x-column, which can be processed on average in one second. Later columns show typical execution times for datasets of some multiple (2, 4 and 8) of the original dataset.

Because of the intractable execution times needed to exhaustively process some large datasets, heuristic (rule-of-thumb) algorithms^x have been developed that can execute much more quickly but do not guarantee the very best result to a given problem. Shoppers often apply heuristic algorithms as they approach the checkout line to minimize wait time. Rather than do a detailed calculation to determine which line gets them through check-out as quickly as possible, shoppers often select the shortest line based on number of people currently waiting or the shortest line based on the total number of items to be scanned, or simply join the closest line open for checkout.

The sad thing about artificial intelligence is that it lacks artifice and therefore intelligence.

— Jean Baudrillard

Combinatorial Optimization Problems

Heuristics are often applied to combinatorial optimization problems, such as the O(n!) example in the table on page 47. Combinatorial problemstypically involve finding an optimal or best-fit solution from a finite set of possible solutions. The Traveling Salesperson Problem (TSP) is a common example. The basic problem is to travel to each of X number of cities, covering the least amount of distance and never visiting a city twice. The brute force approach of evaluating all possible routes to find the shortest one grows factorially, (X-1)!. For example, there are (8-1)! or 5040 possible sequences to visit eight cities. Numerous heuristic algorithms have been developed to provide good routes in a tolerable amount of time. Some have been animated, available on YouTube, to illustrate the approaches.^xi

Some inventions are carried out as search problems. Consider Thomas Edison’s invention of the incandescent light bulb. It was known that passing electricity through a material could cause it to heat up and glow. The problem was getting the material hot enough to provide a useful amount of light without quickly burning out. Edison’s approach was largely trial and error. He and his staff evaluated somewhere between 1,600 and 6,000 filament materials (reports on the number vary) until, in 1879, they settled on a carbon filament made from charred bamboo. It wasn’t until 1904 that tungsten was identified as a superior filament material, and then later the tungsten filament was wound to reduce the space it occupied, reducing the size of light bulbs producing a given amount of light. Putting the filament in a vacuum and later a neutral gas to extend its life was another dimension of light-bulb evolution.

The process was largely searching through a finite number of filament materials, dimensions, geometries and enclosures. The combinatorial space was large but could be reduced by pruning the search space based on empirical rules (heuristics), such as finer conductors will glow with less current and organic material tends to burn when heated in an oxygen environment but not in other environments. Yet there was the initial inventive step/observation that the problem of generating light with electricity might be solved (better) by running a current through a material, causing it to glow brightly. While carbon arc lamps were first demonstrated in the early 1800s, as was electrical incandescence, Edison iteratively mitigated deficiencies of the initial concepts, resulting in a commercial product.

Computer-aided identify-and-test is in use today, particularly with drug and material discovery.^xii The molecular shape of, say, a biological receptor is identified. Candidate molecules to fit the receptor are generated by software and tested to see if they fit and have the desired patient benefit. Chemicals passing that test are then evaluated for potential side effects on humans. Combinatorial growth in computation occurs based on molecular complexity, the number of candidates and number of evaluation dimensions.

Physical Limits

Besides AI’s algorithmic limits, related to the halting question and to processing growth functions, there are physical limits on computational devices that execute AI algorithms. For the last 60 years, computing devices have been dominated by semiconductor electronics that have steadily improved in performance. Silicon circuit features are approaching size limits where instead of following circuit paths, electrons tunnel between paths and the technology no longer works. Also notable is the speed of light, limiting the transmission of data within and between circuit elements to approximately one foot per nanosecond in a vacuum. In media other than a vacuum, the signal velocity is reduced.

Quantum computers have demonstrated superior processing capability for problem domains, such as encryption and optimization. These are still in the experimental stage with some challenges to overcome, including error rates; computer-size scalability; costs, such as the need for cryogenic cooling; and the development of programming environments and algorithm libraries. Some aspects of invention might be radically impacted by quantum computers, but it is difficult at this time to predict the full extent.

Philosophical Limits of AI

Beyond the mathematical and physical constraints on AI invention, there is a set of limitations presented here as philosophical since they are related to assumptions, values and beliefs.

The first two limitations are brain/mind vs computer/ program and the role of lived human experience in invention. As described above, some inventions can be handled by exhaustive generate-and-test, but that approach is limited by existing knowledge and by the combinatorial explosion of processing required for trial-and-error approaches. If the solution search-space can be trimmed, then computation requirements may be reduced. Also as noted earlier, invention involves identifying and applying new knowledge or perspective, sometimes called “thinking outside the box” or “taking the inventive step.”

One assumption in the belief that computers may at some point equal or exceed human capabilities in taking inventive steps is that computers can mimic the human brain/mind: that algorithms executed on silicon, quantum or other artificial structures can replicate the essential thinking of a human inventor.

“Replicate” can mean that inorganic circuits and software could be designed to mimic or simulate the biochemical processes of the human brain with its labyrinth of ~100 billion neurons and ~100 trillion synaptic connections. “Replicate” can also mean that any problem solvable by the human brain/mind could also be solved by a computer, possibly by a different approach. An example of this is chess, in which human grandmasters rely on patterns and looking ahead several moves, while computers rely on a different approach, looking at exact board configurations and game outcomes many moves into the future.

Lived Human Experience

Much invention is directed to improving human living, and how lived human experience plays into identifying those opportunities. Can this lived experiential knowledge be replicated by AI? Consider an example in which a robot checks into a hotel and walks into its room, which has a temperature of 87° F, well within its own and human operating limits. End of story.

Now consider the same story, but a human walks into the room and is uncomfortable until the air conditioner cools the room to 72° F (Where is the thermostat? How is it operated? Now I will be sweaty for my dinner meeting … .)

The human, who is an inventor, recognizes that having the thermostat set high reduces air-conditioner use and consequently reduces hotel energy costs and greenhouse-gas footprint. Yet these benefits are offset by customer discomfort when the room is initially occupied. How can the customer be comfortable when in the room while saving energy when absent? One solution is to control the thermostat based on a check-in time captured from the customer when the reservation is made or by using GPS and estimating time of arrival from a phone app.

This problem can easily be identified by a human, especially one who finds 87° F excessively warm and who can initiate an inventive solution.

How would a computer observe the problem and initiate a search for a solution? Big data and data analytics are possibilities. Guest comments that say, “Room was too warm,” are not as helpful as comments saying, “Room was too warm on arrival.” Then there is the need to collect thermostat data to quantify the phrase “too warm.” This may be difficult if the customer comments are anonymized for privacy, or if the thermostat can’t communicate with the customer-feedback software.

To solve the problem, a contradiction may be formulated along the lines of, “The room must be hot to save air-conditioning costs, AND the room must be cool for customer comfort.” Contradictions are often resolved through optimization, say, finding a room temperature of 76° F that balances energy costs and unhappy customer costs. In other approaches, the contradiction is resolved by recognizing that hot and cool can be time-shifted based on occupancy, leading to the solution of having the room warm when unoccupied and cool when occupied with a cool down occurring before the guest arrives.

Unspoken Needs

The hotel room example addresses spoken needs, but many needs go unspoken. Sometimes this is because data isn’t collected. Other times, people aren’t aware of the need, often because a suboptimal situation is accepted as the way things are. It’s not clear that computers can generally identify unspoken human needs and follow up with solutions requiring an inventive step. As Henry Ford is attributed to saying, “If I had asked people what they wanted, they would have said faster horses.” Horse riding or pulling a carriage set the standard for human transportation for millennia and so was seen as a fixed element. A horseless carriage didn’t make sense because it wouldn’t move. In the 19th century, steam engines propelled trains over fixed tracks, but off-track steam-traction engines were unwieldy and struggled on hills. They also required two people to operate: one steering and the other stoking the boiler. The opportunity to have personal transportation with something other than horses wasn’t seen until inventors subtracted horses and replaced them with steam and then petroleum engines.

Cultural, Social & Legal Limits

Another set of philosophical limits are cultural, social and legal. These limits are highly dynamic as AI technology evolution races faster than cultural and legal systems. To be viable, trained AI systems, such as LLMs, require large amounts of low-cost data.

However, businesses, including news organizations and book publishers, are pushing back on the uncompensated use of their copyrighted text and images. Lawsuits have been filed against generative AI companies regarding use of copyrighted material as training data without notification, compensation and consent.^xiii Individuals have privacy concerns about the use of their social-media posts, as well as concerns over defamation and other damages that can occur when generative-AI systems provide information or guidance, which is false and/or defaming^xiv or biased.^xv Restrictions on free training data will impact the capabilities and costs of AI systems used for invention. The U.S. Congress—as well as European lawmakers^xvi—is starting to consider legislation that promotes technical and commercial advancement of generative AI, while protecting the economic and privacy interests of data sources and those impacted by false outputs.

Society also faces the potential displacement of at least some humans involved in invention, as well as other jobs, by AI systems. A significant rise in unemployment from AI may result in deployment restrictions, service taxes, consumer boycotts or other actions, limiting the use of AI for invention. These actions are hard to predict, especially since the type and magnitude of pushback might depend on a catastrophic gray or black swan event (that is, a significant partially or fully unforeseen event) involving AI.

Other philosophical limits result from realities beyond current science and mathematics. New knowledge about the natural world through scientific or mathematical research could play significant roles. A better understanding of how the human brain/mind invents is a major subcategory. Advances in computer hardware and algorithms are others. Some philosophical limits might lie beyond the physical electrochemical activity of the brain/mind—for example, engaging a human soul. Unique creative abilities of humans are seen emerging from within and rising beyond cellular biology and electrochemical activity. Others posit unique creativity coming from a source external to humans and even to the natural world. Such inventive sources, if valid, are unlikely to be reproduced by computers. Assuming such elements exist, it is unclear what problems would have solutions reachable by humans and unreachable by computers.

AI-Assisted invention

Coming back to the original question—‘Will AI replace inventors?’—my current answer is that, in general, it will not happen in the immediate future. There are some specific areas, such as material and drug discovery, where generate-and-test algorithms in constrained problem/solution spaces will outperform human inventors sooner rather than later. For the rest, I see AI-assisted invention as the norm for the foreseeable future, say, at least five to 10 years.

For patentable inventions, the inventive process includes:

1. Problem identification

2. Solution identification

3. Technical development of the solution^xvii

4. Invention documentation write-up

5. Patent application write-up and filing

6. Patent examination

As noted earlier, the inventive step of identifying a novel and non-obvious solution (step 2, above) to a problem (step 1) may be difficult if not impossible for AI in the general case. Brute force approaches, such as trial-and-error, face the combinatorial explosion of computation and the physical limitations of computers. It is not clear how a computer could be programmed to take inventive steps given structural and lived experience differences between the human brain/mind/ soul and semiconductor or quantum computers.

However, AI can accelerate all six steps in the inventive process and potentially reduce the cost of invention. The combination of higher speed and lower cost computation provides an opportunity for early movers to leap ahead and block competitors from emerging product and technology spaces. This creates a unique opportunity for new entrants and existing companies seeking to increase market share and pricing control.

Currently, invention speed is often limited by the rates at which humans can wade through computer-search results and either find specific pieces of information or summarize a collection of information. LLMs are now doing the reading and stand ready to do search and summary. Humans ask a question, AI provides an answer, the human (for now) fact-checks the answer and then digs deeper with the next question.

There are three classes of people: Those who see; those who see when they are shown; and those who don’t see.

— Leonardo da Vinci

AI, particularly LLMs, can play the roles of a customer to enumerate common articulated problems, a technical expert to help develop solutions, a writer to describe inventions, and a patent analyst to find prior art for solution evaluation and patent examination. All these roles might be carried out quickly, inexpensively and with 24/7 availability. Days and weeks can be eliminated from the invention and patent application timeline to get customer information from marketing, solution help from a technical expert, and then the write-up of invention disclosures and patent applications.

Globally, patents for identical inventions are awarded to the first party to file an application at a national patent office. If and when granted, the patent prevents others from making, using or selling the covered invention for 20 years from the filing date in the countries in which the patent is granted. It is a high-stakes race with a patent prize only for first place.

Core of Invention

Even with the acceleration provided by AI in certain elements of invention, there is still a critical need for humans who can take inventive steps beyond the strides of current AI technology. The core of invention is seeing: seeing once to identify a need and seeing again to identify a novel and non-obvious solution to the need. It is seeing that requires the human brain/mind/soul nourished by lived experience and strengthened by cognitive training. And now it is seeing, enhanced by magnification provided by AI. ◉

REFERENCES

 i https://www.bitlaw.com/patent/requirements.html
 ii https://ipwatchdog.com/2021/07/29/dabus-gets-first-patent-south-africa- formalities-examination/id=136116/
 iii In February 2024, the US Patent and Trademark Office issued guidance on the patentability of AI-assisted inventions in the U.S.: https://www.uspto.gov/ subscription-center/2024/uspto-issues-inventorship-guidance-and-examples- ai-assisted-inventions
 iv In 2023, the US Supreme Court declined to hear the appeal of the district court’s decision on AI as an inventor. The UK supreme court also decided against Thaler in that jurisdiction. https://cafc.uscourts.gov/opinions-orders/21-2347. OPINION.8-5-2022_1988142.pdf and https://www.supremecourt.uk/cases/ docs/uksc-2021-0201-judgment.pdf, respectively.
 v https://ipwatchdog.com/2024/02/29/brazilian-lawmaker-introduces-bill- allow-ai-inventor/id=173809/
 vi The World Intellectual Property Organization (WIPO) and the US Patent and Trademark Office (USPTO) have been working on AI inventorship for the past five years. https://www.wipo.int/wipo_magazine/en/2019/06/ article_0002.html
 vii https://www.britannica.com/biography/Alan-Turing
 viii https://www.britannica.com/science/perfect-number
 ix https://britannica.com/science/time-complexity
 x  https://optimization.cbe.cornell.edu/index.php?title=Heuristic_algorithms
  xi TSP heuristic solution examples include https://www.youtube.com/ watch?v=SC5CX8drAtU and with more background https://www.youtube.com/watch?v=GiDsjIBOVoA
 xii Two examples including metals: https://www-technologyreview-com.cdn.ampproject.org/c/s/www.technologyreview.com/2022/10/25/1062104/machine- learning-new-metals/amp/ and another related to drugs: https://med.stanford.edu/news/all-news/2024/03/ai-drug-development.html accessed July 6, 2024
 xiii Examples of art lawsuits: https://itsartlaw.org/2024/02/26/artificial- intelligence-and-artists-intellectual-property-unpacking-copyright- infringement-allegations-in-andersen-v-stability-ai-ltd/ and text lawsuits filed by newspapers: https://www.npr.org/2024/04/30/1248141220/lawsuit-openai-microsoft-copyright-infringement-newspaper-tribune-post
 xiv https://www.cjr.org/analysis/ai-sued-suit-defamation-libel-chatgpt-google-volokh.php
 xv https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-april/navigating-ai-employment-bias-maze/#:~:text=The%20 Equal%20Employment%20Opportunity%20Commission,to%20 understand%20relevant%20EEOC%20guidance
 xvi https://bipartisanpolicy.org/blog/legal-challenges-against-generative-ai- key-takeaways/
 xvii Note that patents do not require construction or operation of a prototype. Thus, a sufficiently developed concept qualifies for patent protection.