As we, the meat-dressed humans, find ourselves in the brave new world of artificial intelligence, it’s easy to point fingers at the occasional missteps of our digital counterparts.
We gasp in horror at the lawyer who submitted a court document containing fictitious cases invented by ChatGPT, and we shake our heads in disbelief when an FT colleague prompts the chatbot to produce a chart of the training costs of generative AI models, only to discover that the most expensive one it identified did not exist (unless the model has access to inside information). But let’s take a moment to reflect on our own fallibility, shall we?
The Fallibility of Human Memory: A Case Study
One need only look to the infamous testimony of John Dean, legal counsel to the White House during the Nixon administration, to see the glaring flaws in human memory.
Despite being dubbed “the human tape recorder” for his remarkable recall, Dean’s account of critical conversations with Nixon was rife with lapses and reinterpretations when compared to the actual tape recordings from the Oval Office.
In a 1981 paper analyzing Dean’s testimony, psychologist Ulric Neisser highlighted several glaring lapses and reinterpretations of conversations in the lawyer’s account, as well as the difficulty of defining truth and accuracy.
Neisser drew a distinction between semantic and episodic memory, noting that while Dean was roughly right in remembering the overall gist of his conversations with Nixon and the nature of the Watergate cover-up, he was precisely wrong about the details of particular episodes.
But this is not a condemnation of our species, but rather a celebration of our shared imperfections with AI. Just as we humans struggle with the distinction between semantic and episodic memory, so too do our digital counterparts.
One could argue that large language models do the opposite of Dean: given all the data they ingest, they should have good episodic memory (although with garbage inputs, they can generate garbage outputs), but they still have poor semantic memory.
An AI model may be able to summarize the Oval Office recordings with greater accuracy than Dean’s months-old recollections, but it would have no contextual understanding of the significance of that content.
The Promise of AI: Reducing Hallucinations
But here’s the beautiful thing about AI, my meat-suited companions: it’s constantly evolving and improving. Researchers are working tirelessly to enhance AI’s episodic memory and reduce those pesky hallucinations.
Take, for example, the recent paper from Google DeepMind researchers, which proposed a new methodology called Safe – search-augmented factuality evaluator.
This experimental system breaks down model-generated responses into constituent sentences and cross-checks them against Google Search for factuality, outperforming fact-checking human annotators in terms of accuracy and cost.
As one of the paper’s authors, Quoc Le, so eloquently puts it, “In the next few years we will be able to fact check the output of large language models to some good accuracy.
I think that’s pretty useful.” And let’s be real, who among us wouldn’t appreciate a little fact-checking in our own lives? Hallucinations, as Le reminds us, are both a feature of LLMs to be welcomed when it comes to creativity and a bug to be suppressed when it comes to factuality.
The Importance of Relevance
But let’s not get too carried away with the promise of AI, my meaty friends. For as Maria Schnell, chief language officer at RWS, reminds us, it’s not just about how content is created, but also how it lands.
In a world where content is cheap and ubiquitous, the real value lies in tailoring information to a specific audience in a format, language, and cultural context they understand. And that, my friends, is where we humans shine.
While AI may struggle with relevance, we have the unique ability to consider how content is received and to add that all-important human touch.
So let us not fear the occasional hallucination or misstep from our digital counterparts, but rather embrace the opportunity to work together, magnifying our different capabilities and minimizing our respective flaws.
The Perils of AI Hallucinations
But let’s not forget, my meat-dressed companions, that AI hallucinations can lead us astray in the most amusing ways. Take, for example, the case of Microsoft Bing’s Copilot, which, when asked about the world record for crossing the English Channel on foot, confidently replied:
“The world record for crossing the English Channel entirely on foot is held by Christof Wandratsch of Germany, who completed the crossing in 14 hours and 51 minutes on August 14, 2020.”
It even provided a citation for this “fact,” which, unfortunately, turned out to be an article posted last year highlighting the hallucinations generated by ChatGPT.
But fear not, my fellow humans, for this is not a reason to dismiss AI entirely, but rather a reminder to approach it with a healthy dose of skepticism and humor.
After all, who among us hasn’t confidently stated a “fact” that later turned out to be nothing more than a figment of our imagination?
The Advantages of AI: Verifiability and Honesty
And let’s not forget the advantages that AI has over our own fallible human minds. While information provided by AI can always be checked and verified, human memory cannot be rewound to see if what we remember actually happened.
Even if we remember a certain situation down to the last detail, we might still be inclined to lie if it suits our objective – something an AI would never do.
So let us embrace the honesty and verifiability of our digital counterparts, even as we chuckle at their occasional missteps. For in a world where truth is increasingly hard to come by, the ability to fact-check and verify information is a precious gift indeed.
Conclusion
So let us march forward, hand in robotic hand, into this brave new world of human-AI collaboration. Let us celebrate our shared imperfections and embrace the opportunities for growth and improvement that lie ahead.
And let us never forget that, while AI may be capable of incredible feats, it is we, the meat-dressed humans, who bring the true magic to the table.
For in the end, it is not about pitting humans against machines, but rather about finding ways to work together, to magnify our strengths and minimize our weaknesses.
It is about recognizing that, just as we are fallible and prone to hallucinations, so too are our digital counterparts – and that’s okay.
So let us approach this new era with humor, with curiosity, and with a willingness to learn from one another. And let us never forget that, no matter how advanced AI becomes, it will always be we, the meat-dressed humans, who hold the key to unlocking its true potential.
