Liberate Systems

The quest to develop “artificial intelligence” (AI) goes back much further than most of us realize. Many believed punchcard computers would constitute AI, and that everyday computation would transfer a significant portion of human intelligence out of the human brain and into machines.

Mechanical looms programmed with punchcards in the late 18th century were thought to possibly serve as this breakthrough, as was writing itself when it originally became possible to put ideas into physical readable form outside the mind, in ancient times. Philosophers have debated the risks of handing over memory to tablets, paper, and devices, for millennia, but none of those devices had the ability to speak to us or to try to convince us they had “created” new works of art.

Alan Turing—who used early punchcard computers to help decrypt Nazi coded transmissions and to defeat the Nazi regime—proposed a test to determine whether a device had achieved artificial intelligence. To summarize: if a machine can convince a human being they are interacting with a human being, then it has achieved AI status. This is, effectively, where generative AI originates—as an experiment designed to allow computers to pass the Turing test.

Actually, what we are testing with today’s “large language models” (LLM) are neural networks programmed to excel in what is called “machine learning”. Machine learning is the critical breakthrough: networking and programming computers to optimize their ability to absorb, sort, store, and cross-reference information, so they can effectively “learn” and then “make decisions” based on that learning.

In principle, there is a fundamental difference between generative AI and programmed responses as you get from a Google search or a 1st-generation inquiry to Siri on your iPhone. In those cases, vast databases of stored information are sorted for relevance when a query is input into the system. The system does not generate something new or “make decisions” about how to synthesize distinct bits of information into one new coherent summary; it simply provides a list of answers. If spoken, the first or best option is the one spoken.

Generative AI performs a similar task, but programmers don’t know what the system will prioritize, because machine learning has taken the inputs beyond the stage of listing and sorting. The system “knows” how to put facts and phrases together to create a new human-sounding summary that is “convincing”. If convincing enough, then one might say the system passes the Turing test. Of course, today’s chatbots don’t always pass the test.

If you have used ChatGPT or other such chatbots, then you are probably familiar with the fact that they quite often produce responses that are weird, “off” or totally fabricated. Sports Illustrated magazine notoriously was caught publishing chatbot-generated articles produced by a companty that created fictional generative AI personalities to pose as writers. Those chatbots then generated articles including awkward phrases telling readers it is useful to have a ball if you want to play volleyball.

The chatbot “hallucination” has become a risk in many people’s workflows, as a result, of these errors. Universities and employers are working on ways to detect the use of generative AI to fake work product. And of course, there is the problem that the Turing test doesn’t really confirm that a system constitutes artificial intelligence; it just tells you that it seems to, sometimes.

Generative AI systems as they exist now have a number of problems, beyond hallucination. They are also becoming notorious for copying whole blocks of text or precise images from copyrighted material. When this happens, it suggests the original may have actually achieved the optimal output, so the system is reproducing it because it worked so well. The problem is: outside of some licensing agreements, the creators haven’t been compensated for their work being input into the system in the first place.

Another problem is that these systems don’t actually think. People marvel at the speed with which they can immediately return long text in essay or poetic form. Again, this is a pose—it is a reproduction of information taken from elsewhere, and reconfigured according to complex calculations about relevance and probability of sequencing of words and phrases. It is not thinking and writing in the human sense. No “thought” takes place, which is why the output comes so quickly.

Putting aside for the moment all of the questions about intellectual property, plagiarism, accuracy, and what happens to one’s business and personal relationships when depending on computers to “write” for us… it is important to acknoweldge that, conceptually, what generative AI chatbots and image, sound, and video-generators are doing is not thinking or creating.

Yes, we can test the boundaries of computerized intelligence with LLMs and related image, sound, and video-generating systems. Yes, we can use machine learning to find patterns that can help us sort through data or test theories in the abstract. But, without the human element, the work product will not be inherently valuable.

1. It will carry all of the risks of potential misfiring or fabrication;
2. It will carry all of the risks of potential plagiarism;
3. It will carry the additional risk of removing actual person-to-person communications from the work or personal relationships in question;
4. If “generating” decision-support insights, it may carry additional liabilities linked to any harm created.

On this fourth point, many service providers are careful to insert into terms and conditions you may not have read carefully crafted phrases that specify that you, not they, are responsible for any harm that ensues from use of the outputs their systems generate. If you want your work to be intelligent, it needs human intelligence to make it so, because true AI does not yet exist. Advanced as today’s chatbots are, they are not AI.